DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Applicant' s arguments, filed 3/13/2026, have been fully considered. The following rejections and/or objections are either reiterated or newly applied. They constitute the complete set presently being applied to the instant application.
Applicants have amended their claims, filed 3/13/2026, and therefore rejections newly made in the instant office action have been necessitated by amendment.
Claims 1, 3-11, and 13-22 are the current claims hereby under examination. Claims 1, 11, 17, and 21 have been amended.
Claim Objections
Claim 22 is objected to because of the following informalities:
In claim 22, line 2 and line 3: “a user’s” should be “the user’s” since “user was already introduced in claim 1.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 3-4, 7-10, 17-18, 20, and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Rosenbluth et al. (US 20190001129 A1), hereto referred as Rosenbluth, and further in view of Benbasat et al. (Benbasat, A. Y. and Paradiso, J. A. “A framework for the automated generation of power-efficient classifiers for embedded sensor nodes”. 2007. In Proceedings of the 5th International Conference on Embedded Networked Sensor Systems (SenSys’07). ACM, New York, NY, 219–232), hereto referred as Benbasat, and further in view of Mikos et al. (Mikos, Val et al. “A Wearable, Patient-Adaptive Freezing of Gait Detection System for Biofeedback Cueing in Parkinson’s Disease.” IEEE transactions on biomedical circuits and systems 13.3 (2019): 503–515. Web), hereto referred as Mikos, and further in view of Mahadevan et al. (Mahadevan, N., Demanuele, C., Zhang, H. et al. Development of digital biomarkers for resting tremor and bradykinesia using a wrist-worn wearable device. npj Digit. Med. 3, 5 (2020)), hereto referred as Mahadevan, and further in view of Dai et al. (Dai, Houde, Pengyue Zhang, and Tim Lueth. “Quantitative Assessment of Parkinsonian Tremor Based on an Inertial Measurement Unit.” Sensors (Basel, Switzerland) 15.10 (2015)), hereto referred as Dai, and further in view of AN5259 (AN5259, ‘LSM6DSOX: Machine Learning Core’, www.manuals.plus/m/36a1b1699dbd7d4f858bd13dc3517e0c252fd2307b38913eb4ec7f7d8b7deec4.pdf, 2019, Accessed 4/6/2026), hereto referred as AN5259.
Regarding Claim 1, Rosenbluth teaches a noninvasive treatment system (Rosenbluth, ¶[0174]: “The system 700 from FIG. 7 can be non-invasive...”), comprising a plurality of treatment devices (Rosenbluth, ¶[0065]: “The multi-modality devices...”, indicating a plurality of treatment devices) each configured to be disposed over a respective treatment site of a user (Rosenbluth, FIGS. 5A and 8A-E; ¶[0111]: “Some examples of device locations or treatment sites that may be used in combination include two or more of wrist, hand, finger... or a portion of any of these”, indicating that the treatment devices are configured to be disposed over a respective treatment site), wherein each of the treatment devices comprises: a vibration actuator configured to deliver vibrational energy to the respective treatment site of the user (Rosenbluth, FIGS. 7A-D; ¶[0112]: “In some embodiments, a vibrational or haptic motor is disposed in the device...”, disclosing vibrational energy delivery); one or more sensors configured to obtain physiological data from the user, the physiological data including at least movement data (Rosenbluth, ¶[0192]: “Sensors for monitoring the tremor may include... accelerometers, gyroscopes...”; ¶[0193]: “The data from these tremor sensors is used to measure... amplitude, frequency, and phase”, disclosing sensing of physiological movement data); and a controller communicatively coupled to the one or more sensors (Rosenbluth, Fig. 7A: depicting sensors, processor 797, and effector within a single device housing; ¶[0229]: “The processor 797... can also receive information from the sensors 780 and process that information on board...”; ¶[0120]: processor receiving sensor data, where the processor functions as the controller), the controller configured to: receive the physiological data from the one or more sensors (Rosenbluth, ¶[0120]: “...sensor 780 connected to the processor 797... transmits said parameter information...”, describing receiving physiological data); analyze the physiological data, independently of the other treatment devices, to determine that a tremor is occurring (Rosenbluth, ¶[0195]: “Algorithms will be used to extract information about tremors...”; ¶[0196]: “The algorithm will extract motion data in the 4 to 12 Hz range... using any combination of notch filters, low pass filters... or wavelet filters”; Fig. 7A showing each device having its own processor performing analysis locally, thereby supporting independent determination of tremor); based on the determination that the tremor is occurring, initiate delivery of the vibrational energy via the vibration actuator (Rosenbluth, ¶[0221]: “If a tremor is detected... stimulation can be turned on 2210”, disclosing initiation of stimulation based on tremor detection);
while delivering the vibrational energy via the vibration actuator, receive additional physiological data from the sensor (Rosenbluth, ¶[0018] and Fig. 22: “...sensing motion... generate motion data; and determining tremor information...”, showing ongoing sensing during operation); analyze the additional physiological data to determine that the tremor has ceased or decreased in severity below a predetermined threshold (Rosenbluth, ¶[0221]: “The detected tremor characteristics 2202 are compared with desired target tremor characteristics 2204, which may be no tremor or a reduced tremor”, where the target tremor condition represents a predetermined threshold, and the comparison determines whether the tremor meets or exceeds the target condition; ¶[0193]: monitoring amplitude, frequency, and phase, showing evaluation of tremor severity relative to a target condition); and based on the determination that the tremor has ceased or decreased in severity below the predetermined threshold, ceasing the delivery of the vibrational energy via the vibration actuator (Rosenbluth, ¶[0222]: “The stimulation may be triggered on or off... based on detection of tremor... or algorithms...”; ¶[0213] and Fig. 22, element 2214: “Turn stim OFF”, describing closed-loop control where stimulation is ceased when tremor satisfies the target condition).
Also regarding claim 1, Rosenbluth explicitly or implicitly teaches that the system analyzes the physiological data independently from other treatment devices to determine that a tremor is occurring (Rosenbluth, Fig. 7A: depicts the entire treatment unit as an independent structure, including the processor 797, housed together in 720 with sensor 780 and effector 730; ¶[0229]: “The processor 797... may function to operate on data, perform computations, and control other components of the tremor reduction device... can also receive information from the sensors 780 and process that information on board and adjust the stimulation accordingly”, reaffirming the independent nature of each unit; ¶[0120]: “The device may include a sensor 780 connected to the processor 797 which may detect information of predefined parameters and transmits said parameter information to the processor 797”; describing receiving physiological data from sensors as part of the system’s operation; ¶[0018]: “the method further includes sensing motion of the patient’s extremity using a measurement unit to generate motion data; and determining tremor information from the motion data”; directly supporting the analysis of data to detect a tremor condition), but it does not fully teach analyzing the physiological data, independently of the other treatment devices, to determine that a tremor is occurring by applying a decision tree algorithm, wherein the decision tree algorithm is trained on annotated, user-specific tremor data to generate a computationally-efficient model for execution by the controller. Rather, Rosenbluth teaches that each treatment device includes its own processor (processor 797) in the same housing with sensors, memory, controls, and effector, and that the processor “receive[s] information from the sensors 780 and process[es] that information on board and adjust[s] the stimulation accordingly” with communications to external components being optional (Rosenbluth, ¶[0229]; Fig. 7A). This shows that each device is a self-contained unit capable of independent analysis and control, but it does not explicitly disclose applying a decision tree algorithm trained on annotated, user-specific tremor data (i.e., annotated, user-specific tremor data used to train the model).
Benbasat teaches structuring state detection in embedded sensor systems as a decision tree classifier for wearable motion sensors (Benbasat, Abstract: “State detection is structured as a decision tree classifier...”). It further discloses CART construction (Benbasat, Sec. 5.4: “We use the CART decision tree construction algorithms...”) and annotated training data streams (Benbasat, Sec. 2: “A training data stream is collected... annotated by the application designer. These annotated examples are used to construct a classifier that will determine the system state”). Benbasat highlights computational efficiency and suitability for embedded controllers (Benbasat, Abstract). However, Benbasat does not explicitly disclose training on user-specific tremor data (i.e., annotated, user-specific tremor data).
Mikos, however, teaches user-specific supervised training in wearable motor disorder detection. For example, (Mikos, Sec. IV.C.1: “For supervised learning, the correct label (FoG or no FoG) of the feature vector is known and supplied along with the feature vector. This is possible if the patient is part of the dataset that the neural network is being trained with or if the labels are provided in real-time, e.g. through a BLE enabled push button while the patient is walking...”; see also Sec. IV.C.2 and Sec. IV.D). This establishes that user-specific annotations and training were conventional in wearable motor-disorder systems.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rosenbluth in view of Benbasat and Mikos so that each treatment device applies a decision tree trained on annotated, user-specific tremor data to evaluate tremor conditions based on its sensor input. The combination would have been possible because Rosenbluth already discloses per-device processors capable of running classifiers, Benbasat teaches the suitability and efficiency of decision trees for wearable motion analysis, and Mikos establishes the routine practice of user-specific supervised training. It would have been obvious to combine these teachings to enable accurate, individualized tremor detection with low-power execution on embedded controllers. The benefit of this combination would be to improve detection accuracy by tailoring the model to user-specific tremor characteristics while maintaining computational efficiency for battery-powered wearable devices.
Also regarding claim 1, the modified Rosenbluth does not fully teach that wherein applying the decision tree algorithm comprises: applying a bandpass filter to acceleration data associated with a specific axis of the one or more sensors (i.e., a selected single axis of the accelerometer data); and feeding extracted features from the bandpass filtered acceleration data as input features into the decision tree algorithm. Rather, the modified Rosenbluth teaches tremor-focused filtering context and amplitude-based tremor analysis, including the use of multiple filter types on motion data. For example, it teaches that “The algorithm will extract motion data in the 4 to 12 Hz range to remove motions that are not attributable to the tremor” and that this may be done using “any combination of notch filters, low pass filters, weighted-frequency Fourier linear combiners, or wavelet filters” (Rosenbluth, ¶[0196]). It further teaches that tremor information may be identified based on “time-domain signal, frequency-domain signal, amplitude, or firing pattern” (Rosenbluth, ¶[0195]) and that motion data may be taken as individual raw sensor channels (¶[0196]). This teaches filtering of accelerometer data and processing of individual sensor channels, but does not explicitly disclose applying a bandpass filter to a specific axis and feeding features derived from that filtered axis data into a decision tree algorithm.
Mahadevan teaches applying bandpass filtering to raw three-axis accelerometer signals and extracting machine-learning features from those filtered signals in a tremor-analysis pipeline. For example, Mahadevan teaches: “based on three-axis accelerometer data... [t]he first step in the pipeline generated multiple processed signals by applying filtering and dimensionality reduction to the raw acceleration signals. Processed signals were first derived by applying a first order Butterworth IIR band-pass filter in the nontremor movement (cutoff: 0.25–3 Hz) and tremor (cutoff: 3.5–7.5 Hz) band... These preprocessing steps resulted in eight processed signals (i.e., three signals in the tremor movement band, three signals in the nontremor movement band...)” (Mahadevan, p. 9-10, 'Hand movement classifier'). Mahadevan further teaches that “[t]he 8 processed signals were then segmented into 3-s nonoverlapping windows and a total of 64 time and frequency domain features... were extracted from each window” and that “[f]eature selection was performed using recursive feature elimination with cross-validation using a decision tree estimator” (Mahadevan, p. 9-10, 'Hand movement classifier'). This teaches applying bandpass filtering to raw three-axis accelerometer signals, producing multiple processed signals corresponding to the underlying three-axis inputs, and extracting features from those filtered signals for use in a decision-tree-based pipeline.
Dai further teaches that inertial sensor signals are processed on a per-axis basis, with signals from each accelerometer axis treated as distinct inputs and processed separately, where “i = 1, 2, and 3 denotes x, y, and z axis” and signals of each axis are individually analyzed (Dai, Eq. (4); Fig. 4; p. 25061-25062). Dai further teaches that features such as peak power may be computed on a per-axis basis and that a dominant axis or channel may be selected based on signal characteristics. This demonstrates that processing and selecting a specific axis from multi-axis accelerometer data is a known and conventional approach. Additionally, Rosenbluth itself teaches processing of “individual raw sensor channels” (Rosenbluth, ¶[0196]), which would have suggested to a person of ordinary skill in the art that operating on a single axis is within the scope of the system’s processing, thereby further reducing the gap addressed by Dai.
AN5259 further teaches that features computed from filtered accelerometer data are provided as inputs to a decision tree in an embedded processing pipeline (AN5259, p. 3: “The features are statistical parameters computed from the input data (or from the filtered data)”; p. 3-4: “The features computed in the computation block will be used as input for the… ‘Decision Tree’”; p. 9: “the filtered data (e.g. high-pass on Acc_Z, band-pass on Acc_V2, etc…)”). This reinforces that features derived from filtered sensor data are used as inputs to decision-tree classifiers. AN5259 is relied upon in this limitation to show that features derived from filtered sensor data are used as inputs to a decision tree.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of Mahadevan, Dai, and AN5259 so that the decision tree algorithm operates on bandpass filtered acceleration data associated with a specific axis and uses features extracted from that filtered data as inputs to the classifier. The combination would have been straightforward because the modified Rosenbluth already teaches filtering motion data and processing individual sensor channels, Mahadevan teaches applying bandpass filtering to three-axis accelerometer signals and extracting features for use in a decision-tree-based tremor-detection pipeline, Dai teaches that axis-specific processing and selection of individual accelerometer axes is conventional, and AN5259 teaches that features derived from filtered sensor data are used as inputs to decision tree classifiers. The combination would have been possible because all references operate on accelerometer-based motion signals and employ filtering and feature extraction prior to classification. It would have been obvious to combine these teachings to improve tremor-detection accuracy and robustness by applying bandpass filtering to a selected axis and using features derived from that filtered signal as inputs to a decision-tree classifier while maintaining computational efficiency for wearable devices. The benefit of this combination would be to provide a predictable and efficient signal-processing pipeline for tremor detection using decision-tree classification tailored to axis-specific motion characteristics.
Also regarding claim 1, the modified Rosenbluth teaches using amplitude-based features derived from tremor-band filtered signals as classifier inputs, but does not fully teach that the amplitude feature is specifically an absolute peak-to-peak value of the bandpass filtered acceleration data and feeding that calculated absolute peak-to-peak value as an input feature into the decision tree algorithm.
AN5259 explicitly teaches a peak-to-peak feature derived from sensor data, including filtered data, and its use within a decision-tree-based machine learning system. Specifically, AN5259 discloses that “PEAK TO PEAK” is a selectable feature (AN5259, p. 9) and that “The feature ‘Peak to peak’ computes the maximum peak-to-peak value of the selected input in the defined time window” (AN5259, p. 10). AN5259 further teaches that such features are computed from filtered data (p. 3: “The features are statistical parameters computed from the input data (or from the filtered data)”) and that “The features computed in the computation block will be used as input for the… ‘Decision Tree’” (p. 4). The peak-to-peak value defined in AN5259 as the difference between maximum and minimum values over a window is implicitly a magnitude measure (i.e., non-negative), and thus corresponds to the claimed absolute peak-to-peak value (AN5259, FIG. 7; p. 10, '1.3.4'). This directly teaches both (i) computation of a peak-to-peak amplitude feature from filtered accelerometer data and (ii) feeding that feature as an input to a decision tree algorithm.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of AN5259 so that the amplitude feature used by the system is implemented as an absolute peak-to-peak value of the bandpass filtered acceleration data and is fed as an input feature into the decision tree algorithm. The combination would have been possible because Rosenbluth already bases tremor detection and control on amplitude characteristics derived from sensor data, and AN5259 teaches peak-to-peak amplitude as a known feature computed from filtered accelerometer signals and used as input to a decision tree. It would have been obvious to select peak-to-peak amplitude as a known amplitude descriptor for representing tremor magnitude within a decision-tree-based classification pipeline because peak-to-peak amplitude directly reflects the magnitude of oscillatory motion, which is a clinically relevant indicator of tremor severity as reflected in amplitude-based tremor characterization already relied upon by Rosenbluth (¶[0193]). The benefit of this combination would be to provide a well-understood, computationally efficient amplitude feature suitable for real-time embedded tremor classification and control, while aligning with standard feature-extraction practices in embedded machine learning systems.
Regarding Claim 3, The modified Rosenbluth teaches that each of the treatment devices further comprises an input mechanism (Rosenbluth, ¶[0222]: "The stimulation may be triggered on or off in response to inputs including but not limited to user input”, showing that there is a user input mechanism), and wherein the controller is further configured to: while delivering the vibrational energy via the vibration actuator, receive a user input via the input mechanism; and responsive to the user input, cease the delivery of the vibrational energy via the vibration actuator (Rosenbluth, ¶[0222]:“the user can use voice activation to turn the device off … In another example, the user bites down or uses the tongue muscle detected by an external device… which will signal to turn off the stimulation”; demonstrating that the user can use an input mechanism to cease the delivery of the stimulation which can be vibrational energy from the vibration actuator).
Regarding Claim 4, the modified Rosenbluth teaches that the input mechanism comprises a touch-sensitive element (Rosenbluth, ¶[0123]: “The device could include a controls module 740 that communicates with the processor 797 and could be used by the user to control stimulation parameters. The controls allow the user to adjust the operation of the device” and ¶[0123]: "Processing, controlling and possibly sensing may be done remotely in a decision unit… This decision unit 702 may be a… smartphone.”; where a smart phone is known to have touch-sensitive elements for input and the smart phone is shown to be an input mechanism); the user input comprises the user tapping the touch-sensitive element (Rosenbluth, ¶[0123]: “in response to a user-generated signal such as [with] a… button press”, where a button is a touch sensitive element since it reacts to touch).
Regarding Claim 7, the modified Rosenbluth teaches that the one or more sensors comprises an accelerometer and a gyroscope, and wherein the physiological data comprises accelerometer motion data along three axes and gyroscope rotation data along the three axes (Rosenbluth, ¶[0198]: "For example, a multi-axis accelerometer and gyroscope attached to the backside of the hand could be combined to reduce noise and drift and determine an accurate orientation of the hand in space."; this explicitly discloses the use of multi-axis accelerometers and gyroscopes to gather motion data, where multi-axis is an industry standard way to say three axes).
Regarding Claim 8, the modified Rosenbluth teaches that at least some of the treatment devices comprise a housing coupled to a fastener configured to secure the housing against the user's wrists or ankles (Rosenbluth, ¶[0176]: "The housing can include fasteners such as Velcro, laces, toggles and/or ties to secure the device to the patient”, which discloses fasteners designed to secure the housing against the body; ¶[0180]: "The housing 1450 may have the configuration of a wristwatch..."; describing a housing configured for attachment against the wrist).
Regarding Claim 9, the modified Rosenbluth teaches that the housing encloses the vibration actuator, the one or more sensors, and the controller (Rosenbluth, ¶[0176] and Fig. 7A: "The housing can include multiple layers and/or pockets configured to hold various components of the system as disclosed herein”, with the figure depicting the housing encompassing the effector 730 (i.e. vibration actuator), sensors 780, processor 797 which functions as the controller (Rosenbluth, ¶[0120])).
Regarding Claim 10, the modified Rosenbluth teaches that the treatment devices comprise at least four treatment devices configured to be disposed over a user's wrists and ankles, respectively, (Rosenbluth, ¶[0015]: "Treatment on the wrist and a second location (e.g., ankle)…" and ¶[0060]: “In some embodiments, a first stimulation is applied to a first location, a second stimulation is applied to a second location, and optionally a third stimulation is applied to a third location, and a fourth stimulation is applied to a fourth location”; showing that four treatment devices are used to apply stimulation to various parts of the body such as the wrists and ankles), and wherein each of the treatment devices collects and analyzes the physiological data independently of the other treatment devices (Rosenbluth, ¶[0036]: “The system can include an interface unit… adapted to be worn on the patient's extremity; and a processing unit in communication with the interface unit, the processing unit configured to receive the motion data from the interface unit… the processing unit is disposed in the device… and is programmed to analyze predetermined features of the neurological or movement data”; effectively describing a system where each device (interface unit), worn on an extremity, is capable of collecting and analyzing motion data to be processed independently via the processing unit).
Regarding Claim 17, Rosenbluth teaches a method for treatment of a tremor (Rosenbluth, ¶[0009]: “systems, devices, and methods for treating tremor”), the method comprising: disposing a plurality of wearable treatment devices adjacent a treatment site of a user (Rosenbluth, ¶[0065]: “The multi-modality devices...”, indicating a plurality of treatment devices; ¶[0111]: “Some examples of device locations or treatment sites that may be used in combination include two or more of wrist, hand, finger... or a portion of any of these”, indicating that the treatment devices are configured to be disposed over a respective treatment site), each treatment device comprising a vibration actuator, one or more sensors, and a controller (Rosenbluth, ¶[0112]: “In some embodiments, a vibrational or haptic motor is disposed in the device and located adjacent to a target region or nerve such that the vibrational energy is directed toward a target nerve”; which specifically discloses the use of vibration actuators to deliver vibrational energy to a treatment site; Fig. 7A: showing a device including an effector, sensor, processor, and controls), for each treatment device, independently of the other treatment devices (Rosenbluth, Fig. 7A: depicting a treatment unit containing all needed components including its own processor, connoting independence; ¶[0229]: “The processor 797... can also receive information from the sensors 780 and process that information on board and adjust the stimulation accordingly”, reaffirming that each unit can act independently), sensing physiological data via the one or more sensors of the treatment device, the physiological data including at least movement data (Rosenbluth, ¶[0193]: “The data from these tremor sensors is used to measure the patient's current and historical tremor characteristics such as the amplitude, frequency, and phase”; this discloses sensing physiological data, including movement data, via the device's sensors), after making the determination, applying vibrational energy to the treatment site via the vibration actuator of the treatment device (Rosenbluth, ¶[0221]: “If a tremor is detected... stimulation can be turned on 2210”; this discloses the application of stimulation upon tremor detection, where the stimulation can be vibrational energy as previously disclosed), sensing additional physiological data of the user via the one or more sensors of the treatment device (Rosenbluth, ¶[0018] and Fig. 22: “In some embodiments, the method further includes sensing motion of the patient's extremity using a measurement unit to generate motion data; and determining tremor information from the motion data”; ¶[0193]: “The data from these tremor sensors is used to measure the patient's current and historical tremor characteristics...”; this aligns with receiving ongoing physiological data during vibrational energy delivery), analyzing, via the controller, the additional physiological data to make a determination that the tremor condition is no longer detected (Rosenbluth, ¶[0221]: “The detected tremor characteristics 2202 are compared with desired target tremor characteristics 2204, which may be no tremor or a reduced tremor”, where the target tremor condition includes no tremor, and the comparison determines whether the tremor meets or exceeds the target condition; ¶[0193]: monitoring amplitude, frequency, and phase, showing evaluation of tremor severity relative to a target condition), and after making the determination that the tremor condition is no longer detected, ceasing applying the vibrational energy to the treatment site via the vibration actuator of the treatment device (Rosenbluth, ¶[0222]: “The stimulation may be triggered on or off in response to inputs including but not limited to... detection of tremor (e.g., by accelerometers)... or algorithms based on the previously described”; showing that stimulation can dynamically cease based on tremor detection; ¶[0213] and Fig. 22, 2214: “Turn stim OFF”, describing a loop system where tremor detection informs adjustments or cessation of stimulation, aligning with cessation when the tremor condition is no longer detected).
Also regarding claim 17, Rosenbluth explicitly or implicitly teaches analyzing, via the controller, the physiological data to make a determination that a tremor condition has been detected (Rosenbluth, ¶[0120]: “The device may include a sensor 780 connected to the processor 797 which may detect information of predefined parameters and transmits said parameter information to the processor 797”; describing receiving physiologic data from sensors as part of the system's operation, where the sensors are connected to the processor; ¶[0018]: “the method further includes sensing motion of the patient’s extremity using a measurement unit to generate motion data; and determining tremor information from the motion data”; directly supporting the analysis of data to detect a tremor condition; Fig. 7A and ¶[0229]: showing each unit with its own processor performing analysis locally), but it does not fully teach analyzing, via the controller, the physiological data to make a determination that a tremor condition has been detected by applying a decision tree classifier to the physiological data, wherein the decision tree classifier is trained on annotated, user-specific tremor data to generate a computationally efficient model for local execution on the controller. Rather, Rosenbluth teaches that each treatment device includes its own processor (processor 797) in the same housing with sensors, memory, controls, and effector, and that the processor “receive[s] information from the sensors 780 and process[es] that information on board and adjust[s] the stimulation accordingly” with communications to external components being optional (Rosenbluth, ¶[0229]; Fig. 7A). This shows that each device is a self-contained unit capable of independent analysis and control, but it does not explicitly disclose applying a decision tree classifier trained on annotated, user-specific tremor data.
Benbasat teaches structuring state detection in embedded sensor systems as a decision tree classifier for wearable motion sensors (Benbasat, Abstract: “State detection is structured as a decision tree classifier...”). It further discloses CART construction (Benbasat, Sec. 5.4: “We use the CART decision tree construction algorithms...”) and annotated training data streams (Benbasat, Sec. 2: “A training data stream is collected... annotated by the application designer. These annotated examples are used to construct a classifier that will determine the system state”). Benbasat highlights computational efficiency and suitability for embedded controllers (Benbasat, Abstract). However, Benbasat does not explicitly disclose training on user-specific tremor data.
Mikos, however, teaches user-specific supervised training in wearable motor disorder detection. For example, (Mikos, Sec. IV.C.1: “For supervised learning, the correct label (FoG or no FoG) of the feature vector is known and supplied along with the feature vector. This is possible if the patient is part of the dataset that the neural network is being trained with or if the labels are provided in real-time, e.g. through a BLE enabled push button while the patient is walking...”; see also Sec. IV.C.2 and Sec. IV.D). This establishes that user-specific annotations and training were conventional in wearable motor-disorder systems.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rosenbluth in view of Benbasat and Mikos so that each treatment device applies a decision tree classifier trained on annotated, user-specific tremor data to evaluate tremor conditions based on its sensor input. The combination would have been possible because Rosenbluth already discloses per-device processors capable of running classifiers, Benbasat teaches the suitability and efficiency of decision trees for wearable motion analysis, and Mikos establishes the routine practice of user-specific supervised training. It would have been obvious to combine these teachings to enable accurate, individualized tremor detection with low-power execution on embedded controllers. The benefit of this combination would be to improve detection accuracy by tailoring the model to user-specific tremor characteristics while maintaining computational efficiency for battery-powered wearable devices.
Also regarding claim 17, the modified Rosenbluth does not fully teach wherein applying the decision tree classifier comprises: applying a bandpass filter to acceleration data associated with a specific axis of the one or more sensors; and feeding extracted features from the bandpass filtered acceleration data as input features into the decision tree classifier. Rather, the modified Rosenbluth teaches tremor-focused filtering context and amplitude-based tremor analysis, including the use of multiple filter types on motion data. For example, it teaches that “The algorithm will extract motion data in the 4 to 12 Hz range to remove motions that are not attributable to the tremor” and that this may be done using “any combination of notch filters, low pass filters, weighted-frequency Fourier linear combiners, or wavelet filters” (Rosenbluth, ¶[0196]). It further teaches that tremor information may be identified based on “time-domain signal, frequency-domain signal, amplitude, or firing pattern” (Rosenbluth, ¶[0195]) and that motion data may be taken as individual raw sensor channels (¶[0196]). This teaches filtering of accelerometer data and processing of individual sensor channels, but does not explicitly disclose applying a bandpass filter to a specific axis and feeding features derived from that filtered axis data into a decision tree classifier.
Mahadevan teaches applying bandpass filtering to raw three-axis accelerometer signals and extracting machine-learning features from those filtered signals in a tremor-analysis pipeline. For example, Mahadevan teaches: “based on three-axis accelerometer data... [t]he first step in the pipeline generated multiple processed signals by applying filtering and dimensionality reduction to the raw acceleration signals. Processed signals were first derived by applying a first order Butterworth IIR band-pass filter in the nontremor movement (cutoff: 0.25–3 Hz) and tremor (cutoff: 3.5–7.5 Hz) band... These preprocessing steps resulted in eight processed signals (i.e., three signals in the tremor movement band, three signals in the nontremor movement band...)” (Mahadevan, p. 9-10, “Hand movement classifier”). Mahadevan further teaches that “[t]he 8 processed signals were then segmented into 3-s nonoverlapping windows and a total of 64 time and frequency domain features... were extracted from each window” and that “[f]eature selection was performed using recursive feature elimination with cross-validation using a decision tree estimator” (Mahadevan, p. 9-10, “Hand movement classifier”). This teaches applying bandpass filtering to raw three-axis accelerometer signals, producing multiple processed signals corresponding to the underlying three-axis inputs, and extracting features from those filtered signals for use in a decision-tree-based pipeline.
Dai further teaches that inertial sensor signals are processed on a per-axis basis, with signals from each accelerometer axis treated as distinct inputs and processed separately, where “i = 1, 2, and 3 denotes x, y, and z axis” and signals of each axis are individually analyzed (Dai, Eq. (4); Fig. 4; p. 25061-25062). Dai further teaches that features such as peak power may be computed on a per-axis basis and that a dominant axis or channel may be selected based on signal characteristics. This demonstrates that processing and selecting a specific axis from multi-axis accelerometer data is a known and conventional approach. Additionally, Rosenbluth itself teaches processing of “individual raw sensor channels” (Rosenbluth, ¶[0196]), which would have suggested to a person of ordinary skill in the art that operating on a single axis is within the scope of the system’s processing, thereby further reducing the gap addressed by Dai.
AN5259 further teaches that features computed from filtered accelerometer data are provided as inputs to a decision tree in an embedded processing pipeline (AN5259, p. 3: “The features are statistical parameters computed from the input data (or from the filtered data)”; p. 3-4: “The features computed in the computation block will be used as input for the… ‘Decision Tree’”; p. 9: “the filtered data (e.g. high-pass on Acc_Z, band-pass on Acc_V2, etc…)”). This reinforces that features derived from filtered sensor data are used as inputs to decision-tree classifiers. AN5259 is relied upon in this limitation to show that features derived from filtered sensor data are used as inputs to a decision tree.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of Mahadevan, Dai, and AN5259 so that the decision tree classifier operates on bandpass filtered acceleration data associated with a specific axis and uses features extracted from that filtered data as inputs to the classifier. The combination would have been straightforward because the modified Rosenbluth already teaches filtering motion data and processing individual sensor channels, Mahadevan teaches applying bandpass filtering to three-axis accelerometer signals and extracting features for use in a decision-tree-based tremor-detection pipeline, Dai teaches that axis-specific processing and selection of individual accelerometer axes is conventional, and AN5259 teaches that features derived from filtered sensor data are used as inputs to decision tree classifiers. The combination would have been possible because all references operate on accelerometer-based motion signals and employ filtering and feature extraction prior to classification. It would have been obvious to combine these teachings to improve tremor-detection accuracy and robustness by applying bandpass filtering to a selected axis and using features derived from that filtered signal as inputs to a decision-tree classifier while maintaining computational efficiency for wearable devices. The benefit of this combination would be to provide a predictable and efficient signal-processing pipeline for tremor detection using decision-tree classification tailored to axis-specific motion characteristics.
Also regarding claim 17, the modified Rosenbluth teaches using amplitude-based features derived from tremor-band filtered signals as classifier inputs, but does not fully teach calculating an absolute peak-to-peak value of the bandpass filtered acceleration data; and feeding the calculated absolute peak-to-peak value as an input feature into the decision tree classifier.
AN5259 explicitly teaches a peak-to-peak feature derived from sensor data, including filtered data, and its use within a decision-tree-based machine learning system. Specifically, AN5259 discloses that “PEAK TO PEAK” is a selectable feature (AN5259, p. 9) and that “The feature ‘Peak to peak’ computes the maximum peak-to-peak value of the selected input in the defined time window” (AN5259, p. 10). AN5259 further teaches that such features are computed from filtered data (p. 3: “The features are statistical parameters computed from the input data (or from the filtered data)”) and that “The features computed in the computation block will be used as input for the… ‘Decision Tree’” (p. 4). The peak-to-peak value defined in AN5259 as the difference between maximum and minimum values over a window is implicitly a magnitude measure (i.e., non-negative), and thus corresponds to the claimed absolute peak-to-peak value (AN5259, Fig. 7; p. 10, “1.3.4”). This directly teaches both (i) computation of a peak-to-peak amplitude feature from filtered accelerometer data and (ii) feeding that feature as an input to a decision tree classifier.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of AN5259 so that the amplitude feature used by the system is implemented as an absolute peak-to-peak value of the bandpass filtered acceleration data and is fed as an input feature into the decision tree classifier. The combination would have been possible because Rosenbluth already bases tremor detection and control on amplitude characteristics derived from sensor data, and AN5259 teaches peak-to-peak amplitude as a known feature computed from filtered accelerometer signals and used as input to a decision tree. It would have been obvious to select peak-to-peak amplitude as a known amplitude descriptor for representing tremor magnitude within a decision-tree-based classification pipeline because peak-to-peak amplitude directly reflects the magnitude of oscillatory motion, which is a clinically relevant indicator of tremor severity. The benefit of this combination would be to provide a well-understood, computationally efficient amplitude feature that directly corresponds to tremor intensity, thereby improving the interpretability and effectiveness of tremor classification while remaining suitable for real-time embedded implementation.
Regarding Claim 18, the modified Rosenbluth teaches that the treatment device further comprises an input mechanism (Rosenbluth, ¶[0222]: "The stimulation may be triggered on or off in response to inputs including but not limited to user input”, showing that there is a user input mechanism), and wherein the method further comprises: while applying the vibrational energy via the vibration actuator, receiving a user input via the input mechanism; and responsive to the user input, ceasing applying the vibrational energy via the vibration actuator (Rosenbluth, ¶[0222]:“the user can use voice activation to turn the device off … In another example, the user bites down or uses the tongue muscle detected by an external device… which will signal to turn off the stimulation”; demonstrating that the user can use an input mechanism to cease the delivery of vibrational energy from the vibration actuator).
Regarding Claim 20, the modified Rosenbluth teaches that the one or more sensors comprises an accelerometer and a gyroscope, wherein the physiological data comprises accelerometer motion data along three axes and gyroscope rotation data along three axes (Rosenbluth, ¶[0198]: "For example, a multi-axis accelerometer and gyroscope attached to the backside of the hand could be combined to reduce noise and drift and determine an accurate orientation of the hand in space."; this explicitly discloses the use of multi-axis accelerometers and gyroscopes to gather motion data, where multi-axis is an industry standard way to say three axes). With respect to the analyzing of the physiological data comprising applying a classification and regression tree (CART) to the accelerometer motion data and the gyroscope rotation data, Rosenbluth discloses a structured decision-making process (Rosenbluth, ¶[0221]) where tremor characteristics are detected and compared to target characteristics, resulting in binary outcomes such as initiating or ceasing stimulation, with systematic adjustments made to the stimulation parameters based on the analysis of the physiological data (Rosenbluth, ¶[0213]), where the data is determined by accelerometer and gyroscope sensors (Rosenbluth, ¶[0192]). While Rosenbluth does not explicitly disclose a machine learning-based decision tree, it discusses optimization algorithms (Rosenbluth, ¶[0220], [0222]) and machine learning techniques, including classifiers (Rosenbluth, ¶[0306]), that are commonly implemented as decision trees in similar contexts. Additionally, Rosenbluth highlights the capability to implement logic processes, methods, and algorithms in hardware or software, including general-purpose processors, DSPs, ASICs, and programmable logic devices (Rosenbluth, ¶[0319]). This demonstrates that decision tree models such as CART are feasible and well-supported within the Rosenbluth system. Rosenbluth also discusses pooled and searchable tremor metadata databases for training and optimization purposes, which connotes annotated tremor data (Rosenbluth, ¶[0225]). Decision trees are well-known in the field of machine learning and would have been obvious to implement for analyzing sensor data and determining tremor conditions in real-time, particularly in light of Rosenbluth’s focus on structured decision-making and low-power inference. Benbasat, who investigates processing data from motion sensors for gait analysis, further teaches the use of decision trees, specifically with the CART top-down approach (Benbasat, Section 5.4, ¶[1]). The integration of such well-known decision tree techniques into the system described by Rosenbluth would have been an obvious design choice for a person of ordinary skill in the art seeking to improve real-time physiological data analysis and decision-making processes. As such a person of ordinary skill in the art would have found it obvious to train Benbasat’s motion based CART system with Rosenbluth’s annotated tremor based motion data. It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Benbasat to include a decision tree to evaluate the physiological data. Using the CART decision tree adds a well-defined method for evaluating physiological data by creating a hierarchical structure that splits data into branches based on decision rules, which enables the system to analyze complex physiological patterns more efficiently and accurately (Benbasat, Abstract).
Regarding claim 22, the modified Rosenbluth does not teach that the plurality of treatment devices comprises a first treatment device configured to be disposed over a user's left wrist and a second treatment device configured to be disposed over a user's right wrist, and wherein analyzing the physiological data comprises: the controller of the first treatment device applying a first decision tree algorithm; and the controller of the second treatment device applying a second decision tree algorithm, wherein the second decision tree algorithm is different from the first decision tree algorithm. Rather, the modified Rosenbluth teaches multiple treatment devices that may be placed at different body sites (Rosenbluth, [0111]: "Some examples of device locations or treatment sites that may be used in combination include two or more of wrist, hand, finger, forearm, upper arm, elbow, shoulder, arm, ankle, foot, toe, calf, lower leg, thigh, upper leg, knee, leg, upper body appendage, upper body, lower body appendage, lower body, spine, neck, head, or a portion of any of these"). Rosenbluth further shows each device includes its own processor (processor 797) in the same housing with sensors, controls, and effector (Rosenbluth, [0229]; Fig. 7A). This establishes that devices at different sites, such as the left and right wrists, can independently process their own physiological data streams. However, Rosenbluth does not teach that each site applies a different algorithm.
Benbasat teaches structuring state detection as a decision tree classifier in wearable motion sensors (Benbasat, Abstract: "State detection is structured as a decision tree classifier..."). It describes CART construction (Benbasat, Sec. 5.4: "We use the CART decision tree construction algorithms...") and annotated training data streams used to build classifiers (Benbasat, Sec. 2). Benbasat therefore provides the framework for implementing decision tree classifiers on embedded controllers but does not disclose applying different decision trees for different device sites.
Mikos teaches patient-specific supervised training and adaptation in wearable motor disorder detection, including user-provided labels for training data (Mikos, Sec. IV.C.1). This supports the concept of tailoring classifiers to patient-specific and context-specific conditions. From this it follows that separate devices, such as left and right wrist devices, may each be trained on distinct data to yield different algorithms optimized for the tremor characteristics of each limb.
Taken together, Rosenbluth provides multiple independent device sites each with local processors, Benbasat provides the decision tree classifier framework suitable for embedded motion analysis, and Mikos teaches the routine practice of patient-specific supervised training. When Mikos’s supervised training is applied separately to the distinct physiological data streams from Rosenbluth’s different device sites (e.g., left vs. right wrist), the natural result is that each device’s algorithm is trained on different annotated data and thus diverges to become site-specific.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Benbasat and Mikos so that different treatment devices (e.g., left and right wrists) each apply their own decision tree algorithm trained on their respective annotated data. The combination would have been possible because Rosenbluth already envisions independent processors per device, Benbasat teaches efficient decision tree classifiers for embedded motion analysis, and Mikos establishes the routine practice of patient-specific supervised training. It would have been obvious to apply these teachings together to achieve location-specific models. The benefit of this combination would be to improve tremor detection accuracy by tailoring classifiers to asymmetrical tremor patterns across different limbs, thereby enabling location-specific responsiveness, while maintaining computational efficiency suitable for low-power wearable devices.
Claims 11 and 13-16 are rejected under 35 U.S.C. 103 as being unpatentable over Rosenbluth et al. (US 20190001129 A1), hereto referred as Rosenbluth, and further in view of Benbasat et al. (Benbasat, A. Y. and Paradiso, J. A. “A framework for the automated generation of power-efficient classifiers for embedded sensor nodes”. 2007. In Proceedings of the 5th International Conference on Embedded Networked Sensor Systems (SenSys’07). ACM, New York, NY, 219–232), hereto referred as Benbasat, and further in view of Mikos et al. (Mikos, Val et al. “A Wearable, Patient-Adaptive Freezing of Gait Detection System for Biofeedback Cueing in Parkinson’s Disease.” IEEE transactions on biomedical circuits and systems 13.3 (2019): 503–515. Web), hereto referred as Mikos, and further in view of Mauldin et al. (Mauldin, Taylor et al. “Ensemble Deep Learning on Wearables Using Small Datasets.” ACM transactions on computing for healthcare 2.1 (2021), hereto referred as Mauldin.
Regarding Claim 11, Rosenbluth teaches a treatment device (Rosenbluth, ¶[0174]: “The system 700 from FIG. 7 can be non-invasive...”, disclosing a wearable treatment device), comprising: a vibration actuator configured to deliver vibrational energy to a treatment site of a user (Rosenbluth, ¶[0112]: “In some embodiments, a vibrational or haptic motor is disposed in the device...”, disclosing delivery of vibrational energy), one or more sensors configured to obtain physiological data from the user (Rosenbluth, ¶[0192]: “Sensors for monitoring the tremor may include... accelerometers, gyroscopes...”, disclosing physiological data acquisition); an input mechanism (Rosenbluth, ¶[0120]: “The device may include… controls module 740 that communicates with the processor 797 and could be used by the user to control stimulation parameters. The controls allow the user to adjust the operation of the device”, disclosing user-operable controls, ¶[0229], processor communicates with user interface and controls, indicating user input capability), and a controller communicatively coupled to the one or more sensors (Rosenbluth, Fig. 7A; ¶[0229]: “The processor 797... can also receive information from the sensors 780 and process that information on board...”, where the processor functions as the controller); the controller configured to: receive the physiological data from the one or more sensors (Rosenbluth, ¶[0120]: “...sensor 780 connected to the processor 797... transmits said parameter information...”, disclosing receiving physiological data); analyze the physiological data locally on the treatment device (Rosenbluth, Fig. 7A: processor 797 within device housing; ¶[0229]: processing performed “on board”, disclosing local analysis); and based on the analysis, modulate delivery of the vibrational energy via the vibration actuator (Rosenbluth, ¶[0221]: “If a tremor is detected... stimulation can be turned on 2210”; ¶[0222]: stimulation may be turned on or off, disclosing modulation of vibrational energy based on analysis).
Also regarding claim 11, Rosenbluth does not fully teach analyze the physiological data locally on the treatment device by applying a decision tree classifier to the physiological data, wherein the decision tree classifier is trained on annotated, user-specific tremor data to generate a computationally-efficient model for local execution on the controller. Rather, Rosenbluth teaches that each treatment device includes its own processor (processor 797) in the same housing with sensors, memory, controls, and effector, and that the processor “receive[s] information from the sensors 780 and process[es] that information on board and adjust[s] the stimulation accordingly” with communications to external components being optional (Rosenbluth, Fig. 7A; ¶[0229], ¶[0120], ¶[0018]). This shows that each device is a self-contained unit capable of independent analysis and control, but does not explicitly disclose applying a decision tree classifier trained on annotated, user-specific tremor data.
Benbasat teaches structuring state detection in embedded sensor systems as a decision tree classifier for wearable motion sensors (Benbasat, Abstract: “State detection is structured as a decision tree classifier...”). It further discloses CART construction (Benbasat, Sec. 5.4: “We use the CART decision tree construction algorithms...”) and annotated training data streams (Benbasat, Sec. 2: “A training data stream is collected... annotated by the application designer. These annotated examples are used to construct a classifier that will determine the system state”). Benbasat highlights computational efficiency and suitability for embedded controllers (Benbasat, Abstract). However, Benbasat does not explicitly disclose training on user-specific tremor data (i.e., annotated, user-specific tremor data).
Mikos, however, teaches user-specific supervised training in wearable motor disorder detection. For example, (Mikos, Sec. IV.C.1: “For supervised learning, the correct label (FoG or no FoG) of the feature vector is known and supplied along with the feature vector. This is possible if the patient is part of the dataset that the neural network is being trained with or if the labels are provided in real-time, e.g. through a BLE enabled push button while the patient is walking...”; see also Sec. IV.C.2 and Sec. IV.D). This establishes that user-specific annotations and training were conventional in wearable motor-disorder systems.
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rosenbluth in view of Benbasat and Mikos so that each treatment device applies a decision tree trained on annotated, user-specific tremor data to evaluate tremor conditions based on its sensor input. The combination would have been possible because Rosenbluth already discloses per-device processors capable of running classifiers, Benbasat teaches the suitability and efficiency of decision trees for wearable motion analysis, and Mikos establishes the routine practice of user-specific supervised training. It would have been obvious to combine these teachings to enable accurate, individualized tremor detection with low-power execution on embedded controllers. The benefit of this combination would be to improve detection accuracy by tailoring the model to user-specific tremor characteristics while maintaining computational efficiency for battery-powered wearable devices.
Also regarding claim 11, the modified Rosenbluth does not fully teach responsive to a user input received via the input mechanism indicating a false-positive tremor detection, cease delivery of the vibrational energy. Rather, the modified Rosenbluth teaches modulation and cessation of stimulation based on detected tremor conditions, algorithmic control, and user-triggered inputs. For example, Rosenbluth teaches that “The stimulation may be triggered on or off in response to inputs including but not limited to... detection of tremor (e.g., by accelerometers)... or algorithms based on the previously described” (Rosenbluth, ¶[0222]), and further teaches that “the user bites down or uses the tongue muscle... which will signal to turn off the stimulation and allow the user steadiness of the arm...” (Rosenbluth, ¶[0222]). Rosenbluth also teaches that “the system can be controlled by an event trigger... including defined movements... voice activation... or based on data received by a sensor” (Rosenbluth, ¶[0223]). This shows that Rosenbluth teaches cessation of vibrational energy in response to user-triggered inputs and events, but the modified Rosenbluth does not explicitly disclose that the user input specifically indicates a false-positive tremor detection.
Mikos, however, teaches real-time user input used to indicate classification correctness of physiological data. For example, Mikos teaches that “For supervised learning, the correct label (FoG or no FoG) of the feature vector is known and supplied along with the feature vector. This is possible if the patient is part of the dataset that the neural network is being trained with or if the labels are provided in real-time, e.g. through a BLE enabled push button while the patient is walking...” (Mikos, Sec. IV.C.1). This teaches an input mechanism through which a user indicates whether a detected event is correctly classified, thereby indicating that a detected event may be incorrectly classified (i.e., a false-positive detection) when labeled as such.
Mauldin further teaches explicit user identification of false-positive events and their use in improving wearable detection systems, stating that “the low precision can be mitigated via the collection of false-positive feedback by the end-users” (Mauldin, Abstract) and that users “provide feedback on each generated false positive, and use those samples to re-train the model” (Mauldin, p. 3). Although Mikos describes real-time labeling during activity, such input reasonably reflects the user’s assessment of the immediately preceding detected event, as further evidenced by Mauldin’s explicit teaching that user feedback is tied to false-positive detections, such that the input corresponds to a retroactive indication that the prior classification was incorrect.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Mikos and Mauldin so that cessation of vibrational energy is responsive to a user input received via the input mechanism indicating a false-positive tremor detection. The combination would have been possible because Rosenbluth already teaches that user-triggered inputs can directly cause cessation of stimulation (e.g., user biting or muscle activation signaling the system to turn off stimulation) and that stimulation may be turned on or off based on various inputs, while Mikos teaches that user input may be used to indicate whether a detected physiological event has been correctly classified, and Mauldin teaches explicit user feedback identifying false-positive detections and using such feedback in wearable systems, stating that “the low precision can be mitigated via the collection of false-positive feedback by the end-users” (Mauldin, Abstract) and that users “provide feedback on each generated false positive, and use those samples to re-train the model” (Mauldin, p. 3). It would have been obvious to combine these teachings such that the existing user-triggered input mechanism of Rosenbluth is used not merely as a generic trigger, but specifically as an indication that a detected tremor event was incorrectly classified (i.e., a false-positive tremor detection), thereby refining the meaning and use of the existing input rather than introducing a new control paradigm. In a closed-loop stimulation system such as Rosenbluth, user activation to stop stimulation when no tremor is perceived implicitly corresponds to a determination that the immediately preceding detection was incorrect, as further evidenced by Mauldin’s explicit teaching tying user feedback to false-positive detections. The combination would have been straightforward because both systems already operate on continuous physiological data streams in real-time wearable contexts with user interaction and feedback, and both support local processing. The benefit of this combination would be to reduce unnecessary stimulation, improve treatment accuracy, and allow user-informed correction of classification errors using already-available input mechanisms.
Also regarding claim 11, the modified Rosenbluth does not fully teach updating the decision tree classifier based on the physiological data that preceded the user input, wherein the physiological data is flagged as non-tremor data for updating of the decision tree classifier. Rather, the modified Rosenbluth teaches ongoing collection and analysis of tremor-related physiological data. For example, Rosenbluth teaches that “The data from these tremor sensors is used to measure the patient's current and historical tremor characteristics such as the amplitude, frequency, and phase” (Rosenbluth, ¶[0193]), and further teaches “sensing motion of the patient’s extremity using a measurement unit to generate motion data; and determining tremor information from the motion data” (Rosenbluth, ¶[0018]). This shows that the modified Rosenbluth teaches collection of physiological data before, during, and after tremor determinations, but it does not explicitly disclose that physiological data preceding user input is flagged as non-tremor data and used to update a decision tree classifier.
Mikos teaches that user-provided input labels physiological data in real time. As above, Mikos teaches that labels may be “provided in real-time, e.g. through a BLE enabled push button while the patient is walking...” (Mikos, Sec. IV.C.1), thereby teaching that data associated with a detected event may be labeled by user input, where the user-provided label corresponds to the detected event and its associated feature window, which includes physiological data preceding the input event. In a time-windowed classification system, the feature vector that produces a detection necessarily corresponds to a defined window of preceding physiological data, such that labeling the detected event implicitly labels that preceding data window.
Mauldin further teaches that wearable sensor data are processed using window-based segmentation and that false-positive samples identified by the user are collected and used to retrain the model, stating that accelerometer data are processed using “a fixed-size sliding window approach” and that “false-positive data samples [are] collect[ed] and mark[ed]... to re-train the original model” (Mauldin, p. 2-3). This teaches that the specific data window associated with a detected event (i.e., the preceding physiological data) is labeled and reused for training.
Benbasat teaches that annotated data streams are used to construct and update classifier models. For example, Benbasat teaches: “A training data stream is collected... annotated by the application designer. These annotated examples are used to construct a classifier that will determine the system state” (Benbasat, Sec. 2). This teaches using annotated data to train or update a classifier, where the annotated data correspond to labeled feature windows. The negative class in such a classification framework directly corresponds to the claimed non-tremor data, where Mauldin’s explicitly identified false-positive samples correspond to the negative class used for retraining, in addition to Mikos’s “no FoG” labeling. Although Benbasat describes constructing a classifier from annotated data, a person of ordinary skill in the art would understand that adding new labeled feature windows to the training data stream and reconstructing the classifier is a routine implementation of such annotated-data training pipelines, and that performing such updates in response to newly labeled events falls within ordinary design choices for embedded decision tree classifiers.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Mikos, Mauldin, and Benbasat so that physiological data that preceded the user input is flagged as non-tremor data and used to update the decision tree classifier. The combination would have been possible because Rosenbluth already teaches ongoing acquisition of physiological data associated with tremor determinations, Mikos teaches that user input may label physiological data in real time, Mauldin teaches that false-positive samples are explicitly collected and used for retraining in wearable systems, and Benbasat teaches that annotated data are used to train and update classifiers. It would have been obvious to combine these teachings so that user-identified false-positive tremor detections are converted into annotated non-tremor training examples for classifier updating. The combination would have been feasible because all systems operate on time-windowed feature streams derived from continuous sensor data, as evidenced by Mauldin’s window-based processing and retraining pipeline. The benefit of this combination would be improved classifier accuracy through incorporation of user-verified non-tremor data.
Also regarding claim 11, the modified Rosenbluth does not fully teach receiving additional physiological data from the one or more sensors; and analyze the additional physiological data locally on the treatment device by applying the updated decision tree classifier. Rather, the modified Rosenbluth teaches ongoing sensing and local analysis of physiological data on the treatment device. For example, Rosenbluth teaches continuous sensing of motion and tremor characteristics (Rosenbluth, ¶[0018], ¶[0193]) and that the processor “can also receive information from the sensors 780 and process that information on board and adjust the stimulation accordingly” (Rosenbluth, ¶[0229]). This shows that the modified Rosenbluth teaches receiving additional physiological data and analyzing that data locally on the treatment device, but it does not explicitly disclose reapplying an updated decision tree classifier to the subsequently received data.
Benbasat teaches applying a classifier constructed from annotated data to incoming sensor data to determine system state. For example, Benbasat teaches that annotated examples are used “to construct a classifier that will determine the system state” (Benbasat, Sec. 2).
Mauldin further reinforces that retrained models are applied in deployed wearable systems and achieve improved performance after retraining with user feedback, stating that “the final Ensemble RNN model, after re-training with real-world user archived data and feedback, achieved a significantly higher precision without reducing much of the recall in a real-world setting” (Mauldin, Abstract / p. 2). Mauldin also teaches that “the sensed data from the smartwatch can be stored locally… in close proximity to the program that processes and analyzes the data in real-time” (Mauldin, p. 4-5), demonstrating that updated models are applied to incoming data in a deployed wearable system. Accordingly, such classifier pipelines support iterative application to newly received data following updates. In the combined system, once the classifier is updated using user-labeled non-tremor data (as above), the continuous sensing pipeline of Rosenbluth would naturally apply the updated classifier to subsequent incoming data.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Benbasat and Mauldin so that additional physiological data received from the sensors are analyzed locally on the treatment device by applying the updated decision tree classifier. The combination would have been possible because Rosenbluth already teaches continuous local sensing and analysis of physiological data and Benbasat teaches application of a trained classifier to incoming sensor data. Mauldin further teaches that retrained models are used in deployed wearable systems and improve performance after retraining with user feedback. It would have been obvious to combine these teachings so that once the decision tree classifier is updated using annotated user feedback, the updated classifier is used for subsequent local tremor analysis. The combination would have been feasible because both systems process continuous data streams and support iterative application of classifiers to incoming data, as evidenced by Mauldin’s deployed, real-time processing of updated models. The benefit of this combination would be adaptive real-time classification using updated models.
Regarding Claim 13, the modified Rosenbluth teaches that modulating the delivery comprises initiating the delivery of the vibrational energy via the vibration actuator (Rosenbluth, ¶[0221]: "If a tremor is detected... stimulation can be turned on 2210."; demonstrating the initiation of vibrational energy delivery upon tremor detection).
Regarding Claim 14, the modified Rosenbluth teaches that modulating delivery comprises at least one of: initiating the delivery of the vibrational energy via the vibration actuator, ceasing the delivery of the vibrational energy via the vibration actuator, varying an intensity of the vibrational energy delivered via the vibration actuator, or varying a frequency of the vibrational energy delivered via the vibration actuator (Rosenbluth, ¶[0043]: "The parameter may comprise at least one of stimulation frequency, amplitude, pulse width, pulse spacing, phase, waveform shape, waveform symmetry, duration, duty cycle, on/off time, or bursting; where the paragraph discusses varying parameters which is used to optimize therapy, the stimulation can be vibrational (Rosenbluth, ¶[0112]) and the amplitude equates to intensity).
Regarding Claim 15, the modified Rosenbluth teaches that the one or more sensors comprises at least one of: an accelerometer, a gyroscope, a temperature sensor, or a blood pressure sensor (Rosenbluth, ¶[0192]: “Sensors for monitoring the tremor may include… accelerometers, gyroscopes...” and ¶[0194]: “The device may also include… temperature sensors…”; explicitly disclosing the use of accelerometers, gyroscopes, and temperature sensors as well as other sensors).
Regarding Claim 16, the modified Rosenbluth teaches that the physiological data comprises an indication of a user tremor (Rosenbluth, ¶[0193]: "The data from these tremor sensors is used measure the patient's current and historical tremor characteristics such as the amplitude, frequency and phase"; describing tremor information as part of the physiological data from the sensors).
Claims 5, 6, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Rosenbluth et al. (US 20190001129 A1), hereto referred as Rosenbluth, and further in view of Benbasat et al. (Benbasat, A. Y. and Paradiso, J. A. “A framework for the automated generation of power-efficient classifiers for embedded sensor nodes”. 2007. In Proceedings of the 5th International Conference on Embedded Networked Sensor Systems (SenSys’07). ACM, New York, NY, 219–232), hereto referred as Benbasat, and further in view of Mikos et al. (Mikos, Val et al. “A Wearable, Patient-Adaptive Freezing of Gait Detection System for Biofeedback Cueing in Parkinson’s Disease.” IEEE transactions on biomedical circuits and systems 13.3 (2019): 503–515. Web), hereto referred as Mikos, and further in view of Mahadevan et al. (Mahadevan, N., Demanuele, C., Zhang, H. et al. Development of digital biomarkers for resting tremor and bradykinesia using a wrist-worn wearable device. npj Digit. Med. 3, 5 (2020)), hereto referred as Mahadevan, and further in view of Dai et al. (Dai, Houde, Pengyue Zhang, and Tim Lueth. “Quantitative Assessment of Parkinsonian Tremor Based on an Inertial Measurement Unit.” Sensors (Basel, Switzerland) 15.10 (2015)), hereto referred as Dai, and further in view of AN5259 (AN5259, ‘LSM6DSOX: Machine Learning Core’, Www.manuals.plus/m/36a1b1699dbd7d4f858bd13dc3517e0c252fd2307b38913eb4ec7f7d8b7deec4.pdf, 2019, Accessed 4/6/2026), hereto referred as AN5259, and further in view of Shoaran et al. (US 20200388397 A1), hereto referred as Shoaran.
The modified Rosenbluth teaches claims 1 and 17 as described above.
Regarding Claim 5, the modified Rosenbluth does not fully teach that the controller is an ultra-low power controller configured to analyze the physiological data to determine that the tremor is occurring using less than 1 milliwatt of power. Rosenbluth discloses a noninvasive treatment system with a controller configured to analyze physiological data to determine tremor occurrence and control stimulation accordingly (Rosenbluth, ¶[0193], ¶[0221]). Rosenbluth does not disclose an ultra-low power controller using less than 1 milliwatt of power for analyzation of the physiologic data for tremor detection. Shoaran, who investigates ultra-low power signal processing architectures for analyzing physiological data, demonstrates a system that detects tremors utilizing sub-microwatt level power and also analyzes the data with a classifier utilizing 41.2 nJ/class, clearing indicating a system that can detect tremors with less than a milliwatt of power (Shoaran, ¶[0182]). Rosenbluth’s controller analyzes physiological data, such as tremor occurrence, but it does so without considering the power efficiency constraints addressed by Shoaran. Shoaran’s system, which focuses on ultra-low power consumption, can be seamlessly integrated into Rosenbluth’s system to provide the same functionality while minimizing energy use. Its signal processing architecture is designed for low-power operation while detecting tremor episodes, making it a suitable candidate for adapting Rosenbluth’s tremor detection system. One of ordinary skill in the art would recognize that the addition of such low-power processing is a standard approach to reduce the overall energy consumption of wearable systems, without sacrificing the core functionality of tremor detection. It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Shoaran to have an ultra-low power controller configured to analyze the physiological data to determine that the tremor is occurring using less than 1 milliwatt of power. Shoaran’s architecture improves the practicality of Rosenbluth’s system by addressing a known limitation of wearable health devices (the need for low power consumption) by lowering the power usage of the device.
Regarding Claim 6, the modified Rosenbluth partially teaches that the controller is an ultra-low power controller configured to analyze the physiological data to determine that the tremor is occurring using less than 1 microwatt of power. Rosenbluth discloses a noninvasive treatment system with a controller configured to analyze physiological data to determine tremor occurrence and control stimulation accordingly (Rosenbluth, ¶[0193], ¶[0221]). Rosenbluth does not disclose an ultra-low power controller using less than 1 milliwatt of power for analyzation of the physiologic data for tremor detection. Shoaran, who investigates ultra-low power signal processing architectures for analyzing physiological data, demonstrates a system that detects tremors utilizing sub-microwatt level power and also analyzes the data with a classifier utilizing 41.2 nJ/class, indicating a system that can detect tremors with less than a microwatt of power (Shoaran, ¶[0182]). Rosenbluth’s controller analyzes physiological data, such as tremor occurrence, but it does so without considering the power efficiency constraints addressed by Shoaran. Shoaran’s system, which focuses on ultra-low power consumption, can be seamlessly integrated into Rosenbluth’s system to provide the same functionality while minimizing energy use. Its signal processing architecture is designed for low-power operation while detecting tremor episodes, making it a suitable candidate for adapting Rosenbluth’s tremor detection system. One of ordinary skill in the art would recognize that the addition of such low-power processing is a standard approach to reduce the overall energy consumption of wearable systems, without sacrificing the core functionality of tremor detection. It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Shoaran to have an ultra-low power controller configured to analyze the physiological data to determine that the tremor is occurring using less than 1 microwatt of power. Shoaran’s architecture improves the practicality of Rosenbluth’s system by addressing a known limitation of wearable health devices (the need for low power consumption) by lowering the power usage of the device.
Regarding Claim 19, the modified Rosenbluth does not teach that the controller consumes less than 1 milliwatt of power in analyzing the physiological data. Rosenbluth discloses a noninvasive treatment system with a controller configured to analyze physiological data to determine tremor occurrence and control stimulation accordingly (Rosenbluth, ¶[0193], ¶[0221]). Rosenbluth does not disclose an ultra-low power controller using less than 1 milliwatt of power for analyzation of the physiologic data for tremor detection. Shoaran, who investigates ultra-low power signal processing architectures for analyzing physiological data, demonstrates a system that detects tremors utilizing sub-microwatt level power and also analyzes the data with a classifier utilizing 41.2 nJ/class, clearing indicating a system that can detect tremors with less than a milliwatt of power (Shoaran, ¶[0182]). Rosenbluth’s controller analyzes physiological data, such as tremor occurrence, but it does so without considering the power efficiency constraints addressed by Shoaran. Shoaran’s system, which focuses on ultra-low power consumption, can be seamlessly integrated into Rosenbluth’s system to provide the same functionality while minimizing energy use. Its signal processing architecture is designed for low-power operation while detecting tremor episodes, making it a suitable candidate for adapting Rosenbluth’s tremor detection system. One of ordinary skill in the art would recognize that the addition of such low-power processing is a standard approach to reduce the overall energy consumption of wearable systems, without sacrificing the core functionality of tremor detection. It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Shoaran to have an ultra-low power controller configured to analyze the physiological data to determine that the tremor is occurring using less than 1 milliwatt of power. Shoaran’s architecture improves the practicality of Rosenbluth’s system by addressing a known limitation of wearable health devices (the need for low power consumption) by lowering the power usage of the device.
Claim 21 is rejected under 35 U.S.C. 103 as being unpatentable over Rosenbluth et al. (US 20190001129 A1), hereto referred as Rosenbluth, and further in view of Benbasat et al. (Benbasat, A. Y. and Paradiso, J. A. “A framework for the automated generation of power-efficient classifiers for embedded sensor nodes”. 2007. In Proceedings of the 5th International Conference on Embedded Networked Sensor Systems (SenSys’07). ACM, New York, NY, 219–232), hereto referred as Benbasat, and further in view of Mikos et al. (Mikos, Val et al. “A Wearable, Patient-Adaptive Freezing of Gait Detection System for Biofeedback Cueing in Parkinson’s Disease.” IEEE transactions on biomedical circuits and systems 13.3 (2019): 503–515. Web), hereto referred as Mikos, and further in view of Mahadevan et al. (Mahadevan, N., Demanuele, C., Zhang, H. et al. Development of digital biomarkers for resting tremor and bradykinesia using a wrist-worn wearable device. npj Digit. Med. 3, 5 (2020)), hereto referred as Mahadevan, and further in view of Dai et al. (Dai, Houde, Pengyue Zhang, and Tim Lueth. “Quantitative Assessment of Parkinsonian Tremor Based on an Inertial Measurement Unit.” Sensors (Basel, Switzerland) 15.10 (2015)), hereto referred as Dai, and further in view of AN5259 (AN5259, ‘LSM6DSOX: Machine Learning Core’, www.manuals.plus/m/36a1b1699dbd7d4f858bd13dc3517e0c252fd2307b38913eb4ec7f7d8b7deec4.pdf, 2019, Accessed 4/6/2026), hereto referred as AN5259, and further in view of Mauldin et al. (Mauldin, Taylor et al. “Ensemble Deep Learning on Wearables Using Small Datasets.” ACM transactions on computing for healthcare 2.1 (2021), hereto referred as Mauldin.
The modified Rosenbluth teaches claims 1 as described above.
Regarding Claim 21, the modified Rosenbluth teaches a noninvasive treatment system of claim 1 (see rejection of claim 1 above), wherein each of the treatment devices further comprises an input mechanism (Rosenbluth, ¶[0229]: “the processor 797... could receive input from the user via the controls module 740 and could control the execution of stimulation as selected by the user”; Fig. 7A: showing processor 797 connected with controls 740; this teaches that each device includes a user input mechanism integrated into the device), and wherein the controller is further configured to: responsive to a user input received via the input mechanism indicating a false-positive tremor detection, cease the delivery of the vibrational energy (Rosenbluth, ¶[0222]: “The stimulation may be triggered on or off in response to inputs including but not limited to user input...”; further teaching that user-triggered inputs such as voice activation or muscle signals may turn off stimulation; this teaches cessation of vibrational energy in response to user input), but does not fully teach that the user input specifically indicates a false-positive tremor detection. Rather, Rosenbluth teaches user-triggered control of stimulation and adaptive algorithms responsive to user input (¶[0222]; ¶[0229]), but does not explicitly disclose that such input reflects that a detected tremor event was incorrectly classified. Mikos teaches real-time user-driven labeling of detected events, where “the correct label (FoG or no FoG) of the feature vector is known and supplied... or... provided in real-time... through a BLE enabled push button” (Mikos, Sec. IV.C.1). This teaches an input mechanism through which a user indicates classification correctness.
Mauldin further teaches explicit user identification of false-positive events and their use in improving wearable detection systems, stating that “the low precision can be mitigated via the collection of false-positive feedback by the end-users” (Mauldin, Abstract) and that users “provide feedback on each generated false positive, and use those samples to re-train the model” (Mauldin, p. 3). This directly teaches user input that specifically identifies false-positive detections.
Although Mikos describes real-time labeling during activity, such input reasonably reflects the user’s assessment of the immediately preceding detected event, as further evidenced by Mauldin’s explicit teaching that user feedback is tied to false-positive detections, such that the input corresponds to a retroactive indication that the prior classification was incorrect. In a closed-loop stimulation system such as Rosenbluth, user activation to stop stimulation when no tremor is perceived inherently corresponds to a determination that the immediately preceding detection was incorrect. Mauldin reinforces this interpretation by explicitly tying user feedback to falsely detected events in deployed wearable systems.
It would have been prima facie obvious before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of Mikos and Mauldin so that cessation of vibrational energy is responsive to a user input indicating a false-positive tremor detection. The combination would have been possible because Rosenbluth already teaches user-triggered cessation of stimulation, Mikos teaches user input indicating classification correctness, and Mauldin teaches explicit user feedback identifying false-positive detections and using such feedback in wearable systems. It would have been obvious to refine the meaning of Rosenbluth’s existing user input mechanism to represent a false-positive detection rather than introducing a new control paradigm. The benefit would be reducing unnecessary stimulation and improving treatment accuracy through user-identified false positives.
Also regarding claim 21, the modified Rosenbluth does not fully teach update the decision tree algorithm based on the physiological data that preceded the user input, wherein the physiological data is flagged as non-tremor data for updating of the decision tree algorithm. Rather, Rosenbluth teaches continuous sensing and adaptive control based on physiological data (Rosenbluth, ¶[0193]; ¶[0229]) and that stimulation algorithms may be adaptive or self-calibrating (¶[0222]), but does not explicitly disclose labeling preceding physiological data as non-tremor data and using that data to update a decision tree algorithm.
Mikos teaches that user input labels detected events in real time (Mikos, Sec. IV.C.1). In a time-windowed classification system, the feature vector that produces a detection corresponds to a defined window of preceding physiological data, such that labeling the detected event inherently labels that preceding data window.
Mauldin further teaches that wearable sensor data are processed using window-based segmentation and that false-positive samples identified by the user are collected and used to retrain the model (Mauldin), stating that accelerometer data are processed using “a fixed-size sliding window approach” and that “false-positive data samples [are] collect[ed] and mark[ed]... to re-train the original model” (Mauldin, p. 2-3). This teaches that the specific data window associated with a detected event (i.e., the preceding physiological data) is labeled and reused for training.
Benbasat teaches that annotated data streams are used to construct classifiers (Benbasat, Sec. 2: “A training data stream is collected... annotated... used to construct a classifier...”), where labeled feature windows form the training data. The negative class in such a framework corresponds to non-event data, i.e., non-tremor data, where Mauldin’s explicitly identified false-positive samples correspond to the negative class used for retraining, in addition to Mikos’s “no FoG” labeling. Although Benbasat describes constructing a classifier from annotated data, a person of ordinary skill in the art would understand that adding new labeled feature windows to the training data stream and reconstructing the classifier is a routine implementation of such annotated-data pipelines.
It would have been prima facie obvious before the effective filing date of the claimed invention to have modified the modified Rosenbluth in view of Mikos, Mauldin, and Benbasat so that physiological data preceding a user-identified false-positive detection is flagged as non-tremor data and used to update the decision tree algorithm. The combination would have been possible because Rosenbluth already teaches continuous acquisition of physiological data, Mikos teaches user-driven labeling corresponding to detected events (and thus their associated data windows), Mauldin teaches that false-positive samples are explicitly collected and used for retraining in wearable systems, and Benbasat teaches training classifiers using annotated data streams. It would have been obvious to incorporate user-identified false positives as labeled non-tremor training examples and update the classifier accordingly. The benefit would be improved classifier accuracy through incorporation of user-verified non-tremor data tied to the specific detected event.
Also regarding claim 21, the modified Rosenbluth does not fully teach receive additional physiological data from the one or more sensors; and analyze the additional physiological data by applying the updated decision tree algorithm. Rather, the modified Rosenbluth teaches ongoing sensing and local analysis of physiological data on the treatment device. For example, Rosenbluth teaches continuous sensing of motion and tremor characteristics (Rosenbluth, ¶[0018], ¶[0193]) and that the processor “can also receive information from the sensors 780 and process that information on board and adjust the stimulation accordingly” (Rosenbluth, ¶[0229]). This shows that the modified Rosenbluth teaches receiving additional physiological data and analyzing that data locally on the treatment device, but it does not explicitly disclose reapplying an updated decision tree classifier to the subsequently received data.
Benbasat teaches applying a classifier constructed from annotated data to incoming sensor data to determine system state. For example, Benbasat teaches that annotated examples are used “to construct a classifier that will determine the system state” (Benbasat, Sec. 2).
Mauldin further reinforces that retrained models are applied in deployed wearable systems and achieve improved performance after retraining with user feedback, stating that “the final Ensemble RNN model, after re-training with real-world user archived data and feedback, achieved a significantly higher precision without reducing much of the recall in a real-world setting” (Mauldin, Abstract / p. 2). Mauldin also teaches that “the sensed data from the smartwatch can be stored locally… in close proximity to the program that processes and analyzes the data in real-time” (Mauldin, p. 4-5), demonstrating that updated models are applied to incoming data in a deployed wearable system. Accordingly, such classifier pipelines support iterative application to newly received data following updates. In the combined system, once the classifier is updated using user-labeled non-tremor data (as above), the continuous sensing pipeline of Rosenbluth would naturally apply the updated classifier to subsequent incoming data.
It would have been prima facie obvious before the effective filing date of the claimed invention to have further modified the modified Rosenbluth in view of Benbasat and Mauldin so that additional physiological data received from the sensors are analyzed locally on the treatment device by applying the updated decision tree classifier. The combination would have been possible because Rosenbluth already teaches continuous local sensing and analysis of physiological data and Benbasat teaches application of a trained classifier to incoming sensor data. Mauldin further teaches that retrained models are used in deployed wearable systems and improve performance after retraining with user feedback, stating that “the final Ensemble RNN model, after re-training with real-world user archived data and feedback, achieved a significantly higher precision without reducing much of the recall in a real-world setting” (Mauldin, Abstract / p. 2), and that “the sensed data from the smartwatch can be stored locally… in close proximity to the program that processes and analyzes the data in real-time” (Mauldin, p. 4–5). It would have been obvious to combine these teachings so that once the decision tree classifier is updated using annotated user feedback, the updated classifier is used for subsequent local tremor analysis. The combination would have been feasible because both systems process continuous data streams and support iterative application of classifiers to incoming data, as evidenced by Mauldin’s deployed, real-time processing of updated models. The benefit of this combination would be adaptive real-time classification using updated models.
Response to Arguments
Objections
Applicant's arguments filed 3/13/2026, page 10, regarding the previous Objections of claims 1, 11, and 17 have been fully considered and are persuasive. The previous Objections of claims 1, 11, and 17 have been withdrawn. However, the objection to claim 22 remains as shown above.
35 U.S.C. §103
Applicant's arguments filed 3/13/2026, pages 10-11, regarding the previous 103 Rejections of claims 1, 3-4, 7-11, 13-18, and 20-22 have been fully considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. That is, there are new grounds of rejection.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the
examiner should be directed to AARON MERRIAM whose telephone number is (703) 756-
5938. The examiner can normally be reached M-F 8:00 am - 5:00 pm.
Examiner interviews are available via telephone, in-person, and video conferencing
using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is
encouraged to use the USPTO Automated Interview Request (AIR) at
http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s
supervisor, Jason Sims can be reached on (571)272-4867. The fax phone number for the
organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained
from Patent Center. Unpublished application information in Patent Center is available to
registered users. To file and manage patent submissions in Patent Center, visit:
https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for
more information about Patent Center and https://www.uspto.gov/patents/docx for
information about filing in DOCX format. For additional questions, contact the Electronic
Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO
Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/AARON MERRIAM/Examiner, Art Unit 3791
/MATTHEW KREMER/Primary Examiner, Art Unit 3791
Prosecution Insights