Prosecution Insights
Last updated: November 15, 2026
Application No. 18/282,321

DEEP LEARNING-BASED HEART RATE DETECTION METHOD AND WEARABLE DEVICE

Final Rejection §101§103§112
Filed
Sep 15, 2023
Priority
Aug 11, 2021 — CN 202110921225.0 +1 more
Examiner
GROSS, JASON PATRICK
Art Unit
3797
Tech Center
3700 — Mechanical Engineering & Manufacturing
Assignee
Beijing Honor Device Co. Ltd.
OA Round
2 (Final)
64%
Grant Probability
Moderate
3-4
OA Rounds
99%
With Interview

Examiner Intelligence

Grants 64% of resolved cases
64%
Career Allowance Rate
16 granted / 25 resolved
-6.0% vs TC avg
Strong +43% interview lift
Without
With
+43.3%
Interview Lift
resolved cases with interview
Typical timeline
Avg Prosecution
32 currently pending
Career history
68
Total Applications
across all art units

Office Action

§101 §103 §112
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 . Claim Objections Claims 1, 4, 9, 26, and 31 are objected to because of the following informalities: Claims 1, 4, 9, 26, and 31 recite “deep attention Attention network.” There does not appear to be any difference between a “deep attention network” and a “deep attention Attention network.” Please remove the redundant use of “attention.” Claim 26 should read “A wearable device, comprising a processor, wherein the processor is coupled to a memory, and the processor is configured to execute a computer program or instructions stored in the memory, so that the wearable device implements a method comprising:…” Claim 31 depends from claim 26, which is directed to a wearable device. As such, claim 31 should read “31. The wearable device according to claim 26, wherein the obtaining heart rate….” Appropriate correction is required. Claim Interpretation Claims 1, 2, 11, 26 and 29 recite the newly amended limitation of “when a user wears a wearable device, obtaining a photoplethysmography” which in an interpretation it may be construed as a conditional limitation where the conditional limitations may not be given a full weight in light of the below decisions as for considering the other case scenario of “when” wearable device not being worn which the claim would not require this limitation to be a positive recitation. Claims 2, 11, 26 and 29 also recite similar limitations. In the recent Ex parte Gopalan decision, the PTAB addressed a claim where all of the features were recited in a conditional manner. A first step of “identifying … an outlier” was performed if “traffic is outside of a prediction interval.” A second step of “identifying” was performed “only when a count of outliers … is greater than or equal to two, and exceeds an anomaly threshold.” These were the only two elements of the independent claim. Thus, if the traffic is never outside Gopalan’s prediction interval, then the steps of the method are never performed. However, the PTAB distinguished Schulhauser and noted that this construction “would render the entire claim meaningless.” Gopalan at p. 5. The Board went on to state, “Although each of these steps is conditional, they are integrated into one method or path and do not cause the claim to diverge into two methods or paths, as in Schulhauser. Thus, we conclude that the broadest reasonable interpretation of claim 1 requires the performance of both steps…” Id. at p. 6.” Claim Rejections - 35 USC § 112 The following is a quotation of 35 U.S.C. 112(b): (b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention. The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph: The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention. Claims 1-8, 26, and 29-31 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention. Claims 1 and 26 each recite “obtaining heart rate information and scenario information based on [a] ACC effective spectrum data corresponding to the ACC signal, [b] PPG effective spectrum data corresponding to the PPG signal, and [c] a target prediction model...”. Presently, claim 1 only requires that the heart rate information and scenario information are obtained “based on” [a] and [b] and [c]. According to the application, the target prediction model can either be a single prediction model or a combination of prediction models. In either case, [a] and [b] would be inputs for [c] (i.e., the model or models). For example, the target prediction model can be a “first prediction model” such that the method includes “inputting the ACC effective spectrum data and the PPG effective spectrum data to a first prediction model to obtain the heart rate information and the scenario information” [0024]-[0026]). Alternatively, the target prediction model can be a “second prediction model” and “third prediction model” that receive the ACC effective spectrum data and the PPG effective spectrum data, respectively. ([0029], [0030]). The third prediction model would also receive scenario information. As such, it is not clear to one having ordinary skill in the art if claim 1 requires that [a] and [b] are input into [c] or if [a] and [b] and [c] are all different factors in the final determination of the heart rate information and the scenario information. “If the language of the claim is such that a person of ordinary skill in the art could not interpret the metes and bounds of the claim so as to understand how to avoid infringement, a rejection of the claim under 35 U.S.C. 112(b) or pre-AIA 35 U.S.C. 112, second paragraph, is appropriate.” (MPEP 2173.02). Examiner is interpreting claim 1 based on [0024]-[0026] and [0029], [0030], as discussed above. Specifically, claim 1 is being interpreted as “obtaining heart rate information and scenario information by inputting ACC effective spectrum data corresponding to the ACC signal and PPG effective spectrum data corresponding to the PPG signal into a target prediction model....” The scope of “target prediction model” includes each of the described embodiments (i.e., a single prediction model or a combination of prediction models). Because dependent claims 2-8 depend directly or indirectly from claim 1 and dependent claims 29-31 depend directly or indirectly from claim 26, claims 2-8 and 29-31 are also rejected for being indefinite. Claim Rejections - 35 USC § 101 35 U.S.C. 101 reads as follows: Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title. Claims 1-16, 26, and 29-31 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The claims recite: [a] obtaining heart rate information and scenario information based on ACC effective spectrum data corresponding to the ACC signal, PPG effective spectrum data corresponding to the PPG signal, and a target prediction model, wherein the target prediction model is a model that is obtained by training a deep attention network by using ACC sample data and PPG sample data as inputs and using a heart rate label and an exercise scenario label as target variables; (claims 1 and 26) [b] extracting ACC sample data, PPG sample data, and a heart rate label from the multi-scenario sample set; (claim 9) [c] training a deep attention network by using the ACC sample data and the PPG sample data as inputs and using the heart rate label and an exercise scenario label as target variables. (claim 9) Claim limitation [a], as drafted and under its broadest reasonable interpretation, recites a mental process. Claim limitation [a] recites a mental process because it involves “observations, evaluations, judgments, and opinions” with or without a physical aid. (MPEP 2106.04(a)(2), III). In this case, the prediction model is essentially performing signal analysis of PPG signal data and motion signal data to determine a heart rate and a type of activity that the user is currently performing. (see the recently decided Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437, Federal Circuit, 18 April 2025: “[C]laims that do no more than apply established methods of machine learning to a new data environment” are not patent eligible.) Claim limitation [b], as drafted and under its broadest reasonable interpretation, recites a mental process. Claim limitation [b] recites a mathematical concept because it involves mathematical calculations. (MPEP 2106.04(a)(2), I, C). The MPEP provides several examples of mathematical calculations, which include: “performing a resampled statistical analysis to generate a resampled distribution.” In this case, the mathematical calculations include applying the fast fourier transform (FFT) and then converting to a magnitude or power spectrum, filtering, and calculating certain relationships (e.g., amplitude ratio, peak-to-peak difference). Claim limitation [c], as drafted and under its broadest reasonable interpretation, recites a mental process. Claim limitation [c] recites a mathematical concept because it involves mathematical calculations. (MPEP 2106.04(a)(2), I, C). The MPEP provides several examples of mathematical calculations, which include: “performing a resampled statistical analysis to generate a resampled distribution.” In this case, training a machine-learning model involves transforming windows of raw samples into feature vectors by calculating mean, variance, etc. and transforming to a frequency-domain. Moreover, for each training example, loss functions are typically calculated. (see also the recently decided Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437, Federal Circuit, 18 April 2025: “[C]laims that do no more than apply established methods of machine learning to a new data environment” are not patent eligible.) This judicial exception is not integrated into a practical application. Additional elements to consider include: (1) the judicial exception is implemented using a wearable device; (2) that the heart rate information and the scenario information are displayed on a screen; (3) that the data includes PPG spectrum data and ACC spectrum data; (4) that the training is based on a multi-scenario sample set based on a plurality of exercise scenarios; and (5) a processor coupled to a memory and configured to execute a computer program or instructions stored in the memory. With respect to (1) use of a wearable device for acquiring and analyzing the signal data does no more than generally link a judicial exception to a particular technological environment. (MPEP 2106.04(d), I) (see, e.g., Affinity Labs of Texas v. DirecTV, LLC, 838 F.3d 1253 (Fed. Cir. 2016) in which the abstract idea of providing out-of-region access to regional broadcast content was not rendered patent-eligible simply because the technological environment was limited to cellular telephones (as opposed to any and all electronic devices)). With respect to (2), displaying the heart rate information and the scenario information is an extra-solution activity (i.e., the post-solution activity of displaying the results). (MPEP 2106.04(d), I). Likewise, using data that includes PPG spectrum data and ACC spectrum data as recited in (3) does no more than generally link a judicial exception to a particular technological environment. (MPEP 2106.04(d), I). With respect to (4), obtaining training data is a necessary step in training a machine-learning model and is an extra-solution activity (i.e., the pre-solution activity of obtaining data). With respect to (5), the processor and memory are recited at a high-level of generality. “Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception, e.g., a claim to an abstract idea requiring no more than a generic computer to perform generic computer functions that are well-understood, routine and conventional activities previously known to the industry….” (MPEP 2106.03, III, A). The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As explained above, the additional elements either recite extra-solution activity, do nothing more than generally link a judicial exception to a particular technological environment, or recite generic computer elements. Moreover, using PPG spectrum data and ACC spectrum data are well-understood, routine, and conventional activities. (see, e.g., Section 103 rejections discussed below). These are not meaningful limitations that transform the exception into a patent-eligible application, such that the claim does not amount to significantly more than the exception itself. The dependent claims do not render the claimed invention patent eligible. With respect to claims 2 and 29, the step of recording, in a data buffer, spectrum peak-point data is post-solution activity. Claims 3 and 30 recite generic steps of building a model by associating specific features with certain outputs. Claims 4 and 30 recite patent-ineligible subject matter (e.g., a prediction model) for similar reasons as discussed above, particularly with respect to claims 1 and 26. Claim 5 recites that the decision-making process of the trained model is separated into different stages. However, each of the separate stages corresponds to a separate trained model and, as such, recites an abstract idea for similar reasons as discussed above. Claims 6 and 14 recite well-understood, routine, and conventional pre-processing steps as discussed below in the Section 103 rejection. Claim 7 recites generic components for acquiring the PPG signal data and the ACC signal data. They are not meaningful limitations that transform the exception into a patent-eligible application. Claim 8 recites a well-understood, routine, and conventional operation in which a wearable device is wirelessly connected to another computing device (e.g., personal computer) as discussed below in the Section 103 rejection. Claim 10 recites a well-understood, routine, and conventional concept of cross-validation as discussed below in the Section 103 rejection. (see Section 103 rejection below). Claims 11, 12, and 14 recite the well-understood, routine, and conventional processes for obtaining training data as discussed below in the Section 103 rejection. Claim 13 merely recites that the same heart rate data is obtained from another device, which is a redundant operation that is a well-understood, routine, and conventional process. (see Section 103 rejection below). Claim 15 recites a well-understood, routine, and conventional pre-process step of filtering signal data. (see Section 103 rejection below). Accordingly, claims 1-16, 26, and 29-31 do not recite patent eligible subject matter as the claims recite an abstract idea without significantly more. Claim Rejections - 35 USC § 103 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, 4-7, 9, 26, and 31 rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Appl. Publ. No. 2016/0058367 A1 (hereinafter “RAGHURAM”) and U.S. Patent Appl. Publ. No. 2021/0093917 A1 (hereinafter “DERVISOGLU”) and U.S. Patent Appl. Publ. No. 2019/0328243 A1 (hereinafter “NEMATI”). RAGHURAM teaches a device that can estimate the heart rate of an active user by using a physiological model to refine a measurement of the user's heart rate obtained using a PPG sensor. (Abstract and [0008]). The device may be a wearable device, such as a smartwatch or watch-like device as shown in Figure 2A. A touch-screen display 210 can display a user’s heart rate. ([0024]). “Other types of information can also be presented, such as duration of an activity (e.g., a workout), distance traveled or steps taken during the activity, calories burned, and/or other parameters related to the activity.” ([0024]). Regarding claim 1, RAGHURAM teaches a heart rate detection method (Abstract), comprising: when a user wears a wearable device ([0022]: “user 100 wearing a wearable device 102”), obtaining a PNG media_image1.png 770 629 media_image1.png Greyscale PNG media_image2.png 1004 394 media_image2.png Greyscale photoplethysmography (PPG) signal ([0033]: “System 400 can include PPG sensor 402,….”; see also [0030] regarding “PPG signal”) and an acceleration ACC signal ([0033]: “System 400 can include PPG sensor 402, motion sensor 404,….”; see also [0034]: “motion sensor 404 can include an accelerometer, a gyroscopic sensor, or the like.”) in response to a heart rate detection command ([0036], [0046]); obtaining heart rate information ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”) and scenario information ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) based on ACC effective spectrum data corresponding to the ACC signal, PPG effective spectrum data corresponding to the PPG signal (both the PPG signal and the ACC signal are filtered to attenuate unwanted signal, see e.g., [0048], and then transformed into the frequency space, see e.g., [0049]), and a target prediction model ([0039] in which the “final estimator 416” determines a final heart rate estimate that is based on PPG signal and the motion signals)); and displaying the heart rate information on a screen of the wearable device (Figure 2A and [0024]), wherein the target prediction model is a model that is obtained by training a model by using ACC sample data and PPG sample data as inputs ([0037], [0039]) and using a heart rate label and an exercise scenario label as target variables (see, e.g., [0037] in which the “activity and intensity classifier” uses derived heart rate data and motion data), and the target prediction model has a scenario identification function ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) and a heart rate prediction function ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”). While RAGHURAM does teach using a display to display information to the user ([0024]), RAGHURAM does not explicitly teach displaying the scenario information. In the same field of endeavor, DERVISOGLU teaches “wearable devices and techniques for detecting walking workouts. By accurately and promptly detecting the start of walking workouts activities and automatically distinguishing between walking workout and causal walking activities, the disclosure enables wearable devices to accurately calculate user performance information when users forget to start and/or stop recording walking workouts.” (Abstract). “[T]he wearable device 100 may use sensed and collected motion information to predict a user's activity. Examples of activities may include, but are not limited to walking, running, cycling, swimming, skiing, cardio machine activities, and the like.” ([0043]). DERVISOGLU also teaches that the device can notify the user of the detected activity and request that the user confirm the activity (e.g., confirm that a workout has begun). (see, e.g., [0068]). It would have been obvious to one having ordinary skill in the are to modify the method to include displaying the scenario information to the user. One would have been motivated to display the scenario information to the user, as taught in DERVISOGLU, so that the user would be notified that the wearable device has begun tracking a workout and/or so that the user could confirm that a certain activity has begun or ended. There would have been a reasonable expectation of success as RAGHURAM already detects the scenario information and DERVISOGLU demonstrates that watches can display such information. While RAGHURAM teaches using machine-learning models to analyze the data and classify the activity ([0037]), RAGHURAM does not explicitly teach that the target prediction model is obtained by training a deep attention network. In the same field of endeavor, NEMATI teaches “systems and methods can accurately and efficiently determine abnormal cardiac activity from motion data and/or cardiac data using techniques that can be used for long-term monitoring of a patient.” (Abstract). Monitoring devices may include wearable devices, such as a smart watch. “For example, the one sensor collection device 110 may be a smart watch including at least the cardiac and motion sensors that can be attached to an individual's wrist, for example, using a wrist band.” ([0044]). NEMATI teaches using a target prediction model that is obtained training a deep attention network. Specifically, NEMATI’s device is “configured for multi-class classification of cardiac activity using at least the motion data and the cardiac data.” ([0049]). NEMATI uses a “deep learning architecture” that may include “an attention network.” (Abstract). The attention network is used in a series of neural networks. “[T]he resulting networks (CNN, RNN and attention network) may then trained together (end-to-end), for example, training the attention network and RNN using the learning features extracted by the CNN, so as to optimize the classifier.” ([0086]). These networks are trained using “training samples of motion data and/or cardiac data having known cardiac activity features, such as known abnormal cardiac activity, known subject information (e.g., medical and/or demographic information (e.g., age, medication use, device use, etc.)), among others, or combination thereof.” ([0048]). “[T]he attention network can determine the optimal sections of data to analyze, or how to relatively weigh combinations of windows to provide a final diagnosis or burden. It can also identify the time(s) at which the cardiac activity occurred.” ([0098]). The attention network classifies cardiac activity for a period of time by weighting outputs from another network. ([[0099]). “This weighting, or ‘attention’ can [be] determined by optimization of the weights over the entire training set, using back propagation.” ([0099]). It would have been obvious to one having ordinary skill in the art to modify the method such that the target prediction model is obtained by training a deep attention network. One would be motivated to use a deep attention network, as taught in NEMATI, because such attention networks can determine the optimal sections of data to analyze or how to relatively weigh combinations of sections of data to provide the correct outputs. More specifically, the deep attention network would optimize weights to apply to sections of PPG data. There would be a reasonable expectation of success as NEMATI teaches that deep attention networks may be used to train models that are ultimately implemented by wearable devices that track bio-signals. With respect to claim 4, RAGHURAM teaches inputting the ACC effective spectrum data and the PPG effective spectrum data to a first prediction model to obtain the heart rate information ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”) and the scenario information ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”), wherein the target prediction model is the first prediction model, the first prediction model is a model that is obtained by training the deep attention network by using the ACC sample data and the PPG sample data as inputs ([0037], [0039]) and using the heart rate label and the exercise scenario label as target variables (see, e.g., [0037] in which the “activity and intensity classifier” uses derived heart rate data and motion data), and the target prediction model has a scenario identification function ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) and a heart rate prediction function ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”). NOTE: Other than re-naming the “target prediction model” to “first prediction model,” claim 4 does not recite any limitations that further distinguish the prediction model. With respect to claim 5, RAGHURAM teaches wherein the obtaining heart rate information and scenario information based on ACC effective spectrum data corresponding to the ACC signal, PPG effective spectrum data corresponding to the PPG signal, and a target prediction model comprises: inputting the ACC effective spectrum data to a second prediction model to obtain the scenario information, (see, e.g., Figure 4 of RAGHURAM in which the activity/intensity classifier provided with “motion signals” and outputs the activity, which is used indirectly by the final estimator 416.) wherein the second prediction model is a model that is obtained by training a deep neural network by using the ACC sample data as an input and using the exercise scenario label as a target variable (“Activity and intensity classifier 412 can implement machine learning algorithms that identify characteristic motion-sensor features associated with various types of activity, and the algorithms can be trained by gathering sensor data during known user activities prior to deploying activity and intensity classifier 412.” [0037]), and the second prediction model has the scenario identification function (“Activity and intensity classifier 412 can analyze the received data to determine the type of activity in which the user is most likely engaged, e.g., walking, sitting or standing still, jogging, running, cycling, riding in a motor vehicle and so on.” ([0037]); and inputting the PPG effective spectrum data and the obtained scenario information to a third prediction model to obtain the heart rate information (see Figure 4 and final estimate 416 that determines the heart rate information based on PPG signals and motion signals), and the third prediction model has the heart rate detection function (“Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).” ([0039]), wherein the target prediction model comprises the second prediction model and the third prediction model (the processing subsystem 406 includes both the activity/intensity classifier 412 and the final estimator 416). RAGHURAM does not explicitly teach wherein the third prediction model is a model that is obtained by training the deep attention network by using the PPG sample data, the ACC sample data, and the exercise scenario label as inputs and using the heart rate label as a target variable. However, as discussed above with respect to claim 1, NEMATI teaches using a target prediction model that is obtained training a deep attention network. Specifically, NEMATI’s device is “configured for multi-class classification of cardiac activity using at least the motion data and the cardiac data.” ([0049]). NEMATI uses a “deep learning architecture” that includes “an attention network.” (Abstract). The attention network is used in a series of neural networks. “[T]he resulting networks (CNN, RNN and attention network) may then trained together (end-to-end), for example, training the attention network and RNN using the learning features extracted by the CNN, so as to optimize the classifier.” ([0086]). These networks are trained using “training samples of motion data and/or cardiac data having known cardiac activity features, such as known abnormal cardiac activity, known subject information (e.g., medical and/or demographic information (e.g., age, medication use, device use, etc.)), among others, or combination thereof.” ([0048]). “[T]he attention network can determine the optimal sections of data to analyze, or how to relatively weigh combinations of windows to provide a final diagnosis or burden. It can also identify the time(s) at which the cardiac activity occurred.” ([0098]). The attention network classifies cardiac activity for a period of time by weighting outputs from another network. ([[0099]). “This weighting, or ‘attention’ can [be] determined by optimization of the weights over the entire training set, using back propagation.” ([0099]). It would have been obvious to one having ordinary skill in the art to modify the method such that the target prediction model is obtained by training a deep attention network using the PPG sample data, the ACC sample data, and the exercise scenario label as inputs and using the heart rate label as a target variable. One would be motivated to use a deep attention network, as taught in NEMATI, because such attention networks can determine the optimal sections of data to analyze or how to relatively weigh combinations of sections of data to provide the correct outputs. More specifically, the deep attention network would optimize weights to apply to sections of PPG data. There would be a reasonable expectation of success as NEMATI teaches that deep attention networks may be used to train models that are ultimately implemented by wearable devices that track bio-signals. With respect to claim 6, RAGHURAM teaches wherein after the obtaining a PPG signal and an ACC signal, the method further comprises: separately performing first preprocessing on the ACC signal and the PPG signal to obtain the ACC effective spectrum data and the PPG effective spectrum data, wherein the first preprocessing comprises fast Fourier transform FFT and filtering processing. (see, e.g., [0048]: “At blocks 508 and 509, the dark-channel-subtracted PPG signal and the motion-sensor signal can be separately bandpass filtered to reduce frequency content outside the range where most of the energy of a signal driven by a heart pumping would be expected to resid.”; and [0049]: “At blocks 510 and 512, the processed signals can be transformed to frequency space, e.g., using a discrete Fourier transform (DFT).”). With respect to claim 7, RAGHURAM teaches wherein the obtaining a PPG signal and an ACC signal comprises: collecting the PPG signal by using a PPG sensor (see “PPG sensor 300” in [0029] and [0030]), and collecting the ACC signal by using an acceleration sensor (see, e.g., [0034]: “motion sensor 404 can include an accelerometer, a gyroscopic sensor, or the like.”). With respect to claim 9, RAGHURAM teaches a method for training a model used for heart rate detection (see, e.g., [0037]: “Activity and intensity classifier 412 can implement machine learning algorithms that identify characteristic motion-sensor features associated with various types of activity, and the algorithms can be trained by gathering sensor data during known user activities prior to deploying activity and intensity classifier 412.”, comprising: obtaining a multi-scenario sample set, wherein the multi-scenario sample set is a data sample set obtained through detection based on a plurality of exercise scenarios (See [0037] above. Also, the “known user activities” described in [0037] would include the activities that the device is designed to identify, which include a plurality of exercise scenarios. See, e.g., [0009]: “The device can also determine an activity context for the user, such as the type of activity in which the user is engaged (e.g., walking, running, jogging, swimming, cycling, sitting, sleeping, etc.) and, where appropriate, an intensity associated with the activity.”; extracting ACC sample data, PPG sample data, and a heart rate label from the multi-scenario sample set. RAGHURAM teaches “gathering sensor data during known user activities” ([0037]). The sensors include a PPG sensor and a motion sensor.([0036]). As one purpose of RAGHURAM is to determine the heart rate based on the PPG and motion data, the sample set would necessarily include a heart rate label (i.e., measured heart rate during activities) (see also, e.g., [0020] in which the “direct” measurement of the user’s heart rate is estimated using PPG data and user’s current activity can also be based on “previous heart rate measurements.”) obtaining a first prediction model, wherein the first prediction model has a scenario identification function ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) and a heart rate prediction function ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”). While RAGHURAM generally teaches training a machine-learning algorithm by “gathering sensor data during known user activities” ([0037]) and that sensor data would include PPG data and ACC data (see, e.g., [0033]), RAGHURAM does not explicitly teach, training a deep attention network by using the ACC sample data and the PPG sample data as inputs and using the heart rate label and an exercise scenario label as target variables. In the same field of endeavor, NEMATI teaches “systems and methods can accurately and efficiently determine abnormal cardiac activity from motion data and/or cardiac data using techniques that can be used for long-term monitoring of a patient.” (Abstract). Monitoring devices may include wearable devices, such as a smart watch. “For example, the one sensor collection device 110 may be a smart watch including at least the cardiac and motion sensors that can be attached to an individual's wrist, for example, using a wrist band.” ([0044]). NEMATI teaches using a target prediction model that is obtained training a deep attention network. Specifically, NEMATI’s device is “configured for multi-class classification of cardiac activity using at least the motion data and the cardiac data.” ([0049]). NEMATI uses a “deep learning architecture” that may include “an attention network.” (Abstract). The attention network is used in a series of neural networks. “[T]he resulting networks (CNN, RNN and attention network) may then trained together (end-to-end), for example, training the attention network and RNN using the learning features extracted by the CNN, so as to optimize the classifier.” ([0086]). These networks are trained using “training samples of motion data and/or cardiac data having known cardiac activity features, such as known abnormal cardiac activity, known subject information (e.g., medical and/or demographic information (e.g., age, medication use, device use, etc.)), among others, or combination thereof.” ([0048]). “[T]he attention network can determine the optimal sections of data to analyze, or how to relatively weigh combinations of windows to provide a final diagnosis or burden. It can also identify the time(s) at which the cardiac activity occurred.” ([0098]). The attention network classifies cardiac activity for a period of time by weighting outputs from another network. ([[0099]). “This weighting, or ‘attention’ can [be] determined by optimization of the weights over the entire training set, using back propagation.” ([0099]). It would have been obvious to one having ordinary skill in the art to modify the method to train a deep attention network by using the ACC sample data and the PPG sample data as inputs and using the heart rate label and an exercise scenario label as target variables. In RAGHURAM, the ultimate goal is to identify the heart rate information and the scenario information using ACC data and PPG data. One would be motivated to train a deep attention network, as taught in NEMATI, because such attention networks can determine the optimal sections of the ACC and PPG sample data to analyze or how to relatively weigh combinations of sections of data to provide the correct outputs. More specifically, the deep attention network would optimize weights to apply to at least sections of PPG data. There would be a reasonable expectation of success as NEMATI teaches that deep attention networks may be trained using PPG and motion data. With respect to claim 26, RAGHURAM teaches a wearable device (see, e.g., smartwatch in Figure 2A), comprising a processor (“processing unit 406” in Figure 4 and “processing subsystem 904” in Figure 9), wherein the processor is coupled to a memory (see, e.g., [0079]), and the processor is configured to execute a computer program or instructions stored in the memory ([0079]), so that the wearable device implements the method comprising: when a user wears a wearable device ([0022]: “user 100 wearing a wearable device 102”), obtaining a photoplethysmography (PPG) signal ([0033]: “System 400 can include PPG sensor 402,….”; see also [0030] regarding “PPG signal”) and an acceleration ACC signal ([0033]: “System 400 can include PPG sensor 402, motion sensor 404,….”; see also [0034]: “motion sensor 404 can include an accelerometer, a gyroscopic sensor, or the like.”) in response to a heart rate detection command ([0036], [0046]); obtaining heart rate information ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”) and scenario information ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.) based on ACC effective spectrum data corresponding to the ACC signal, PPG effective spectrum data corresponding to the PPG signal (both the PPG signal and the ACC signal are filtered to attenuate unwanted signal, see e.g., [0048], and then transformed into the frequency space, see e.g., [0049]), and a target prediction model ([0039] in which the “final estimator 416” determines a final heart rate estimate that is based on PPG signal and the motion signals)); and displaying the heart rate information on a screen of the wearable device (Figure 2A and [0024]), wherein the target prediction model is a model that is obtained by training the model using ACC sample data and PPG sample data as inputs ([0037], [0039]) and using a heart rate label and an exercise scenario label as target variables (see, e.g., [0037] in which the “activity and intensity classifier” uses derived heart rate data and motion data), and the target prediction model has a scenario identification function ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) and a heart rate prediction function ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”). While RAGHURAM does teach using a display to display information to the user ([0024]), RAGHURAM does not explicitly teach displaying the scenario information. In the same field of endeavor, DERVISOGLU teaches “wearable devices and techniques for detecting walking workouts. By accurately and promptly detecting the start of walking workouts activities and automatically distinguishing between walking workout and causal walking activities, the disclosure enables wearable devices to accurately calculate user performance information when users forget to start and/or stop recording walking workouts.” (Abstract). “[T]he wearable device 100 may use sensed and collected motion information to predict a user's activity. Examples of activities may include, but are not limited to walking, running, cycling, swimming, skiing, cardio machine activities, and the like.” ([0043]). DERVISOGLU also teaches that the device can notify the user of the detected activity and request that the user confirm the activity (e.g., confirm that a workout has begun). (see, e.g., [0068]). It would have been obvious to one having ordinary skill in the art to modify the method to include displaying the scenario information to the user. One would have been motivated to display the scenario information to the user, as taught in DERVISOGLU, so that the user would be notified that the wearable device has begun tracking a workout and/or so that the user could confirm that a certain activity has begun or ended. There would have been a reasonable expectation of success as RAGHURAM already detects the scenario information and DERVISOGLU demonstrates that watches can display such information. While RAGHURAM teaches using machine-learning models to analyze the data and classify the activity ([0037]), RAGHURAM does not explicitly teach that the target prediction model is obtained by training a deep attention network. In the same field of endeavor, NEMATI teaches “systems and methods can accurately and efficiently determine abnormal cardiac activity from motion data and/or cardiac data using techniques that can be used for long-term monitoring of a patient.” (Abstract). Monitoring devices may include wearable devices, such as a smart watch. “For example, the one sensor collection device 110 may be a smart watch including at least the cardiac and motion sensors that can be attached to an individual's wrist, for example, using a wrist band.” ([0044]). NEMATI teaches using a target prediction model that is obtained training a deep attention network. Specifically, NEMATI’s device is “configured for multi-class classification of cardiac activity using at least the motion data and the cardiac data.” ([0049]). NEMATI uses a “deep learning architecture” that may include “an attention network.” (Abstract). The attention network is used in a series of neural networks. “[T]he resulting networks (CNN, RNN and attention network) may then trained together (end-to-end), for example, training the attention network and RNN using the learning features extracted by the CNN, so as to optimize the classifier.” ([0086]). These networks are trained using “training samples of motion data and/or cardiac data having known cardiac activity features, such as known abnormal cardiac activity, known subject information (e.g., medical and/or demographic information (e.g., age, medication use, device use, etc.)), among others, or combination thereof.” ([0048]). “[T]he attention network can determine the optimal sections of data to analyze, or how to relatively weigh combinations of windows to provide a final diagnosis or burden. It can also identify the time(s) at which the cardiac activity occurred.” ([0098]). The attention network classifies cardiac activity for a period of time by weighting outputs from another network. ([[0099]). “This weighting, or ‘attention’ can [be] determined by optimization of the weights over the entire training set, using back propagation.” ([0099]). It would have been obvious to one having ordinary skill in the art to modify the method such that the target prediction model is obtained by training a deep attention network. One would be motivated to use a deep attention network, as taught in NEMATI, because such attention networks can determine the optimal sections of data to analyze or how to relatively weigh combinations of sections of data to provide the correct outputs. More specifically, the deep attention network would optimize weights to apply to sections of PPG data. There would be a reasonable expectation of success as NEMATI teaches that deep attention networks may be used to train models that are ultimately implemented by wearable devices that track bio-signals. With respect to claim 31, RAGHURAM teaches that the obtaining heart rate information and scenario information based on ACC effective spectrum data corresponding to the ACC signal, PPG effective spectrum data corresponding to the PPG signal, and a target prediction model comprises inputting the ACC effective spectrum data and the PPG effective spectrum data to a first prediction model to obtain the heart rate information ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”) and the scenario information ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”), wherein the target prediction model is the first prediction model, the first prediction model is a model that is obtained by training the deep attention network by using the ACC sample data and the PPG sample data as inputs ([0037], [0039]) and using the heart rate label and the exercise scenario label as target variables (see, e.g., [0037] in which the “activity and intensity classifier” uses derived heart rate data and motion data), and the target prediction model has a scenario identification function ([0037]: “Activity and intensity classifier 412 can analyze inputs including signals from motion sensor 404 to determine the user's current activity and intensity of the activity.”) and a heart rate prediction function ([0039]: “Final estimator 416 can use both direct heart rate estimate HR_D and modeled heart rate estimate HR_M to determine a final heart rate estimate (HR_F).”). NOTE: Other than re-naming the “target prediction model” to “first prediction model,” claim 4 does not recite any limitations that further distinguish the prediction model. 1, 4-7, 9, 26, and 31 Claims 2, 3, 29, and 30 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Appl. Publ. No. 2016/0058367 A1 (hereinafter “RAGHURAM”) and U.S. Patent Appl. Publ. No. 2021/0093917 A1 (hereinafter “DERVISOGLU”) and U.S. Patent Appl. Publ. No. 2019/0328243 A1 (hereinafter “NEMATI”) as applied to claims 1 and 26 above, and further in view U.S. Patent Appl. Publ. No. 2024/0176425 A1 (hereinafter “WANG”) and Park K, Kim DH. Accelerating image classification using feature map similarity in convolutional neural networks. Applied Sciences. 2018 Dec 29;9(1):108 (hereinafter “PARK”). Claims 2, 3, 29, and 30 recite concepts that are commonly found in signal processing and machine-learning. In particular, these concepts include (a) associating signal data (i.e., inputs) with a certain result (i.e., label); (b) analyzing that signal data to extract features from the signal data; and (c) storing those extracted features with the result in order to predict future results. None of RAGHURAM, DERVISOGLU, or NEMATI explicitly teach each of these concepts as recited in claims 2 and 3. However, RAGHURAM generally refers to these concepts. (see, e.g., [0037]: “Activity and intensity classifier 412 can implement machine learning algorithms that identify characteristic motion-sensor features associated with various types of activity, and the algorithms can be trained by gathering sensor data during known user activities prior to deploying activity and intensity classifier 412.”). Likewise, DERVISOGLU teaches these concepts at a high level. (see, e.g., [0044]: “In various embodiments, one or more machine learning techniques and/or predictive models trained to predict the user's activity and/or detect activity start and end points. Training the one or more predictive models may include surveying a plurality of datasets including motion data, rotational data, heart rate data, and the like collected during activities having a known activity type (e.g., walking workouts, casual walking activities, running activities, and the like).”; see also [0045]: “For example, the step counter 308 may determine a step count based on motion features extracted from motion data and or rotational data that indicate steps. The motion features that indicate steps may be determined by surveying a plurality of datasets including motion data and or rotational data collected during known walking workouts. Motion features that are specific to a particular user may be determined by limiting the walking workouts included in the plurality of datasets to walking workouts performed by the particular user….”). In the same field of endeavor, WANG teaches a method for controlling a wearable device which includes detecting an abnormal touch event on a display screen of a wearable device and enabling gesture recognition in response to the abnormal touch event. (Abstract). With gesture recognition enabled, the wearable device is configured to detect a first user gesture and a different, second user gesture. ([0013] and [0014]). To detect these user gestures, WANG teaches training a separate machine-learning model for each user gesture. and using the separate trained model to recognize the user gesture. ([0120]-[0123]). The machine-learning models and the trained models are configured to receive a “feature vector” as an input. “The first feature vector is used as an input of a first AI model, and the first feature vector is processed by a trained first AI model, to output a specific gesture.” ([0122]). “The second feature vector is used as an input of a second AI model, and the second feature vector is processed by a trained second AI model, to output a specific gesture.” ([0123]). Notably, the feature vectors are based on PPG and ACC signal data. “To generate the first feature vector, the following processing may be performed on the foregoing obtained ACC data (that is, raw ACC data) and PPG data (that is, raw PPG data): (1) performing normalization processing on the raw ACC data and the raw PPG data to form normalized ACC data and normalized PPG data; (2) performing frequency-domain transformation on the raw ACC data and the raw PPG data respectively to form ACC frequency-domain data and PPG frequency domain data, (3) performing time-frequency transformation (wavelet transformation) on the raw ACC data and the raw PPG data to form ACC time-frequency data and PPG time-frequency data: and (4) performing correlation transformation between the raw PPG data and the raw ACC data to form ACC-PPG related data.” ([0125]-[0129]). In addition to the ACC-PPG related data, WANG teaches obtaining various waveforms of the ACC and PPG data and extracting features from those waveforms. “Feature information is extracted from each (each dimension of) data waveform, and several extracted features jointly form a group of feature vectors. The extracted feature data includes but is not limited to a feature of data, such as an average value, a standard deviation, a waveform factor, a peak-to-peak value, a gradient, skewness, waveform short-time energy, a short-time average amplitude difference, an energy entropy, an autocorrelation coefficient, an approximate entropy, an average absolute deviation, a cliff, a power spectral density, and spectral data. The feature data is spliced to form the first feature vector.” ([0130]). The feature vectors are used as training data to train a model to recognize certain gestures. “The training data may include a training feature vector of a specific gesture and corresponding training label information (for example, a finger is rubbed or a fist is held).” ([0138]). When the wearable device is in use, feature vectors are generated from data acquired during the gesture and those feature vectors are used as inputs to the trained model. “Then, the generated first feature vector and second feature vector may be respectively input into the trained first AI model and the trained second AI model.” ([0137]). To summarize, WANG teaches using feature vectors to train a model and then, when the wearable device is in use, generating those same features as inputs to recognize the gesture. Moreover, WANG teaches that the machine-learning models may include “a K-nearest neighbor algorithm,” which is well-known in the art and operates by comparing an input feature vector to a stored feature vector. With respect to claim 2, it would have been obvious to one having ordinary skill in the art to modify the method to record spectrum peak-point data corresponding to the heart rate information and the scenario information (i.e., the outputs), wherein the spectrum peak-point data corresponding to the heart rate information and the scenario information comprises a peak point location and an amplitude that are of a PPG spectrum and a peak point location and an amplitude that are of an ACC spectrum. One would have been motivated to train a model and use a trained model as taught in WANG in order to accurately determine the heart rate information and scenario information using PPG and ACC data. RAGHURAM and DERVISOGLU suggest using trained models and WANG teaches that the trained models may accurately identify the output/label using a feature vector derived from PPG and ACC signal data. There would have been a reasonable expectation of success as each of RAGHURAM, DERVISOGLU, and WANG teach that PPG and ACC data may be used to generate desired output from trained models. NOTE: Examiner is interpreting as “spectrum peak-point data” to include at least one of a peak-to-peak value, a gradient, skewness, a short-time average amplitude difference, a cliff or a power spectral density as these features necessarily comprise “a peak point location and an amplitude that are of a PPG spectrum and a peak point location and an amplitude that are of an ACC spectrum.” (see, e.g., [0130] of WANG). However, none of the prior art teaches recording the heart rate information and the scenario information in a label buffer and recording the spectrum peak-point data in a data buffer. PARK is concerned with the computational load sustained by low performance devices. “[T]he extensive time required for classification owing to the large amount of computation involved, makes it unsuitable for application to low-performance devices.” (Abstract). PARK teaches classifying input data based on similarities with previous input data. “In this study, we cache class labels and feature vectors extracted from feature maps for images classified by the CNN. Then, when a new image is input, its class label is output based on its similarity with the cached feature vectors. This process can be performed at each layer; hence, if the classification is successful, there is no need to perform the remaining convolution layer operations. This reduces the required classification time.” (emphasis added) (Abstract). Notably, the class labels are cached (i.e., added to a label buffer) and the feature vectors are cached (i.e., added to a data buffer). It would have been obvious to one having ordinary skill in the art to cache class labels and feature vectors, as taught in PARK. Like the devices that concern PARK, wearable devices are (relatively) low-performance devices that are challenged to classify new data within a practical amount of time. Yet wearable devices often receive data that is similar or identical to recently received data (e.g., in situations where the heart rate and/or activity is not changing) like the devices discussed in PARK. One would have been motivated to cache (i.e., buffer) the labels and corresponding feature vectors in order to reduce the classification time and computational load as taught in PARK. There would have been a reasonable expectation of success as buffering simply involves storing recently outputted labels and corresponding feature vectors. With respect to claim 3, WANG also teaches “[d]uring implementation, a person skilled in the art may first determine to-be-used AI models, and model parameters of these AI models are all default values. Therefore, training data needs to be used to determine model parameters applicable to embodiments of this application.” ([0137]). It would have been obvious to one having ordinary skill in the art to modify the method such that, upon obtaining first N pieces of peak-point spectrum data based on the PPG effective spectrum data and the ACC effective spectrum data (i.e., a feature vector), if no peak-point spectrum data in the first N pieces of peak-point spectrum data is prestored in the data buffer, obtaining the heart rate information and the scenario information based on the ACC effective spectrum data corresponding to the ACC signal, the PPG effective spectrum data corresponding to the PPG signal, and the target prediction model or if any peak-point spectrum data in the first N pieces of peak-point spectrum data is prestored in the data buffer, reading, from the label buffer, heart rate information and scenario information that correspond to the any peak-point spectrum data. One would have been motivated to make this modification because, as taught in WANG, it is necessary to build (i.e., train) the model by inputting peak-point spectrum data that the model may or may have not seen yet. There would have been a reasonable expectation of success as each of RAGHURAM, DERVISOGLU, and WANG teach that PPG and ACC data may be used to train a model. With respect to claim 29, as discussed above with respect to claim 2, RAGHURAM does not explicitly teach when the heart rate information and the scenario information are obtained, recording the heart rate information and the scenario information in a label buffer; and recording, in a data buffer, spectrum peak-point data corresponding to the heart rate information and the scenario information, wherein the spectrum peak-point data corresponding to the heart rate information and the scenario information comprises a peak point location and an amplitude that are of a PPG spectrum and a peak point location and an amplitude that are of an ACC spectrum. However, WANG teaches a method for controlling a wearable device which includes detecting an abnormal touch event on a display screen of a wearable device and enabling gesture recognition in response to the abnormal touch event. (Abstract). With gesture recognition enabled, the wearable device is configured to detect a first user gesture and a different, second user gesture. ([0013] and [0014]). To detect these user gestures, WANG teaches training a separate machine-learning model for each user gesture. and using the separate trained model to recognize the user gesture. ([0120]-[0123]). The machine-learning models and the trained models are configured to receive a “feature vector” as an input. “The first feature vector is used as an input of a first AI model, and the first feature vector is processed by a trained first AI model, to output a specific gesture.” ([0122]). “The second feature vector is used as an input of a second AI model, and the second feature vector is processed by a trained second AI model, to output a specific gesture.” ([0123]). Notably, the feature vectors are based on PPG and ACC signal data. “To generate the first feature vector, the following processing may be performed on the foregoing obtained ACC data (that is, raw ACC data) and PPG data (that is, raw PPG data): (1) performing normalization processing on the raw ACC data and the raw PPG data to form normalized ACC data and normalized PPG data; (2) performing frequency-domain transformation on the raw ACC data and the raw PPG data respectively to form ACC frequency-domain data and PPG frequency domain data, (3) performing time-frequency transformation (wavelet transformation) on the raw ACC data and the raw PPG data to form ACC time-frequency data and PPG time-frequency data: and (4) performing correlation transformation between the raw PPG data and the raw ACC data to form ACC-PPG related data.” ([0125]-[0129]). In addition to the ACC-PPG related data, WANG teaches obtaining various waveforms of the ACC and PPG data and extracting features from those waveforms. “Feature information is extracted from each (each dimension of) data waveform, and several extracted features jointly form a group of feature vectors. The extracted feature data includes but is not limited to a feature of data, such as an average value, a standard deviation, a waveform factor, a peak-to-peak value, a gradient, skewness, waveform short-time energy, a short-time average amplitude difference, an energy entropy, an autocorrelation coefficient, an approximate entropy, an average absolute deviation, a cliff, a power spectral density, and spectral data. The feature data is spliced to form the first feature vector.” ([0130]). The feature vectors are used as training data to train a model to recognize certain gestures. “The training data may include a training feature vector of a specific gesture and corresponding training label information (for example, a finger is rubbed or a fist is held).” ([0138]). When the wearable device is in use, feature vectors are generated from data acquired during the gesture and those feature vectors are used as inputs to the trained model. “Then, the generated first feature vector and second feature vector may be respectively input into the trained first AI model and the trained second AI model.” ([0137]). To summarize, WANG teaches using feature vectors to train a model and then, when the wearable device is in use, generating those same features as inputs to recognize the gesture. Moreover, WANG teaches that the machine-learning models may include “a K-nearest neighbor algorithm,” which is well-known in the art and operates by comparing an input feature vector to a stored feature vector. It would have been obvious to one having ordinary skill in the art to modify the method to record spectrum peak-point data corresponding to the heart rate information and the scenario information (i.e., the outputs), wherein the spectrum peak-point data corresponding to the heart rate information and the scenario information comprises a peak point location and an amplitude that are of a PPG spectrum and a peak point location and an amplitude that are of an ACC spectrum. One would have been motivated to train a model and use a trained model as taught in WANG in order to accurately determine the heart rate information and scenario information using PPG and ACC data. RAGHURAM and DERVISOGLU suggest using trained models and WANG teaches that the trained models may accurately identify the output/label using a feature vector derived from PPG and ACC signal data. There would have been a reasonable expectation of success as each of RAGHURAM, DERVISOGLU, and WANG teach that PPG and ACC data may be used to generate desired output from trained models. NOTE: Examiner is interpreting as “spectrum peak-point data” to include at least one of a peak-to-peak value, a gradient, skewness, a short-time average amplitude difference, a cliff or a power spectral density as these features necessarily comprise “a peak point location and an amplitude that are of a PPG spectrum and a peak point location and an amplitude that are of an ACC spectrum.” (see, e.g., [0130] of WANG). However, none of the prior art teaches recording the heart rate information and the scenario information in a label buffer and recording the spectrum peak-point data in a data buffer. PARK is concerned with the computational load sustained by low performance devices. “[T]he extensive time required for classification owing to the large amount of computation involved, makes it unsuitable for application to low-performance devices.” (Abstract). PARK teaches classifying input data based on similarities with previous input data. “In this study, we cache class labels and feature vectors extracted from feature maps for images classified by the CNN. Then, when a new image is input, its class label is output based on its similarity with the cached feature vectors. This process can be performed at each layer; hence, if the classification is successful, there is no need to perform the remaining convolution layer operations. This reduces the required classification time.” (emphasis added) (Abstract). Notably, the class labels are cached (i.e., added to a label buffer) and the feature vectors are cached (i.e., added to a data buffer). It would have been obvious to one having ordinary skill in the art to cache class labels and feature vectors, as taught in PARK. Like the devices that concern PARK, wearable devices are (relatively) low-performance devices that are challenged to classify new data within a practical amount of time. Yet wearable devices often receive data that is similar or identical to recently received data (e.g., in situations where the heart rate and/or activity is not changing) like the devices discussed in PARK. One would have been motivated to cache (i.e., buffer) the labels and corresponding feature vectors in order to reduce the classification time and computational load as taught in PARK. There would have been a reasonable expectation of success as buffering simply involves storing recently outputted labels and corresponding feature vectors. With respect to claim 30, and as discussed above with respect to claim 3, WANG also teaches “[d]uring implementation, a person skilled in the art may first determine to-be-used AI models, and model parameters of these AI models are all default values. Therefore, training data needs to be used to determine model parameters applicable to embodiments of this application.” ([0137]). It would have been obvious to one having ordinary skill in the art to modify the method such that, upon obtaining first N pieces of peak-point spectrum data based on the PPG effective spectrum data and the ACC effective spectrum data (i.e., a feature vector), if no peak-point spectrum data in the first N pieces of peak-point spectrum data is prestored in the data buffer, obtaining the heart rate information and the scenario information based on the ACC effective spectrum data corresponding to the ACC signal, the PPG effective spectrum data corresponding to the PPG signal, and the target prediction model or if any peak-point spectrum data in the first N pieces of peak-point spectrum data is prestored in the data buffer, reading, from the label buffer, heart rate information and scenario information that correspond to the any peak-point spectrum data. One would have been motivated to make this modification because, as taught in WANG, it is necessary to build (i.e., train) the model by inputting peak-point spectrum data that the model may or may have not seen yet. There would have been a reasonable expectation of success as each of RAGHURAM, DERVISOGLU, and WANG teach that PPG and ACC data may be used to train a model. Claims 11-16 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Appl. Publ. No. 2016/0058367 A1 (hereinafter “RAGHURAM”) and U.S. Patent Appl. Publ. No. 2021/0093917 A1 (hereinafter “DERVISOGLU”) and U.S. Patent Appl. Publ. No. 2019/0328243 A1 (hereinafter “NEMATI”) as applied to claim 9 above, and further in view Chung H, Ko H, Lee H, Lee J. Deep learning for heart rate estimation from reflectance photoplethysmography with acceleration power spectrum and acceleration intensity. IEEE Access. 2020 Mar 19;8:63390-402 (hereinafter “CHUNG”) (cited in Information Disclosure Statement dated 6/11/2024). With respect to claim 11, the prior art does not explicitly teach that the obtaining a multi-scenario sample set as recited. However, RAGHURAM suggests that the models may be “trained by gathering sensor data during known user activities.” ([0037]). Likewise, DERVISOGLU teaches that “[t]raining the one or more predictive models may include surveying a plurality of datasets including motion data, rotational data, heart rate data, and the like collected during activities having a known activity type (e.g., walking workouts, casual walking activities, running activities, and the like).” (emphasis added) ([0044]). As such, RAGHURAM and DERVISOGLU teach labeling the training data according to the type of activity or exercise. In the same field of endeavor, CHUNG teaches a new deep neural network based on multiclass and non-uniform multilabel classification for HR estimation that can be used by a wearable device. (Abstract). CHUNG teaches connecting a wearable device and a heart rate band device to a data collection module (see p.63391, right-column: “We used two datasets to evaluate our proposed model for HR estimation—the IEEE Signal Processing Cup (ISPC) 2015 dataset (n = 23) and direct measurements obtained by our developed device (n = 48). Both sets of data include multichannel PPG signals and three-axis accelerometer signals simultaneously measured by devices worn on the wrist during intensive physical exercise. For true HR reference, ECG signals were simultaneously measured on the chest… For both datasets, the reference true HRs were measured by ECG data simultaneously recorded by a 24-h Holter monitor (SEER Light, GE Healthcare, Milwaukee, WI, USA).”); when a user wearing the wearable device and the heart rate band device does a first exercise, obtaining, by the data collection module, first heart rate detection data of the wearable device and second heart rate detection data of the heart rate band device; and when the user does a second exercise, obtaining, by the data collection module, third heart rate detection data of the wearable device and fourth heart rate detection data of the heart rate band device, and the multi-scenario sample set comprises the first heart rate detection data, the second heart rate detection data, the third heart rate detection data, and the fourth heart rate detection data (see p.63391, right-column explaining the different exercise intensity levels at which the signal data was acquired by the device and heart-band. “Both sets of data include multichannel PPG signals and three-axis accelerometer signals simultaneously measured by devices worn on the wrist during intensive physical exercise. For true HR reference, ECG signals were simultaneously measured on the chest.”). It would have been obvious to one having ordinary skill in the art to obtain a multi-scenario training dataset by performing the steps recited in claim 11. One would be motivated to obtain training data as taught in CHUNG because, as suggested by RAGHURAM and DERVISOGLU, in order to train the model to identify different activities, the training data should include PPG signals and motion signals collected at known, different activities. There would have been a reasonable expectation of success as CHUNG teaches that training data can be obtained. With respect to claim 12, CHUNG teaches that the extracting ACC sample data and PPG sample data from the multi-scenario sample set comprises: extracting the ACC sample data and the PPG sample data from detection data of the wearable device (see, e.g., pp. 63392 (beginning with B. Preprocessing for Input Layer) to 63393 (top left column)), wherein the detection data of the wearable device comprises the first heart rate detection data and the third heart rate detection data. As discussed above with respect to claim 11, CHUNG teaches that “[b]oth sets of data include multichannel PPG signals and three-axis accelerometer signals simultaneously measured by devices worn on the wrist during intensive physical exercise.”). With respect to claim 13, CHUNG teaches that extracting a heart rate label from the multi-scenario sample set comprises: extracting the heart rate label from detection data of the heart rate band device, wherein the detection data of the heart rate band device comprises the second heart rate detection data and the fourth heart rate detection data. As discussed above with respect to claim 11, CHUNG teaches that the training dataset includes “the reference true HRs…measured by ECG data simultaneously recorded by a 24-h Holter monitor.” (see p.63391, right-column). With respect to claim 14, CHUNG teaches after the extracting ACC sample data, PPG sample data, and a heart rate label from the multi- scenario sample set, the method further comprises: performing fast Fourier transform FFT (see p.63392, top right-column: “After the down-sampling, we computed the power spectra via a 2,048-point Fast Fourier Transform (FFT), where 200 sample vector (8-s data) was padded with trailing zeros to length 2,048, and the frequency bin resolution became 0.012 Hz (0.73 bpm).”) and filtering (see p.63392, bottom left-column: “We filtered all measured signals of Sn (i) [PPG signals] and Am (i) [acceleration signals] using a fourth-order Butterworth band pass filter (BPF) with cutoff frequencies of 0.4 and 4 Hz.”) processing on the ACC sample data and the PPG sample data to obtain filtered ACC sample data and PPG sample data. CHUNG also teaches that the training a deep attention network by using the ACC sample data and the PPG sample data as inputs and using the heart rate label and an exercise scenario label as target variables comprises: training the deep attention network by using the filtered ACC sample data and PPG sample data as inputs (see p.63392, bottom-right column teaching that the inputs are the normalized power spectra for the PPG data and the acceleration data). CHUNG does not teach using the heart rate label and the exercise scenario label as target variables. However, as explained above with respect to claim 11, RAGHURAM and DERVISOGLU clearly suggest that the training data should include PPG signals and motion signals collected at known, different activities so that the trained model can distinguish different activities (i.e., targets). (see, e.g., [0037] of RAGHURAM and [0044] of DERVISOGLU). With respect to claim 15, CHUNG teaches wherein the filtering processing is used to filter out noise data outside [0.7 Hz, 4 Hz] in a spectrum. “We filtered all measured signals of Sn (i) [PPG signals] and Am (i) [acceleration signals] using a fourth-order Butterworth band pass filter (BPF) with cutoff frequencies of 0.4 and 4 Hz.” (p.63392, bottom left-column). Examiner notes that the claimed range constitutes 90% of the range described in CHUNG. With respect to claim 16, CHUNG teaches wherein after the obtaining a first prediction model, the method further comprises: quantizing the first prediction model by using a preset quantization parameter, to obtain a quantized first prediction model. CHUNG notes that the implementing the deep learning network would be difficult because wearable devices require “low power specifications due to the limited computing power.” (p.63401, left column). To address this challenged, CHUNG teaches quantizing using a preset quantization parameter. “we can reduce the model complexity by downsizing the input data. In our proposed model, the input data, which comprised 222 frequency bin-based power spectra, had frequency resolution of 0.0122 Hz (0.73 bpm). Even if we halve the resolution to 111 frequency bins, the frequency resolution is 1.46 bpm, which is also sufficient to provide accurate HR.” (p.63401, bottom left column). It would have been obvious to one having ordinary skill in the art to quantize the prediction model by using a preset quantization parameter, (e.g., halving the number of bins as taught in CHUNG) to obtain a quantized prediction model. One would have been motivated to quantize the prediction model in order to reduce the complexity and/or computing power as taught in CHUNG. There would have been a reasonable expectation of success as CHUNG teaches that the model may be quantized in order to reduce the complexity. Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Appl. Publ. No. 2016/0058367 A1 (hereinafter “RAGHURAM”) and U.S. Patent Appl. Publ. No. 2021/0093917 A1 (hereinafter “DERVISOGLU”) and U.S. Patent Appl. Publ. No. 2019/0328243 A1 (hereinafter “NEMATI”) as applied to claim 1, and in further view of MAXREFDES103# Health Sensor Band User Guide (2020) (hereinafter “MAXIM”). With respect to claim 8, the prior art does not teach sending the heart rate information and the scenario information to a terminal device, wherein the terminal device is an electronic device wirelessly connected to the wearable device; and displaying the heart rate information and the scenario information on a screen of the terminal device. In the same field of endeavor, MAXIM is a user guide provides information about preparing and running the MAXREFDES103# health sensor band. (Abstract, page 1). “This platform uses a high-sensitivity PPG biosensor, power-management IC (PMIC), and microcontrollers from Maxim Integrated® in a wrist-worn design that captures biometric signals important to healthcare.” (Abstract). MAXIM teaches that the wristband may communicate with “evaluation GUIs” installed on a computer (“PC”) that permits the user to “view and log the processed/raw data.” (p. 17). The wristband and the PC may communicate wirelessly through Bluetooth. (p.17). Using the evaluation GUI, one can view the “algorithm data” which includes HR (bpm), HR confidence (%), and Activity, among other data. (p.24). The activity can be Rest, Other, Walk, Run, Bike, Rhythmic. (p.26). It would have been obvious to one having ordinary skill in the art to modify the method so that the wearable device could wirelessly connect with an external device so that the heart rate information and scenario information can be displayed. One would have been motivated to enable wireless connection with a PC or other device so that the user could view data and/or reconfigure settings of the wearable device as taught in MAXIM. There would have been a reasonable expectation of success as MAXIM teaches that a wristband can connect wirelessly to another device. Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Appl. Publ. No. 2016/0058367 A1 (hereinafter “RAGHURAM”) and U.S. Patent Appl. Publ. No. 2021/0093917 A1 (hereinafter “DERVISOGLU”) and U.S. Patent Appl. Publ. No. 2019/0328243 A1 (hereinafter “NEMATI”) as applied to claim 9, and in further view of Reiss et al. “Deep PPG: Large-scale heart rate estimation with convolutional neural networks,” Sensors. 2019 Jul 12;19(14):3079. (hereinafter “REISS”). With respect to claim 8, the prior art does not teach training a deep attention network by using the ACC sample data and the PPG sample data as inputs and using the heart rate label and an exercise scenario label as target variables by adjusting a parameter of a model through cross-validation, so that the model learns to predict heart rate information in different exercise scenarios. In the same field of endeavor, REISS reviews various techniques for estimating heart rate from PPG signals corrupted by motion artifacts. (p.11, 3.1 Algorithms). “We implemented each of the three above described approaches: SpaMa, SpaMaPlus, and Schaeck2017. All three algorithms have in common that they include several adjustable parameters, such as the number of highest peaks to consider in the PPG and acceleration spectra, or the minimum required frequency difference for removing motion-induced peaks in the PPG spectrum. It is common practice in related work to tune these parameters for each dataset-session specifically, instead of applying a cross-validation scheme (e.g., see evaluation results for SpaMa [15] and Schaeck2017 [31]).” (pp.11-12, 3.2 Evaluation Metric and Schemes). REISS notes that the “practical relevance” of these techniques is limited because “[w]hen deploying PPG-based heart rate estimation algorithms in real-life settings, there is no access to ground truth information (using an off-the-shelf chest-worn heart rate monitor would defeat the purpose of wrist-worn heart rate estimation, while only providing an approximate ground truth value with an unknown averaging method).” (p.12, 3.2 Evaluation Metric and Schemes). REISS contends that “[i]nstead of reporting session-optimised results, we argue that leave-one-session-out (LOSO) cross-validation should be preferred. During this cross-validation scheme, parameter optimisation is performed on all data except of one session, and the left-out session is used as test data. This procedure is repeated so that each session is used as test data exactly once. Thus, assuming a dataset with adequate (population, situation, etc.) variety, results reported with LOSO cross-validation can reflect the generalisation capabilities of the developed algorithms.” REISS concludes that “[b]y using our newly introduced dataset and relying on a cross-validation evaluation scheme, a better understanding of the generalisation capabilities of different approaches can be obtained.” (p.23, Conclusions). It would have been obvious to one having ordinary skill in the art to train the deep attention network by adjusting a parameter of a model through cross-validation so that the model learns to predict heart rate information in different exercise scenarios. One would have been motivated to apply REISS’s method of LOSO cross-validation so the reported results would reflect “the generalisation capabilities of the developed algorithms.” (p.12, right before 3.3. Evaluation Results). There would have been a reasonable expectation of success as REISS teaches the LOSO cross-validation scheme can be applied during training. Conclusion Any inquiry concerning this communication or earlier communications from the examiner should be directed to JASON P GROSS whose telephone number is (571)272-1386. The examiner can normally be reached Monday-Friday 9:00-5:00CT. 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, Anne M. Kozak can be reached at (571) 270-5284. 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. /JASON P GROSS/Examiner, Art Unit 3797 /SERKAN AKAR/ Primary Examiner, Art Unit 3797
Read full office action

Prosecution Timeline

Sep 15, 2023
Application Filed
Dec 29, 2025
Non-Final Rejection mailed — §101, §103, §112
Feb 20, 2026
Response Filed
Nov 13, 2026
Final Rejection mailed — §101, §103, §112 (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12788501
SYSTEM AND METHOD FOR GENERATING AN INITIAL TRANSFORMATION MATRIX FOR REGISTERING INTRAOPERATIVE IMAGING DATA WITH PREOPERATIVE IMAGING DATA
2y 6m to grant Granted Nov 10, 2026
Patent 12773270
Transformation Selection Technique Based On Tracker Movement
3y 2m to grant Granted Oct 27, 2026
Patent 12746071
AUTOMATED REGISTRATION OF PREOPERATIVE VOLUME IMAGE DATA, USING A SEARCH IMAGE
2y 2m to grant Granted Sep 29, 2026
Patent 12714381
MEDICAL ANALYSIS APPARATUS AND RELATIVE METHOD
3y 6m to grant Granted Aug 25, 2026
Patent 12708452
SYSTEMS, DEVICES, AND METHODS FOR DETERMINING POSITIONS OF SURGICAL TOOLS IN ENDOSCOPIC IMAGES
2y 3m to grant Granted Aug 18, 2026
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

3-4
Expected OA Rounds
64%
Grant Probability
99%
With Interview (+43.3%)
Moderate
PTA Risk
Based on 25 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month