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 .
Status of Claims
This action is in response to the amendments filed 08/14/2025. Claims 1-8, 11-16, and 19-20 have been amended, claims 10 and 18 have been cancelled, claims 21-24 have been added. Claims 1-8, 11-16, and 19-24 are currently pending.
Response to Arguments
Claims 10 and 18 have been cancelled, therefore the rejections of claims 10 and 18 no longer stand.
Applicant’s arguments regarding the 101 rejection have been fully considered but they are not persuasive. Applicant argues that “the human mind is not equipped to perform at least, for example, the features of “training a first artificial intelligence (AI) model based on training data. . . training a second model, based on the training data and a specialized tree algorithm, for segmentation. . .the prescriptive tree maximizes the likelihood of the desired outcome based on the segmentation, using a splitting criterion. . .”. Examiner respectfully notes that, as the limitations directed to training the first and second model did not recite any particular models nor particular training steps, these limitations were interpreted mere instructions to apply mental steps related to determining the likelihood of a desired outcome for a given action and using a mentally determined splitting criterion to segment data and maximize this likelihood using generic computer components. Similarly, the limitations directed to “determining, via a teacher model, an expected outcome. . . providing an interpretable prescriptive policy. . .that is more interpretable than a first policy and has a determined difference (between a best outcome and an expected outcome) less than a predetermined threshold. . .and adjusting the prescriptive tree utilizing the determined difference as a loss function, wherein adjusting comprises finetuning the given depth of the prescriptive tree” were not claimed in such a way that precludes interpretation of these steps as mental steps merely implemented by generic computer components.
Applicant also argues that “the instant claim recites technical improvement for improving the accuracy and interpretability of a prescriptive model by adjusting the prescriptive tree based on the determined difference between the best expected outcome and the expected outcome of interpretable prescriptive policies. The claimed method, when viewed as a whole, provides an improvement over existing prescriptive analytics approaches”. Examiner respectfully disagrees and notes that section 2106.05(a) of the MPEP states “a technical explanation as to how to implement the invention should be present in the specification. That is, the disclosure must provide sufficient details such that one of ordinary skill in the art would recognize the claimed invention as providing an improvement. The specification need not explicitly set forth the improvement, but it must describe the invention such that the improvement would be apparent to one of ordinary skill in the art. Conversely, if the specification explicitly sets forth an improvement but in a conclusory manner (i.e., a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology” and later states “It is important to note, the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements”. Examiner notes that Applicant’s alleged improvement is related to the adjustment of a prescriptive tree, which was interpreted as amental step. The claims do not describe finetuning the depth of a prescriptive tree model in a way that requires specific, computer-executed operations or event automated adjustment. The claim instead recites two models at a high level without providing sufficient technical detail to make it clear how the prescriptive tree is adjusted in a way that a person could replicate by adjusting a mental model of a prescriptive tree. Examiner also notes, in response to Applicant’s comparison, that claim 3 of Example 47 of the July 2024 Subject Eligibility Examples was determined to be eligible as “Steps (d)-(f) provide for improved network security using the information from the detection to enhance security by taking proactive measures to remediate the danger by detecting the source address associated with the potentially malicious packets. Specifically, the claim reflects the improvement in step (d), dropping potentially malicious packets in step (e), and blocking future traffic from the source address in step (f)”. In this claim, the improvement was not to steps (a)-(c), which were interpreted as judicial exceptions, but to the additional elements in steps (d)-(f). Applicant’s claim does not recite any additional elements that amount to an improvement nor significantly more than the recited abstract ideas; therefore, the rejection is maintained. The 101 rejections have been updated to include the amended limitations and to clarify the reasoning given for the limitations that were not amended.
Applicant’s arguments regarding the prior art rejection have been fully considered but they are not persuasive. Applicant argues that the Liu reference does not teach “adjusting the prescriptive tree based on one or more pre-determined constraints and the determined difference between the best expected outcome and the expected outcome. . .wherein the determined difference is utilized as a loss function for the adjusting of the prescriptive tree”. Examiner respectfully disagrees and notes that Liu teaches that the maximum depth of the tree model can be adjusted in in section IV C, and that figure 3 shows the how the output accuracy of the tree model is adjusted at different tree depths. Section III C of Liu teaches how the output of the decision tree model is utilized in the loss function (see the derivation of equations 9-12). The prior art rejections have been updated to include the amended limitations and to clarify the reasoning given for the limitations that were not amended
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 1-8, 11-16, and 19-24 are rejected under 35 U.S.C. 101. Claims 1-8 are directed to a method, claims 11-16 and 21 are directed to a system, and claims 19-20 and 22-24 are directed to a non-transitory computer readable medium; therefore, claims 1-8, 11-16, and 19-24 fall within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter). However, claims 1-8, 11-16, and 19-24 fall within the judicial exception of an abstract idea, specifically the abstract ideas of “Mental Processes” (including observation, evaluation, and opinion) and “Mathematical Concepts (including mathematical calculations and relationships)”.
Claim 1:
Claim 1 is directed to a method; therefore, the claim does fall within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
Claim 1 recites the following abstract ideas:
determining, via the teacher model, a first policy that produces an optimal action, wherein the optimal action provides a best expected outcome (mental step directed to evaluation, judgement – a person could determine in their mind a policy that produces an optimal action and/or best expected outcome. As no particular model nor particular steps of using a teacher model to determine a policy are claimed, Examiner is interpreting the teacher model as a generic computer component and interpreting determining a policy via a teacher model as mere instructions to apply an abstract idea using a generic computer component (see MPEP 2106.05(f));
applying, via the prescriptive tree, a recursive segmentation algorithm to generate one or more interpretable prescriptive policies, wherein each interpretable prescriptive policy of the one or more interpretable policies is less complex and more interpretable than the first policy, each recursive split in the prescriptive tree separates data into two data sets, and recursive splitting terminates once the prescriptive tree reaches a given depth (mental step directed to evaluation, judgement – a person could apply a tree-based recursive segmentation algorithm in their mind to separate data into sets, potentially assisted by pen and paper (see MPEP 2106.04(a)(2)(III)), until the tree reaches a predetermined given depth to generate interpretable prescriptive policies in their mind);
determining, via the teacher model, an expected outcome for each interpretable prescriptive policy of the one or more interpretable prescriptive policies (mental step directed to evaluation, judgment – a person could determine in their mind an expected outcome for a generated interpretable prescriptive policy (see above for interpretation of “via the teacher model”));
determining a difference between the best expected outcome and the expected outcome for the interpretable prescriptive policy of the one or more interpretable prescriptive policies (mental step directed to observation, evaluation – a person could determine a difference between an observed best expected outcome and an observed expected outcome of an interpretable prescriptive policy in their mind);
and providing an interpretable prescriptive policy of the one or more prescriptive policies that is more interpretable than the first policy and has the determined difference less than a pre-determined threshold (mental step directed to evaluation, judgment – a person could provide an observed interpretable policy in their mind, potentially assisted by pen and paper, having determined that the difference between a best expected outcome and an expected outcome is less than a pre-determined threshold in their mind);
adjusting the prescriptive tree based on one or more pre-determined constraints and the determined difference, wherein the determined difference is utilized as a loss function for the adjusting of the prescriptive tree, and the adjusting comprises finetuning the given depth of the prescriptive tree (mental step directed to observation, judgement – a person could finetune the depth of, or mentally adjust a prescriptive tree in their mind, potentially assisted by pen and paper, having observed or determined constraints and differences as a loss function in their mind).
Claim 1 recites the following additional elements:
training a first artificial intelligence (Al) model based on training data to determine a likelihood of a desired outcome for a given action, wherein the first AI model comprises a teacher model; training a second model, based on the training data and a specialized tree algorithm for segmentation, the second model comprises a prescriptive tree, and the prescriptive tree maximizes the likelihood of the desired outcome based on the segmentation, using a splitting criterion. Training an artificial intelligence model and a second model comprising a prescriptive tree are claimed without particular details regarding the structure of the models or particular training steps, but only describe the outcomes that the models are trained to perform (determining a likelihood and segmentation). Therefore, these models are interpreted as a generic computer components used to merely implement the mental steps associated with determining a likelihood of an outcome for an action and segmenting data using a splitting criterion to maximize this likelihood. These additional elements do not integrate the abstract ideas from claim 1 into a practical application or amount to significantly more than the abstract ideas recited in claim 1 (see MPEP 2106.05(b), MPEP 2106.05(d), and MPEP 2106.05(f)).
Claim 11 is a system claim and its limitation is included in claim 1. The only difference is that claim 11 requires a system. Therefore, claim 11 is rejected for the same reasons as claim 1.
Claim 19 is a non-transitory computer readable medium claim and its limitation is included in claim 1. The only difference is that claim 19 requires a non-transitory computer readable medium. Therefore, claim 19 is rejected for the same reasons as claim 1.
The independent claims are not patent eligible.
Dependent claims 2-8, 12-16, and 20-24 when analyzed as a whole are held to be patent ineligible under 35 U.S.C. 101 because the additional recited limitations fail to establish that the claims are not directed to an abstract idea, as they recite further embellishment of the judicial exception.
Claim 2 recites wherein the segmentation comprises constructing a decision tree, and wherein the decision tree is customizable based on the splitting criterion that is user-defined and that optimizes the desired outcome. Constructing a customizable decision tree to optimize an outcome based on a user-defined splitting criterion is interpreted as a mental step directed to observation, evaluation – a person could construct a customizable decision tree in their mind, potentially assisted by pen and paper, based on user-defined splitting criteria to optimize an outcome.
Claim 3 recites wherein the teacher model is a neural network. Examiner notes that the neural network is claimed without particular details regarding the structure, training, or use/inference and is therefore interpreted as a generic computer component (see MPEP 2106.05(b).
Claim 4 recites wherein each leaf node of the prescriptive tree represents a corresponding interpretable prescriptive policy of the one or more interpretable prescriptive policies for a particular segment of a population, and demographics of the particular segment are specified by a path from a root of the prescriptive tree to the leaf node. This limitation is interpreted selecting a particular type of data to be manipulated and does not integrate the abstract ideas recited in claim 1 in to a practical application or amount to significantly more than the abstract ideas recited in claim 1 (see MPEP 2106.05(g)).
Claims 5-7 recite different applications for the model which include targeted pricing, promotion, and personalized medicine. These are all interpreted as generally linking the abstract ideas recited in claim 1 (on which claim 4 depends) to a given field of use and technological environment and do not integrate the abstract ideas recited in claim 1 in to a practical application or amount to significantly more than the abstract ideas recited in claim 1 (see MPEP 2106.05(h)).
Claim 8 recites selecting, the interpretable prescriptive policy, from the one or more interpretable prescriptive policies based on the expected outcome for each interpretable prescriptive policy of the one or more interpretable prescriptive policies. Selecting a policy based on an expected outcome is interpreted as a mental step directed to evaluation, judgement – a person could select in their mind a particular policy based on expected outcomes.
Claim 12 is a system claim and its limitation is included in claim 4. Claim 12 is rejected for the same reasons as claim 4.
Claim 13 is a system claim and its limitation is included in claim 5. Claim 13 is rejected for the same reasons as claim 5.
Claim 14 is a system claim and its limitation is included in claim 6. Claim 14 is rejected for the same reasons as claim 6.
Claim 15 is a system claim and its limitation is included in claim 7. Claim 15 is rejected for the same reasons as claim 7.
Claim 16 is a system claim and its limitation is included in claim 8. Claim 16 is rejected for the same reasons as claim 8.
Claim 20 is a non-transitory computer readable medium claim and its limitation is included in claim 4. Claim 20 is rejected for the same reasons as claim 4.
Claim 21 is a system claim and its limitation is included in claim 3. Claim 21 is rejected for the same reasons as claim 3.
Claim 22 is a non-transitory computer readable medium claim and its limitation is included in claim 3. Claim 22 is rejected for the same reasons as claim 3.
Claim 23 is a non-transitory computer readable medium claim and its limitation is included in claim 5. Claim 23 is rejected for the same reasons as claim 5.
Claim 24 is a non-transitory computer readable medium claim and its limitation is included in claim 6. Claim 24 is rejected for the same reasons as claim 6.
Viewed as a whole, these additional claim elements do not provide meaningful limitations to transform the abstract idea into a patent eligible application of the abstract idea such that the claims amount to significantly more than the abstract idea itself. Therefore, the claims are rejected under 35 U.S.C. 101 as being directed to non-statutory subject matter.
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 1-8, 11-16, and 19-24 are rejected under 35 U.S.C. 103 as being unpatentable over Liu et al (“Improving the Interpretability of Deep Neural Networks with Knowledge Distillation”, herein Liu) in view of Gao et al (“Healthcare Supply Chain Network Coordination Through Medical Insurance Strategies with Reference Price Effect”, herein Gao).
Regarding claim 1, Liu teaches a computer-implemented method for integrated segmentation and [prescriptive] policies generation (section 1 para. 9 recites “we employ knowledge distillation for another purpose: interpretation. We resolve the tension between interpretability and accuracy performance by distilling deep neural nets into vanilla decision trees”), the computer-implemented method comprising:
training a first artificial intelligence (Al) model based on training data to determine a likelihood of a [desired] outcome [for a given action], wherein the first AI model comprises a teacher model; and training a second model based on the training data and a specialized tree algorithm, for segmentation, the second model comprises a [prescriptive] tree and the [prescriptive] tree maximizes the likelihood of the [desired] outcome based on the segmentation, using a splitting criterion (section III D para. 1 recites “attempts, we employ the matching logits method when distilling CNN into vanilla decision trees. Fig. 2 illustrates the framework of our method. In this figure, the architecture of the CNN is the one used to train the MNIST data as in [61]. It comprises of two convolutional layers and two pooling layers followed by two fully connected layers: fc1 and fc2. After this deep CNN is trained, we feed the feature part X of the original training data to the trained model to obtain the corresponding logits Z. Then we train CART with X and Z which is treated as the targets”. Section III D para. 3 recites “Once CART is trained, in order to obtain the final prediction results on test cases we need to add a softmax layer over the test results of CART to turn numerical test results into categorical ones. Assuming the test results on CART is k, the final output probabilities for class i therefore is (EQ 16)”. Section IV C para. 1 recites “When we are performing classification tasks applying a decision tree, there are a variety of parameters to tune such as the minimum number of samples per leaf, the strategy used to choose the split at each node (either the best split or the best random split) and so on. We select two parameters that would influence the performance of a decision tree substantially: the maximum depth of the tree and the functions to measure the impurity of a split (either “gini” or “entropy”)”. Section IV D para. 1 recites “the teacher model for the MNIST dataset already has a very high accuracy of 99.25%, but the student model’s highest accuracy in Table II is only 86.55%. However, from our experiments we found that training a good teacher model indeed helped to boost the distillation results. In our experiments, we notice that distillation helps to improve the accuracy by 1% to 5%” (i.e., training a first teacher model to determine the likelihood of an outcome and a second tree model comprising a tree model, wherein the tree model uses a splitting criteria to find the highest possible, or maximum outcome likelihood));
determining, via the teacher model, a first policy that [produces an optimal action, wherein the optimal action] provides a best expected outcome (section III C para. 1 recites “Knowledge distillation transfers the generalization ability of a complex teacher model to a simple student model. Using the teacher model’s soft targets for distillation could produce much better outcome than hard targets. Fig. 1 shows an example of hard and soft targets. Hard targets just contain the information for the predict label while soft targets reveal all the predicted probabilities for all the classes”. Section III D para. 1 recites “we employ the matching logits method when distilling CNN into vanilla decision trees. Fig. 2 illustrates the framework of our method. In this figure, the architecture of the CNN is the one used to train the MNIST data as in [61]. It comprises of two convolutional layers and two pooling layers followed by two fully connected layers: fc1 and fc2”. Section IV C para. 2 recites “For the MNIST dataset, the teacher CNN model achieves an accuracy of 99.25%” (i.e., determining a first output, or policy which provides a best expected outcome, from a teacher model));
applying, via the [prescriptive] tree, a recursive segmentation algorithm to generate one or more interpretable [prescriptive] policies, wherein each interpretable [prescriptive] policy of the one or more interpretable [prescriptive] policies is less complex and more interpretable than the first policy (section III B para. 1 recites “Classification and Regression Trees (CART) was introduced in 1984 by Breiman et al. Although CART and C4.5 were invented by different authors, they follow similar ideas for training decision trees. Owing to the reason that CART supports numerical target values (regression) and the key to our methodology is to solve a multi-output regression problem, we introduce briefly here the algorithm of CART. CART applies a greedy approach which constructs the decision tree in a top-down recursive divide-and-conquer manner. As our experiments applies CART for regression, the descriptions focus on regression tasks” (i.e., using a second model to apply a recursive segmentation algorithm to the output of the teacher model to produce a less complex and more interpretable output)), each recursive split in the [prescriptive] tree separates data into two data sets, and the recursive splitting terminates once the [prescriptive] tree reaches a given depth (section III B para. 2-3 recite “This algorithm partitions the feature space and groups instances with the same labels together. Initially, it constructs a root node with all training samples S with features as xi ϵ Rn for i = 1. . .l and labels as yi ϵ Rl and split the node into two child nodes recursively. The splitting criterion is: C = (a, tn), where a is the attribute to split on and tn is the threshold at node n. This criterion partitions S into Sleft(C) = (x, y)|xa ≤ tn, Sright(C) = S \ Sleft(C). Then the algorithm recursively splits Sleft(C) and Sright(C) until the maximum depth specified by the user is reached” (i.e., the CART algorithm is recursive and terminates when the tree reached a given depth));
determining, via the teacher model, an expected outcome for each interpretable [prescriptive] policy of the one or more interpretable [prescriptive] policies (section III D para. 1 recites “After this deep CNN is trained, we feed the feature part of the original training X data to the trained model to obtain the corresponding logits Z. Then we train CART with X and Z which is treated as the targets”. Section III D para. 3 recites “Once CART is trained, in order to obtain the final prediction results on test cases we need to add a softmax layer over the test results of CART to turn numerical test results into categorical ones. Assuming the test results on CART is k, the final output probabilities for class i therefore is (EQ 16)” (i.e., determining the likelihood of an expected outcome for the interpretable output. Examiner’s Note: Applicant’s claim does not specifically require a framework wherein the output of the prescriptive tree is sent back into the teacher model as an input, but only requires that the teacher model determine an expected outcome for each interpretable policy. Therefore, the broadest reasonable interpretation of “determining, via the teacher model” includes the additional softmax layer added to the model to output the results of the CART decision tree by way of the knowledge distillation from the teacher model));
determining a difference between the best expected outcome and the expected outcome for each interpretable [prescriptive] policy of the one or more [prescriptive] policies; and providing an interpretable [prescriptive] policy of the one or more [prescriptive] policies that is more interpretable than the first policy and has the determined difference less than a predetermined threshold (section III B para. 2 recites “This algorithm partitions the feature space and groups instances with the same labels together. Initially, it constructs a root node with all training samples S with features as xi ϵ Rn for i = i . . .l and labels as yi ϵ Rl and split the node into two child nodes recursively. The splitting criterion is: C = (a, tn), where a is the attribute to split on and tn is the threshold at node n. The optimal attribute and the splitting threshold are found. Then the algorithm recursively splits Sleft(C) and Sright(C) until the maximum depth specified by the user is reached, a node becomes pure, Mn < minsamples or Mn = 1”. Section IV C para. 2 recites “The performance for the student model and the vanilla decision tree classification results are shown in Table IV. “Acc student” represents the accuracy of the student decision tree trained using the logits of the teacher CNN model on TensorFlow”. Section IV D para. 1 recites “the teacher model for the MNIST dataset already has a very high accuracy of 99.25%, but the student model’s highest accuracy in Table II is only 86.55%. However, from our experiments we found that training a good teacher model indeed helped to boost the distillation results. In our experiments, we notice that distillation helps to improve the accuracy by 1% to 5%” (i.e., determining a difference between the best expected outcome output by the teacher model and the expected outcome output by the tree model by way of the distillation from the teacher model and providing an interpretable policy that meets a threshold determined in part based on the splitting criteria));
and adjusting the [prescriptive] tree based on one or more pre-determined constraints and the determined difference, wherein the determined difference is utilized as a loss function for the adjusting of the [prescriptive] tree, and the adjusting comprises finetuning the given depth of the prescriptive tree (section III C para. 2 recites “Take CNN for example, the last hidden layer l before the softmax layer is a fully connected layer with logits z as the output (EQ9). Here zi is the logit for one of the classes: i. j is the number of hidden nodes for layer l – 1. W and b are weights and bias respectively. The softmax layer calculates the output probabilities for each class as (EQ10). The cross-entropy function is then applied to calculate the loss of the model (EQ11). Hence, to avoid the loss of information, it is desirable to use logits z instead of the predicted probabilities q. This method is called “matching logits” and the pioneer work was done in [33]. Hinton et al. [27] extended their work to a more general case by inserting a temperature term T into (EQ10) – (EQ12) and they demonstrated mathematically that in the high temperature limit and when the logits were zero-meaned separately for each training instance of the student model, matching logits was a special case of using the soft targets for distillation” (i.e., adapting the logits output from a decision tree so that they can be utilized in the loss function). Section IV D para. 3 recites “We are also curious about the performances of the student models and the vanilla decision trees when the maximum depth of the tree is not specified. In this situation, for the MNIST dataset, we found that the accuracy for the student model was 88.28% and the decision tree classification achieved 87.4% for the criterion of “gini”. For the Connect-4 dataset, when the teacher model has an accuracy of 83.22% the student model achieves 79.06% and vanilla decision tree has 77.57% when using “gini” as impurity measure. We notice that the accuracy improvements are smaller than the cases where the depth of the trees are specified. This is easy to explain as when the tree levels are not set the vanilla decision tree takes much deeper tree levels than the student model to arrive at the current accuracy results. Hence these decision trees are far less interpretable than the student models because the level of tree depth determines the interpretability for decision trees” (i.e., adjusting, or finetuning, the depth of the tree model based on constraints and a determined difference between best and other expected outcomes determining using a loss function)).
However, Liu does not explicitly teach a model to determine a desired outcome for a given action, a policy that produces an optimal action, or a prescriptive policy.
Gao teaches a model to determine a desired outcome for a given action, a policy that produces an optimal action, and a prescriptive policy (section 7.1 para. 1-2 recites “we establish a dynamic game theory model with reference price effect and analyze the influence of medical insurance reimbursement strategies on healthcare supply chain network coordination. The results show that: The medical insurance reimbursement strategy has a positive correlation with patients’ reference price effect. A higher reference price effect can get more reimbursement and other factors such as the marginal profit and service efficiency also influence reimbursement”. Section 7.1 para. 4 recites “In Table 4 under the centralized decision section, both the subordinate and superior hospital can get more reimbursement under centralized decisions than decentralized decisions of NS and SS (i.e., a desired outcome for a given action). The additional reimbursement will increase the willingness of cooperation between superior and subordinate hospitals. In Table 5, under the centralized decision section, it illustrates that when the number of patient admissions to a subordinate hospital increases, the admission rate decreases at the superior hospital. Under all the medical insurance strategies the centralized strategy performs best in this area. Furthermore, the average price of subordinate hospitals is lower than at superior hospitals, but the reimbursement is higher, which will reduce the patients’ expenses for medical services (i.e., a policy that produces an optimal outcome). 3) It also can be seen from Table 7 that the total utility under the centralized decision making of the healthcare supply chain network is higher than the decentralized one. In conclusion, the centralized decision scenario can further help facilitate a more efficient way of allocating medical resources. With a more strategic allocation of medical resources, healthcare services will be more accessible for patients at affordable prices, so vertical integration of the healthcare supply chain network with centralized decision provides the best outcome for the healthcare supply chain network” (i.e., a model which outputs suggested, or prescriptive policies)).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine these teachings by applying the teacher student model from Liu to the prescriptive policy data from Gao. Liu and Gao are both directed to models for optimizing a process, and while Liu describes an example of using the teacher student model for image classification, one of ordinary skill in the art would understand that the convolutional neural network used in Liu is not limited to image classification but could be used to determine patterns in any kind of input data. Liu teaches the benefits of this using this teacher-student model for uses cases like the prescriptive policy data from Gao, as section I paragraphs 1-2 of Liu recite “Understanding why a specific prediction is made is of utmost importance for the end-users to trust and adopt the model, and for the system designers to refine the model by performing feature engineering and parameter tuning. This is especially true for high stakes domains such as clinical decision support, disaster response and recidivism prediction. For instance, decision trees are preferred over DNN in the health care domain for disease diagnosis due to their ease of interpretation”. The abstract of Gao also supports the benefits of this combination, reciting “a centralized decision- making strategy can stimulate both superior hospital and subordinate hospital’s cooperative intentions which benefits the social healthcare system”.
Regarding claim 2, the combination of Liu and Gao teaches the computer-implemented method of claim 1, wherein the segmentation comprises constructing a decision tree and wherein the decision tree is customizable based on the splitting criterion that is user-defined and that optimizes the desired outcome (Liu section IV C para. 1 recites “For decision tree classifications, we apply the modules in the scikit-learn machine learning tool. When we are performing classification tasks applying a decision tree, there are a variety of parameters to tune such as the minimum number of samples per leaf, the strategy used to choose the split at each node (either the best split or the best random split) and so on. We select two parameters that would influence the performance of a decision tree substantially: the maximum depth of the tree and the functions to measure the impurity of a split (either “gini” or “entropy”)” (i.e., the decision tree has a customizable user defined splitting criterion for optimization)).
Regarding claim 3, the combination of Liu and Gao teaches the computer-implemented method of claim 1, wherein the teacher model is a neural network (Liu section III C para. 1 recites “Knowledge distillation transfers the generalization ability of a complex teacher model to a simple student model. Using the teacher model’s soft targets for distillation could produce much better outcome than hard targets”. Liu section III D para. 1 recites “we employ the matching logits method when distilling CNN into vanilla decision trees. Fig. 2 illustrates the framework of our method. In this figure, the architecture of the CNN is the one used to train the MNIST data as in [61]. It comprises of two convolutional layers and two pooling layers followed by two fully connected layers: fc1 and fc2” (i.e., the teacher model is a neural network)).
Regarding claim 4, the combination of Liu and Gao teaches the computer-implemented method of claim 1, wherein each leaf node of the prescriptive tree represents a corresponding interpretable prescriptive policy of the one or more interpretable prescriptive policies for a particular segment of a population, and demographics of the particular segment are specified by a path from a root of the prescriptive tree to the leaf node (fig. 2 of Liu illustrates the path of an interpretable tree from a root node to one or more leaf nodes representing a given interpretable policy for a particular segment of training data. Gao section 6 para. 1 recites “The data from a region of China is used to analyze effect of reference price on medical insurance reimbursement strategy and how it influences the patients’ choice and utility of healthcare supply chain network” (i.e., determining policies for segments of a given demographic)).
Regarding claim 5, the combination of Liu and Gao teaches the computer-implemented method of claim 4, wherein each model, of the first AI model and the second model, is deployed for use in an application involving targeted pricing, each interpretable prescriptive policy of the one or more interpretable prescriptive policies represents an optimal product price for a segment of customers, and the best expected outcome represents a maximum expected revenue from the targeted pricing (Gao section 4.2 para. 1-3 recite “In a centralized decision model, the government establishes the medical insurance reimbursement policy to maximize the overall interest of the whole healthcare supply chain network. Under the current supply chain structure of vertical integration of medical consortia, if a binding cooperation agreement is reached on the medical policy project, the coordination between the superior and subordinate hospitals in the healthcare supply chain network will be carried out. Then, we use a centralized decision model and according to Equation (8), its Hamiltonian function is (EQ 26), where vsc is an adjoint variable that denotes the reference price effect impact on the supply chain. According to the Equation (26), using the optimal control theory and dynamic programming, we get the best medical insurance subsidies for superior and subordinate hospital, which is shown as Proposition 4. According to the above results, when taking the reference price effect into account in the centralized decision model of the supply chain, the reimbursement amount of medical insurance for superior and subordinate hospitals depends on the patients’ reference price and the marginal profit value of medical institutions. Different from the decentralized decision model, in the vertical integration of medical systems, each medical institution would consider the marginal benefits of its related upstream and downstream institutions when making decisions, which improves the coordination of the healthcare supply chain network. Therefore, Proposition 4 indicates that, in coordination of the supply chain, the medical treatment insurance reimbursement is not only determined by the reference price effect, but also the all cooperative medical institutions’ marginal profit. In this way we can promote the integration of medical institutions and maximize the benefit of entire healthcare supply chain network” (i.e., a model, such as the composite first and second models from Liu, used for targeting pricing which optimizes for maximal expected revenue)).
Regarding claim 6, the combination of Liu and Gao teaches the computer-implemented method of claim 4, wherein each model, of the first AI model and the second model, is deployed for use in an application involving targeted promotion, each interpretable prescriptive policy of the one or more interpretable prescriptive policies represents an optimal product discount for a segment of customers, and the best expected outcome represents a maximum expected revenue from the targeted promotion (Gao section 1 para. 6 recites “In this study, common diseases that can be treated equally well in both subordinate and superior hospitals are discussed. Patients have right to choose an appropriate hospital, but an important factor is that if patients want to get the normal medical insurance reimbursement ratio at a superior hospital, they must be referred by the doctor of the subordinate hospital and obtain the transfer certification. By doing so, patients can enjoy a high ratio of the reimbursement of about 85%. Contrary, patients without the transfer certification can only enjoy a reimbursement ratio of 45%. Due the variation in reimbursement ratio stipulated by different hospitals, the patients’ reference price effect will affect their choice of hospitals. Meanwhile, medical insurance can play a vital role in promoting the implementation of patient triage and hierarchical diagnosis and treatment”. Gao section 4.1.1 para. 1-3 recites “In this strategy, superior and subordinate hospitals do not know each other’s decision. Both participates conduct simultaneous action and obtain their respective optimal decisions to maximize their own interests, which is eligible with Nash equilibrium game conditions. In the process, the medical insurance reimbursement for the superior hospital is Snsh , and for the subordinate hospital is Snsb(t). Using the optimal control theory of dynamic programming, we can obtain the solution as follows: See Proposition 1 and EQ 11-17. It can be seen from Tables 2 and 3 that Snsb is positively correlated with θb and Snsh is positively correlated with θh. At the same time, Snsh , Snsb are both positively correlated with δ. This means that the medical insurance reimbursement will increase with the increasing of reference price effect, and the cognition of overpricing will be alleviated, which solves the problem of the high cost of getting medical treatment. Since Snsh is positively correlated with ρh and Snsb is positively correlated with ρb, we get that medical insurance reimbursements can improve the marginal benefit. Snsh has a negative correlation with ηh which means it is positively correlated with the service efficiency of hospitals, as is Snsb with ηb. This conclusion indicates that medical insurance reimbursements can improve the service efficiency of hospitals and reduce patients’ waiting time” (i.e., a model, such as the composite first and second models from Liu, used for targeted promotion which optimizes for a maximum revenue based on targeted promotions)).
Regarding claim 7, the combination of Liu and Gao teaches the computer-implemented method of claim 4, wherein each model, of the first AI model and the second model, is deployed for use in an application involving personalized medicine, each interpretable prescriptive policy of the one or more interpretable prescriptive policies represents an optimal treatment for a segment of patients, and the best expected outcome represents a maximum success rate from the personalized medicine (Gao section 6 para. 1 recites “The data from a region of China is used to analyze effect of reference price on medical insurance reimbursement strategy and how it influences the patients’ choice and utility of healthcare supply chain network”. Gao section 6 para. 7 recites “Under the centralized strategy, the total profit for this certain disease is reduced while the total medical cost will decrease, which solves the problem that the medical insurance investment continues to increase but the total medical cost of patients increases. In terms of the overall medical system, common diseases can be basically cured in subordinate hospitals, which is a rational use of resources. The medical resources of superior hospitals can be used to treat relatively complex diseases, and the overall utility of the medical system will improve. The total utility under the different decision making of the healthcare supply chain network is shown in Table 7” (i.e., a model, such as the composite first and second models from Liu, used involving optimal treatment for a segment of patients optimized for a maximum success rate of treatment)).
Regarding claim 8, the combination of Liu and Gao teaches the computer-implemented method of claim 1, further comprising: selecting, the interpretable prescriptive policy, from the one or more interpretable prescriptive policies based on the expected outcome for each interpretable prescriptive policy of the one or more interpretable prescriptive policies (Liu section IV C para. 2-3 recites “For the MNIST dataset, the teacher CNN model achieves an accuracy of 99.25%. The performance for the student model and the vanilla decision tree classification results are shown in Table IV. “Acc student” represents the accuracy of the student decision tree trained using the logits of the teacher CNN model on TensorFlow. “Acc gini” is the accuracy of the decision tree without distillation when the impurity measure is “gini” in scikit-learn when trained utilizing the same training and test data as the CNN model. “Acc entropy” is the classification accuracy of the decision tree when the impurity measure is “entropy”. We highlighted the best performance in bold. Under different tree depths, the student model always outperforms the vanilla decision tree. The same conclusion holds true for the Connect-4 dataset in Table V where the accuracy for the MLP teacher model is 86.62%. The reason we limit the tree depth to 10 is that we would like to construct interpretable models and trees over a depth of 10 becomes extremely hard for human cognitions to comprehend” (i.e., selecting a given interpretable outcome from a plurality of interpretable outcomes)).
Claim 11 is a system claim and its limitation is included in claim 1. The only difference is that claim 11 requires a system. Therefore, claim 11 is rejected for the same reasons as claim 1.
Claim 12 is a system claim and its limitation is included in claim 4. Claim 12 is rejected for the same reasons as claim 4.
Claim 13 is a system claim and its limitation is included in claim 5. Claim 13 is rejected for the same reasons as claim 5.
Claim 14 is a system claim and its limitation is included in claim 6. Claim 14 is rejected for the same reasons as claim 6.
Claim 15 is a system claim and its limitation is included in claim 7. Claim 15 is rejected for the same reasons as claim 7.
Claim 16 is a system claim and its limitation is included in claim 8. Claim 16 is rejected for the same reasons as claim 8.
Claim 19 is a non-transitory computer readable medium claim and its limitation is included in claim 1. The only difference is that claim 19 requires a computer program product. Therefore, claim 19 is rejected for the same reasons as claim 1.
Claim 20 is a non-transitory computer readable medium claim and its limitation is included in claim 4. Claim 20 is rejected for the same reasons as claim 4.
Claim 21 is a system claim and its limitation is included in claim 3. Claim 21 is rejected for the same reasons as claim 3.
Claim 22 is a non-transitory computer readable medium claim and its limitation is included in claim 3. Claim 22 is rejected for the same reasons as claim 3.
Claim 23 is a non-transitory computer readable medium claim and its limitation is included in claim 5. Claim 23 is rejected for the same reasons as claim 5.
Claim 24 is a non-transitory computer readable medium claim and its limitation is included in claim 6. Claim 24 is rejected for the same reasons as claim 6.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
US 20200380380 A1 (Ramachandran et al) teaches a method for leveraging unsupervised machine learning to produce interpretable routing rules using decision trees.
US 20210390417 A1 (Shahrzad et al) teaches a method for generating an explainable surrogate-assisted evolutionary optimization method to discover rule-based decision strategies for which actions to take to achieve certain outcomes when historical training data is limited or unavailable.
US 20230289623 A1 (Zhang et al) teaches a method for generating a predictive model and a prescriptive model through an offline learning process at a first system and autonomously updating the predictive model and the prescriptive model from feedback.
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 LEAH M FEITL whose telephone number is (571) 272-8350. The examiner can normally be reached on M-F 0900-1700 EST.
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, Viker Lamardo can be reached on (571) 270-5871. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see https://ppair-my.uspto.gov/pair/PrivatePair. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll- free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/L.M.F./ Examiner, Art Unit 2147
/ERIC NILSSON/Primary Examiner, Art Unit 2151
Prosecution Insights