Lifestyle recommendation system for chronic disease mitigation

Abstract

A computer-implemented, machine learning method for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient includes extracting structured information from unstructured static data. Real-time data is transformed into a homogeneous representation. A portion of data is selected based on a pre-defined time interval or frequency. A patient graph is constructed using the structured information, the homogenous representation, and the portion of data. Whether there is a current risk is determined based on analyzing the real-time data. Whether there is a potential risk is determined based on the patient graph and a classification method. Recommendation candidates are generated based on features of other patients meeting a similarity condition with the patient using the patient graph and a hierarchy search. The method has applications including, but not limited to, use cases in medical AI / healthcare, for example to optimize predictions or treatments, or to support decision-making.

Description

LIFESTYLE RECOMMENDATION SYSTEM FOR CHRONIC DISEASE MITIGATIONCROSS-REFERENCE TO PRIOR APPLICATION

[0001] Priority is claimed to U.S. Provisional Application Serial No. 63 / 602,690 filed on November 27, 2023, the entire contents of which is hereby incorporated by reference herein. FIELD

[0002] The present invention relates to artificial intelligence (Al) and machine learning (ML), and in particular to explainable Al (XAI) and a recommendation system implementing XAI, and to a method, system, data structure, computer program product and computer-readable medium for predicting and explaining lifestyle recommendations for mitigation of a medical condition, such as a chronic disease like diabetes.BACKGROUND

[0003] There are a number of chronic diseases that patients have to live with their entire lives, and consequently would like to mitigate and reduce the symptoms, causes or effects the diseases have on their lives as much as possible. Within each category of such diseases, timely diagnosis, education of patients in self-management, and continuous medical care are essential to prevent acute complications and minimize the risk of long-term complications. For instance, in the case of migraines, it would be beneficial to be able to identify the trigger in order to prevent future episodes. For psoriasis, it would be beneficial to understand what causes the skin condition to worsen. Management of diabetes is even more complicated, and therapeutic decisions require consideration of diverse medical factors and lifestyle choices that, if optimized, would be able to improve the quality of life of diabetic patients.

[0004] The trend of self-management for diabetes is on the rise, and Al-based clinical decision support is proving beneficial for both patients and healthcare professionals involved in diabetes care. The entire paradigm of diabetes management has been transformed due to the integration of new technologies such as continuous glucose monitoring (CGM) devices. Al has been shown to provide useful management tools to deal with these incremental repositories of data and enhancing patient’s commitment to various demanding self-care behaviors that are often very burdensome: carefully scheduling meals, counting carbohydrates, exercising, monitoring blood glucose levels, and adjusting endeavors on a daily basis (see Makroum, M.A.; Adda, M.; Bouzouane, A.; and Ibrahim, H. Machine Learning and Smart Devices for Diabetes Management: Systematic Review. Sensors (Basel). 2022 Feb 25;22(5): 1843. doi: 10.3390 / s22051843; and Contreras, I.; and Vehi I. Artificial Intelligence for Diabetes Management and Decision Support: Literature Review. I Med Internet Res. 2018 May 30;20(5):el0775. doi: 10.2196 / 10775, each of which is hereby incorporated by referenceherein). Particularly, recent studies confirm that Al-based tools and mobile apps appear to be effective interventions to help improve medication adherence in type 2 diabetes patients compared with conventional care strategies (see Shrivastava, T.P.; Goswami, S.; Gupta, R.; and Goyal, R.K. Mobile App Interventions to Improve Medication Adherence Among Type 2 Diabetes Mellitus Patients: A Systematic Review of Clinical Trials. J Diabetes Sci Technol. 2023 Mar;17(2):458-466. doi: 10.1177 / 19322968211060060, which is hereby incorporated by reference herein).SUMMARY

[0005] In an embodiment, the present invention provides a computer-implemented, machine learning method for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient. Structured information is extracted from unstructured static data. Real-time data is transformed from a monitoring device into a homogeneous representation. A portion of data provided by patients is selected from a database based on a pre-defined time interval or pre-defined frequency. A patient graph is constructed using the structured information, the homogenous representation of the real-time data, and the portion of data. Whether there is a current risk is determined based on analyzing the real-time data and one or more conditions associated with measured characteristics of the real-time data. Whether there is a potential risk is determined based on the patient graph and a classification method. Recommendation candidates are generated for the patient to mitigate the medical condition based on one or more features of other patients meeting a similarity condition with the patient using the patient graph and a hierarchy search based on determining that the current risk and / or the potential risk exist. The method has applications including, but not limited to, use cases in medical Al / healthcare, for example to optimize predictions or treatments or to support decision-making.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Embodiments of the present invention will be described in even greater detail below based on the exemplary figures. The present invention is not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or combined in different combinations in embodiments of the present invention. The features and advantages of various embodiments of the present invention will become apparent by reading the following detailed description with reference to the attached drawings which illustrate the following:

[0007] FIG. 1 schematically illustrates a method for real-time explainable lifestyle recommendations and a recommendation system for chronic disease mitigation according to an embodiment of the present invention;

[0008] FIG. 2 schematically illustrates a method for making personalized lifestyle recommendations and a personalized recommendation module according to an embodiment of the present invention; and

[0009] FIG. 3 is a block diagram of an exemplary processing system, which can be configured to perform any and all operations disclosed herein.DETAILED DESCRIPTION

[0010] Embodiments of the present invention provide an Al-based approach for providing explainable recommendations for patients to mitigate medical conditions, such as diseases or long-lasting health conditions (e.g., diabetes), and to monitor their medical condition. It offers patients suggestions and explanations on what to try next, therefore, iteratively, working with the patient to find the best possible strategy to manage their medical condition.

[0011] An embodiment of the present invention can be applied, for example, as an Al tool to support patients of diabetes to monitor their disease. Since diabetes is a lifetime disease, the focus is not on short medical treatment interventions. Rather, an embodiment of the present invention provides for a daily basic help in diabetics management and offers suggestions and explanations to the affected patients on what to try next, therefore, iteratively, working with the patient to find the best possible strategy to manage their disease. Embodiments of the present invention present a solution which provides comprehensive lifestyle recommendations while at the same time offering better accuracy, security, trustworthiness and reliability than existing technology by allowing the user to validate the efficiency of generated recommendation candidates with real-time sensor data.

[0012] In an embodiment, the present invention provides a system to improve chronic disease management (e.g., diabetes) by identifying causes of aggravated conditions and explainable suggestions on lifestyle decisions. From a high-level perspective, the system provides personalized recommendations based on two concurrent and complementary workflows. On one hand, the generation of recommendations is triggered by a reactive workflow based on real-time data. On the other hand, the process of generating suggestions can also be started by a proactive workflow which is run on a regular basis.

[0013] In a first aspect, the present invention provides a computer-implemented, machine learning method for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient. Structured information is extracted from unstructured static data. Real-time data is transformed from a monitoring device into a homogeneous representation. A portion of data provided by patients is selected from a database based on a pre-defined time interval or pre-defined frequency. A patient graph is constructed using the structured information, the homogenous representation of the real-time data, and the portion of data.Whether there is a current risk is determined based on analyzing the real-time data and one or more conditions associated with measured characteristics of the real-time data. Whether there is a potential risk is determined based on the patient graph and a classification method. Recommendation candidates are generated for the patient to mitigate the medical condition based on one or more features of other patients meeting a similarity condition with the patient using the patient graph and a hierarchy search based on determining that the current risk and / or the potential risk exist.

[0014] In a second aspect, the present invention provides the method according to the first aspect, further comprising generating output with an explanation using a large language model (LLM) and the recommendation candidates.

[0015] In a third aspect, the present invention provides the method according to the first aspect or the second aspect, wherein the output with the explanation includes intermediate results of the hierarchy search.

[0016] In a fourth aspect, the present invention provides the method according to any of the first to third aspects, wherein extracting the structured information from the unstructured static data includes using an open information extraction model, and wherein the unstructured static data includes textual data.

[0017] In a fifth aspect, the present invention provides the method according to any of the first to fourth aspects, wherein transforming the real-time data into the homogeneous representation includes using an extract, transform, and load (ETL) algorithm.

[0018] In a sixth aspect, the present invention provides the method according to any of the first to fifth aspects, wherein the database includes at least medical reports for the patients, selfassessments from the patients, laboratory test results from the patients, and time-series data from sensors associated with the patients, wherein the laboratory test results are structured data.

[0019] In a seventh aspect, the present invention provides the method according to any of the first to sixth aspects, further comprising: generating a risk alarm based on determining the current risk and / or the potential risk exist; and transmitting the risk alarm to a user associated with the real-time data and the static data.

[0020] In an eighth aspect, the present invention provides the method according to any of the first to seventh aspects, wherein determining whether there is the current risk is further based on a rule based analysis of the real-time data and whether a risk condition is met, and wherein determining whether there is the potential risk is further based on a classification method and extrapolating health status data determined periodically according to the pre-defined time interval or the pre -defined frequency.

[0021] In a ninth aspect, the present invention provides the method according to any of the first to eighth aspects, wherein constructing the patient graph includes using a certain subset of the structured information and the homogenous representation of the real-time data based on the pre-defined time interval or the pre-defined frequency.

[0022] In a tenth aspect, the present invention provides the method according to any of the first to ninth aspects, wherein the patient graph includes nodes each representing one of the patients, each node being associated with attributes including patient metadata, diagnostics / symptoms, treatment history, lifestyle data, and / or historical time-series data.

[0023] In an eleventh aspect, the present invention provides the method according to any of the first to tenth aspects, further comprising: adding a health status evaluation attribute to the attributes of the patient graph based on the pre-defined time interval or the pre-defined frequency; and validating one or more features of the recommendation candidates using at least the historical time-series data.

[0024] In a twelfth aspect, the present invention provides the method according to any of the first to eleventh aspects, wherein the hierarchy search iteratively groups selected ones of the patients based on graph density with a decreasing threshold, wherein features of a first group of patients that are closest to the patient in the patient graph are used to provide personalized recommendations to the patient and features of a second group of patients that is larger than the first group of patients are used to provide general recommendations to the patient.

[0025] In a thirteenth aspect, the present invention provides the method according to any of the first to twelfth aspects, wherein the similarity condition is based on a predetermined distance between the patients in the patient graph or sharing more attributes in common than other ones of the patients in the patient graph.

[0026] In a fourteenth aspect, the present invention provides a computer system for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient comprising one or more processors, which, alone or in combination, are configured to perform a machine learning method for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient according to any of the first to thirteenth aspects.

[0027] In a fifteenth aspect, the present invention provides a tangible, non-transitory computer-readable medium for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient which, upon being executed by one or more hardware processors, provide for execution of a machine learning method according to any of the first to thirteenth aspects.

[0028] FIG. 1 shows a high-level diagram and data flows in the system. Inputs and (intermediate) output are numbered using letters (such as l.a) and system components arenumbered using numbers (such as 1.1), while lines with a black diamond indicate a real-time data flow. The system is composed of the following main blocks:

[0029] User (1.a) 100: The process is firstly triggered by a user 100, who provides the system the input data of both static and dynamic real-time data. For example, the user 100 is a patient who has a chronic disease such as diabetes.

[0030] Static Data (l.b) 102: The static data 102 associated with a user 100 refers to the information which changes less often than on a daily basis. It includes multi-modal data. For instance, it could be textual data such as in an electronic health records (EHR) which maintains the history and treatment status of the diabetes, and / or it could be structured data such as in laboratory medical testing results which record the testing of biological indicators such as the level of HbAlc.

[0031] Monitoring Device (l.c) 104: In addition to traditional static data 102, the user 100 can provide the system with real-time sensor data via some wearable monitoring devices. For instance, diabetic patients provide real-time blood-sugar data via a continuous glucose monitor (CGM) which is placed under the skin. Via activity or sport trackers such as a FitBit or an Apple Watch, a user 100 can provide real-time exercise and movement information. Via diabetes food trackers, users 100 can provide real-time diet data. Optionally, the user 100 can provide additional input means such as a diary that includes text information. These are all examples of useful inputs for later lifestyle recommendations.

[0032] Patient Database (l.d) 106: The system creates a patient database 106. It collects and stores the following information for each patient (user 100): text data (e.g., medical reports, EHR, patient self-assessments), structured data (e.g., laboratory test results), and time-series data from sensors. Text and tabular data can be imported from a patient’s EHR and / or can be provided by each patient individually. This data is updated on a regular basis. Sensor data is received from wearable devices like glucose sensors, continuous glucose monitoring systems, artificial pancreas, blood pressure monitors, smart watches, monitoring devices 104, etc. Depending on the sensor, the data might come in real-time or be continuously updated within short periods with all updates stored in the system. Although FIG. 1 depicts just a single patient (user 100), it is to be understood that the data flows from multiple patients in the centralized patient database.

[0033] Information Extractor (1.1) 108: The information extractor module (information extractor) 108 is designed to process text data from the medical reports, electronic health records (EHR) or patient self-assessment assays and get relevant information about a patient’s condition in a structured manner. In embodiments, the information extractor module 108 may extract this information from the patient database 106. Text records may comprise information aboutlifestyle activities, meals, diet, physical exercises, and adverse events. Existing Open Information Extraction (OpenlE) models such as MILIE (see Kotnis, B.; Gashteovski, K.; Onoro Rubio, D.; Rodriguez-Tembras, V.; Shaker, A.; Takamoto, M.; Niepert, M.; and Lawrence, C. 2021. MILIE: Modular & Iterative Multilingual Open Information Extraction. arXiv:2110.08144, which is hereby incorporated by reference herein) can be utilized for this task. Information extracted from natural language text records is supplemented by the data that resulted from laboratory tests (e.g., blood tests, metabolic panels, etc.) and structured health records to get information about diagnostics, comorbidities, prescribed treatment, etc. These data can be easily converted and stored in a tabular format.

[0034] Real-time Data Extractor (1.2) 110: There are a variety of wearable devices or apps that a patient 100 might use to obtain information about current health state or physical activities, or to measure certain indicators (e.g., glucose level) in real-time. This type of data is processed using a sensor data extractor 110 which aims to transform the sensor data into homogeneous representation as the data extracted via the information extractor 108 (1.1). At the same time, the real-time measurements are provided as input to the personalized recommendation module (1.5) 112.

[0035] Time Window Filter (1.3) 114: The time window filter 114 allows the system to specify the periods of time from which data should be included, e.g. data within a defined time range or defined regular intervals. For instance, according to individual body function, a user 100 can decide a personal interval as a calculation base to predict potential risk. Alternatively, the user 100 can decide a personal preferred time range (e.g., 3 months or 1 year) as a basis to evaluate the health status of patients in the database 106 and indirectly choose the recommendation source. Further, the user 100 can decide how often the system generates a recommendation .

[0036] Knowledge Graph Constructor (1.4) 116 (l.e) 118: The knowledge graph constructor module 116 converts all the collected data into a graph 118. The graph nodes of graph 118 represent different patients, including patients from the database 106 and the user 100 herself / himself. The criteria for joining two patients (represented by nodes in the graph 118) with an edge is based on their similarity. Here, the patient similarity can be based on the similarity of their metadata and / or distance in terms of geographic location. For example, the geographic location might be mapped into a numerical representation (e.g., geographic coordinates such as Latitude and Longitude). After being mapped into a numerical representations, geographic locations which are close in physical space may translate to being closer / more similar values in their corresponding geographic coordinates. The metadata of a patient 100 includes, for example, a set of features related to the patient’s socio-demographiccontext. In embodiments the socio-demographic information can be encoded in a set of numerical / categorical attributes / features to represent information such as age, gender, education, etc., associated with a patient. This information, along with geographic location and the other information described above is stored in the attributes associated to each patient node. The features associated to pairs of patients are used to compute a similarity measure based on which the corresponding nodes should be or should not be connected with an edge in the patient graph 118. Additionally, as shown in the constructed knowledge graph (l.e) 118, each patient node (the black circles) stores the following information in the associated attributes (the white rectangle together with each node): patient metadata (e.g., stored as tabular data), diagnostics / symptoms (e.g., stored as tabular data), treatment history (e.g., stored as tabular data), lifestyle data, e.g. sport activities, meals habits (e.g., stored as tabular data), and historical time-series data (e.g., stored as tabular data or sequential data).

[0037] Personalized Recommendation Module (1.5) 112: The personalized recommendation module (personalized recommender) 112 is introduced by embodiments of the present invention to provide personalized suggestions to the patients 100. FIG. 2 shows the detailed architecture of this module 112. It receives as input information from multiple data flows to make suggestions to the current patient including: real-time sensor data (2. a) 200, extracted via real-time data extractor 110, and the constructed knowledge graph (l.e) 118 for the patient 100, constructed by the knowledge graph constructor (1.4) 116. The personalized recommendation module 112 includes a pair of sub-modules related to two concurrent and complementary workflows:1. Reactive workflow: The generation of recommendations is triggered by a reactive rulebased risk detector 202 based on real-time data 200.2. Proactive workflow: The generation of the suggestion is triggered by the proactive periodic risk predictor 204 which is run on a regular basis and forecasts the potential risk that might occur in near future.

[0038] Rule-based Risk Detector (2.1) 202 (2.b) 206: Received sensor data flow goes through the rule-based risk detector 202. This submodule 202 is configured to identify the conditions when one or more of the measured characteristics (e.g., blood glucose) are outside of the normal range or rapidly getting worse (e.g., severe hyperglycemia progresses). The rulebased algorithm generates a flag of risks 206 to the end user when the risk condition is met.

[0039] Risk Predictor (2.2) 204 (2.c) 208: The rule-base risk detector 202 detects the risk cooccurring at a moment by reading real-time data 200. On the other hand, the risk predictor 204 forecasts the potential risk 208 that might occur in the near future to warn the user. Taking the patient knowledge graph 118 as input (particularly the attributes information - patient metadata, diagnostics / symptoms, treatment history, lifestyle data, e.g. sport activities, meals habits, andhistorical time-series data), the risk predictor 204 can be either a graph-based classification method such as KBLRN (see Garcia-Duran, A.; and Niepert, M. 2017. KBLRN: End-to-End Learning of Knowledge Base Representations with Latent, Relational, and Numerical Features. arXiv: 1709:04676, which is hereby incorporated by reference herein) or a machine learning classifier such as a decision tree that extrapolates health status data based on previous observations. As a result, the risk predictor 204 will generate the forecasted risk 208. For instance, it could be a percentage which indicates the possibility of potential risk. The periodicity of the prediction depends on the timing configuration set in the time window filter module 114.

[0040] Risk Alarm Module (2.3) 210: By receiving the detected real-time risk flag 206 and risk forecasting 208, the risk alarm module 210 then forwards them as one of the outputs 120 back to user. For instance, the output 120 can be sent to a wearable device and the user 100 will hear an alarm beep or feel vibration when a risk is detected. Meanwhile, the risk alarm module 210 will forward the detected risk signal and trigger the explainable recommendation generator (2.4) 212 to provide lifestyle related countermeasure recommendations 214 to the user.

[0041] Explainable Recommendation Generator (2.4) 212 (2.d) 214: After one of the two workflows triggers the risk alarm module (2.3) 210, either for an ongoing risk (rule-based risk detector (2. 1) 202) or for a potential imminent risk (risk predictor (2.2) 204), lifestyle suggestions and explanations (e.g., recommendations 214) are generated according to the following algorithm composed of two phases. In the first phase, a hierarchy of suggestion is generated, where each level provides a set of important features to achieve an healthy status. The different levels of the hierarchy are characterized by a different level of personalization: the lowest level of the hierarchy provides the most customized suggestions (e.g. more tailored to the specific socio-demographic context of the user and of similar patients), while the top level includes the most generic possible ones. In the second phase, the proposed features are validated by monitoring the time-series data from sensors and by verifying that the feature values correlate with the historical time-series data of similar patients.

[0042] The first phase (graph-enabled bottom-up hierarchy search) according to an embodiment of the present invention comprises:1. Once the system has detected a health risk 206 or the defined time interval in the time window filter is reached (and a risk is forecasted 208), the recommendation generator 212 starts to generate recommendations 214 based on the data associated both to the user 100 and to the other patients provided by the patient knowledge graph 118.2. Starting from the patient-patient graph, a similarity criteria is defined to get a subgraph within a given time window. This would select one or multiple connected components as groupsof patients from the original graph with the patient nodes that are most similar to the current patient. For example, the criteria might consider the amount of features shared by two patients (with similar values) in the corresponding patient node attributes.3. For each patient, a health level is introduced and calculated within the time window (e.g., by considering blood glucose or lab tests done over the time frame).4. Selected patients are iteratively grouped, based on the graph density, with a decreasing threshold on the graph density.A hierarchical method is applied to each group that searches what are the best features related to the activities under the patient’s control (e.g., change in behaviors, meal, treatment, etc.) for the group to distinguish between more healthy and less healthy patients (low -level features). Excluded from consideration are the features for which there is insufficient data to decide about the best split in the following steps.Iterate over the density threshold to involve more patients and repeat step 4 for the aggregated groups. Calculate the most important features.For the final group with the lowest graph density threshold, repeat step 4 and calculate important features to distinguish between more healthy and less healthy patients (top-level features).5. The iterative procedure started with the highest possible threshold, resulting in the smallest possible group of patients similar to the current patient. The resulting best features for more healthy patients in this group (low -level features in the hierarchy) represent the most customized possible suggestion (e.g., the system recommends “swimming”, because a few closest patients share it as an important lifestyle feature and it helps them, so this feature might be also good for a new patient, and there may be a swimming pool nearby). The last iteration (lowest possible density threshold) provides a set of best features (top-level features in the hierarchy) that are instead associated to the most generic possible recommendations.

[0043] The second phase (time series validation) according to an embodiment of the present invention comprises:1. Each patient has time series sensor data recorded, such as glucose level. After the time window filter (1.3) 114, a health status evaluation based on a defined interval will be added as an attribute into the patient graph, for example, it can be a daily glucose level trend over the last thirty days.2. After finding the best features, the system continues to validate the feature with the historical time series data, for example checking if the feature values correlate with the glucose level trends of similar patients.

[0044] As a result, the best found feature or features, such as a specific sport and / or diet habit, will be output as recommendation candidates. The intermediate procedures of hierarchy searching can be directly visualized as an explanation to explain how the system generates certain recommendations. Moreover, the validation process can be output as evidence to verify why the generated recommendations 214 are trustable.

[0045] LLM Generation Module (1.6) 122: The LLM generation module 122 uses both the output of graph-enabled bottom-up hierarchy search and the time series validation algorithms from the personalized recommendation module 112 to generate the suggestions in a human- readable format. This module uses a large language model (LLM) such as ChatGPT to generate output 120 for a patient in a text form based on the identified important features. Each feature may have a positive or negative effect on the medical condition. Features with negative effects on health serve for the risk prediction, while features with positive effects on health are used to generate suggestions. For example, a suggestion could be to decrease consumption of starchy vegetables to half, and the explanation could be that 70% similar patients got improvements by eating fewer white potatoes, com, and peas (following this suggestion) for the last xx days.

[0046] Embodiments of the present invention therefore provide for improvements to Al and ML computer systems to provide enhanced functionality to more accurately and securely predict actions which can improve a condition, thereby enhancing the trustworthiness and reliability of those computer systems. Moreover, embodiments of the present invention can be practically applied to effect further improvements in a number of technical domains, such as medical and healthcare (e.g., as an Al tool for management of a medical condition, for diagnosis / treatment; for lifestyle recommendations, etc.).

[0047] An exemplary embodiment of the present invention discussed above is practically applied for the self-management for diabetes for the common user, i.e. the diabetic patient himself / herself. In this scenario, the diabetic patient interacts with the system in order to receive suggestions related, for example, to changes in the lifestyle habits to improve his / her health conditions. The suggested actions are focused on aspects which fall under the direct control of the person such as physical activities, food, etc. The bottom-up hierarchical approach not only provides a prioritization among the possible suggestions (from most personalized to more general) to guide the user in the iterative process, but also provides explanations that, together with the validation phase, make the whole system more secure, reliable and trustworthy for the final user.

[0048] In the exemplary embodiment for self-management for diabetes discussed above, the focus was on the common user, i.e. the diabetic patient himself / herself. Another exemplary embodiment can instead consider a different scenario where the main actor involved is thedoctor of a diabetic patient rather than the patient itself. Given the different type of background owned by the new actor involved, some steps related to the explainable recommendation generator (2.4) are adapted. In particular, in the exemplary embodiment for the patient, the set of most-important features (used to generate the recommendation) were constrained to be only the ones related to the activities that can be under the direct control of the patients. In this exemplary embodiment for the doctor, this constraint can be relaxed. That is, the explainable recommendation generator (2.4) can now potentially include as part of the suggestions / explanations also features which are related to medical treatments and / or medicaments. For example, by relaxing this constraint, a search conducted as described above with reference to performing a graph-enabled bottom-up hierarchy search, may return any feature as part of the set of best features to distinguish between more healthy and less healthy patients. The features related to medical treatments are included in the features (i.e., the attributes associated to each patient node) which include information related to treatment history. The doctor, based on his / her own expertise and knowledge, can double-check the proposed solution, and confirm the treatment for his patient. Furthermore, the validation phase can provide a methodology to verify the correlation of the suggested features with the real-time measurements from the patient and compare it with the historical data from similar patients, providing further evidence for the doctor.

[0049] The explainable recommendation generator (2.4) thus provides enough flexibility to be easily adapted to extend the set of possible suggestions from self-manageable changes in lifestyle to more involved actions requiring medical treatments, under the supervision of a doctor. Thus, embodiments of the present invention can also provide a system tailored for doctors assisting the patients.

[0050] In an embodiment, the present invention provides a method for making explainable recommendations for a medical condition, the method comprising the steps of: Setup a new system (cold start):1) Prepare monitoring device(s) (l.c) to collect real-time data from a user. For instance, a wearable continuous glucose monitors to record glucose level, a sport bracelet to record exercise, a food tracker app to record diet habits, etc.2) Create an information extractor (1.1) such as MILIE to extract structured information from unstructured input such as textual data.3) Create a real-time data extractor ( 1.2) such as an Extract-Transform-Load (ETL) algorithm to collect and transform data collected via monitoring devices into compatible format for later Al models. Collecting and transforming the data collected via the monitoring devices may include extracting or transforming the data into a feature space.4) Prepare a patient database (l.d) to store both static and time-series data from related patients, used as source data for later recommendations.5) Create a time window fdter (1.3) to allow defining a specific period of time or time intervals for later recommendation generation. For instance, the time window filter can be implemented as a software program which does not require to use Al. For example, the task performed by the time winder filter module is to filter / select the data base on time. As no classification or prediction task is required for this operation, a standard software program can be used for filtering or selecting.6) Create a knowledge graph constructor ( 1.4) to construct a patient knowledge graph ( 1.e) from various data sources. For instance, the knowledge graph constructor can be implemented as a pipeline which enables the extracted results of information extraction model such as Milie to construct a well-connected knowledge graph. Milie is a model that can extract structured e.g., <subject, predicate, object>, triple from text.7) Create a personalized recommendation module (1.5) to generate recommendations and corresponding explanations. In particular, the personalized recommendation module comprises three components: a. Build a rule-based risk detector (2. 1). b. Train a health risk predictor (2.2) such as a graph-based or machine learning based model. c. Build a transparent recommendation generator (2.4) with time-series validation function.8) Create a LLM-based generation module (1.6) such as Llama2 to translate system output into human readable format.When new data comes in (use, inference or application phase):1) The user (l.a) triggers the system by providing both static data (l.b) and real-time data via certain monitoring devices (l.c).2) The system generates recommendations, e.g., lifestyle or treatment, to mitigate disease by the following steps: a. The information extractor (1.1) extracts structured information from various unstructured data such as health records and medical testing reports. b. The real-time data extractor (1.2) transforms real-time data input such as sensor data of glucoses into the required representation for further data analytics. c. Based on the pre-defined time interval or frequency, the time window filter (1.3) keeps only required data from both user and patients from database. The time window filter (1.3) may be used to bound the data provided to the Risk Predictor (2.2) according to its expected timeresolution which in turn depends on the data used to train its internal graph-based or ML-based classification model. The time window filter (1.3) may be used to allow the configuration of a periodic risk prediction. d. The knowledge graph constructor (1.4) then constructs a patient knowledge graph ( 1 .e) from various data sources, which connects both the user and other patients in one graph and retains all their attributes information. e. The personalized recommendation module (1.5) first detects both current and potential risk by analyzing real-time and graph data. Once risk is detected, it will send a risk signal back to the user and further explore the best lifestyle or treatment feature as recommendation candidates. In embodiments, a recommendation may be generated purely on the real-time data analysis. The current disclosure describes providing personalized recommendations based on two concurrent and complementary workflows. f. The LLM generation module (1.6) constructs the final human-readable output (1.f) with explanation(s) based on detected lifestyle or treatment features and presents it to the user.

[0051] Embodiments of the present invention provide for the following improvements and technical advantages over existing technology:1) Automatically provides comprehensive lifestyle suggestions including all kinds of daily life activities (e.g., diet, behavior, sport) to patients based on detected real-time health risk to mitigate disease timely.2) Provides highly personalized recommendations to help patients with chronic disease mitigate illness from lifestyle changes. This is achieved by considering both individual and socio-demographic (e.g., regional) characteristics.3) Provides completely explainable recommendations by using a transparent graph-enabled bottom-up hierarchical searching method.4) Generates reliable recommendations by validating the efficiency of generated recommendation candidates with time-series information such as glucose fluctuations and overall health tendencies.

[0052] Existing Al tools for diabetes management use data from wearable devices and patient records to do blood glucose prediction, classification of glycemic status, detection of adverse events, tracking activity and nutrition. In contrast to existing tools, embodiments of the present invention provide a daily basic help in diabetics management and clear explainable suggestions in a human-understandable manner.

[0053] glUCModel is a daily-life support system that employs case-based reasoning and an integrates blood glucose prediction tool (see Hidalgo. J.I.; Maqueda, E.; Risco-Martin, J.L.;Cuesta-Infante, A.; Colmenar, J.M.; and Nobel, J. glUCModel: a monitoring and modelingsystem for chronic diseases applied to diabetes. J Biomed Inform. 2014 Apr;48: 183-92. doi: 10. 1016 / j.jbi.2013.12.015, which is hereby incorporated by reference herein). It uses a patient’s personal and medical data inside the recommender module that compares patient data with other cases, finds similarities, and creates new suggestions how a user can improve control over the diabetes and what habits to modify to improve quality of life. In contrast, embodiments of the present invention introduce a different recommendation approach enables to provide clear explanations as to why a particular suggestion was generated.

[0054] Tools developed under the METABO project (see Fioravanti, A.; Fico, G.; Arredondo, M.T.; and Leuteritz, J.P. A mobile feedback system for integrated E-health platforms to improve self-care and compliance of diabetes mellitus patients. Annu Int Conf IEEE Eng Med Biol Soc. 2011;2011:3550-3. doi: 10.1109 / IEMBS.2011.6090591, which is hereby incorporated by reference herein) involve methods for monitoring patients with metabolic disorders, prevention of future excursions, optimization of care pathways, extracting patterns via knowledge discovery, and guiding weight loss programs. The platform helps users to follow an adequate care pathway by giving medically approved advices and providing recommendations to achieve prescriptions set by their care professionals. In contrast to this platform, embodiments of the present invention provide for self-management and lifestyle support, and also provide for clear explanations and finding similar patients from a graph.

[0055] The chronic disease management system Sweetch by Sweetch Health Ltd.(<<https: / / www.sweetch.com / >>) is a fully automated, personalized mHealth platform designed to promote adherence to physical activity and weight reduction in people with prediabetes (see Everett, E.; Kane, B.; Yoo, A.; Dobs, A.; and Mathioudakis, N. A Novel Approach for Fully Automated, Personalized Health Coaching for Adults with Prediabetes: Pilot Clinical Trial. J Med Internet Res. 2018 Feb 27;20(2):e72. doi: 10.2196 / jmir.9723, which is hereby incorporated by reference herein). It uses machine learning to automatically translate raw data streams originating from the patient’s mobile and wearable devices into insights about the individual’s life habit schedule, activity patterns, driving and walking routes, surroundings, and more. In contrast to embodiments of the present invention, Sweetch cannot provide recommendations related to diabetes-induced risks, but aimed to provide better adherence to a treatment program and achieving recommended activities, among other technological deficiencies compared to embodiments of the present invention.

[0056] The 5G-smart diabetes system collects structured and unstructured data from patients’ electronic medical health records, daily life data from wearable devices and uses machine learning methods to do risk assessment from diabetes-related features and validate the predictions through blood glucose index collected by the medical devices (see Chen, M.; Yang,J.; Zhou, J.; Hao„ Y.; Zhang, J.; and Youn, C.H. 5G-Smart Diabetes: Toward Personalized Diabetes Diagnosis with Healthcare Big Data Clouds, in IEEE Communications Magazine, vol. 56, no. 4, pp. 16-23, April 2018, doi: 10.1109 / MCOM.2018.1700788, which is hereby incorporated by reference herein). The system is able to produce a therapeutic schedule for patients and provide guidance to improve their self-treatment. In contrast, embodiments of the present invention provide improvements such as enabling to consider lifestyle related features under a patient’s control, finding similar patients, and providing clear explanations for the generated suggestions.

[0057] Referring to FIG. 3, a processing system 300 can include one or more processors 302, memory 304, one or more input / output devices 306, one or more sensors 308, one or more user interfaces 310, and one or more actuators 312. Processing system 300 can be representative of each computing system disclosed herein.

[0058] Processors 302 can include one or more distinct processors, each having one or more cores. Each of the distinct processors can have the same or different structure. Processors 302 can include one or more central processing units (CPUs), one or more graphics processing units (GPUs), circuitry (e.g., application specific integrated circuits (ASICs)), digital signal processors (DSPs), and the like. Processors 302 can be mounted to a common substrate or to multiple different substrates.

[0059] Processors 302 are configured to perform a certain function, method, or operation (e.g., are configured to provide for performance of a function, method, or operation) at least when one of the one or more of the distinct processors is capable of performing operations embodying the function, method, or operation. Processors 302 can perform operations embodying the function, method, or operation by, for example, executing code (e.g., interpreting scripts) stored on memory 304 and / or trafficking data through one or more ASICs. Processors 302, and thus processing system 300, can be configured to perform, automatically, any and all functions, methods, and operations disclosed herein. Therefore, processing system 300 can be configured to implement any of (e.g., all of) the protocols, devices, mechanisms, systems, and methods described herein.

[0060] For example, when the present disclosure states that a method or device performs task “X” (or that task “X” is performed), such a statement should be understood to disclose that processing system 300 can be configured to perform task “X”. Processing system 300 is configured to perform a function, method, or operation at least when processors 302 are configured to do the same.

[0061] Memory 304 can include volatile memory, non-volatile memory, and any other medium capable of storing data. Each of the volatile memory, non-volatile memory, and anyother type of memory can include multiple different memory devices, located at multiple distinct locations and each having a different structure. Memory 304 can include remotely hosted (e.g., cloud) storage.

[0062] Examples of memory 304 include a non-transitory computer-readable media such as RAM, ROM, flash memory, EEPROM, any kind of optical storage disk such as a DVD, a Blu- Ray® disc, magnetic storage, holographic storage, a HDD, a SSD, any medium that can be used to store program code in the form of instructions or data structures, and the like. Any and all of the methods, functions, and operations described herein can be fully embodied in the form of tangible and / or non-transitory machine-readable code (e.g., interpretable scripts) saved in memory 304.

[0063] Input-output devices 306 can include any component for trafficking data such as ports, antennas (i.e., transceivers), printed conductive paths, and the like. Input-output devices 306 can enable wired communication via USB®, DisplayPort®, HDMI®, Ethernet, and the like. Input-output devices 306 can enable electronic, optical, magnetic, and holographic, communication with suitable memory 306. Input-output devices 306 can enable wireless communication via WiFi®, Bluetooth®, cellular (e.g., LTE®, CDMA®, GSM®, WiMax®, NFC®), GPS, and the like. Input-output devices 306 can include wired and / or wireless communication pathways.

[0064] Sensors 308 can capture physical measurements of environment and report the same to processors 302. User interface 310 can include displays, physical buttons, speakers, microphones, keyboards, and the like. Actuators 312 can enable processors 302 to control mechanical forces.

[0065] Processing system 300 can be distributed. For example, some components of processing system 300 can reside in a remote hosted network service (e.g., a cloud computing environment) while other components of processing system 300 can reside in a local computing system. Processing system 300 can have a modular design where certain modules include a plurality of the features / functions shown in FIG. 3. For example, I / O modules can include volatile memory and one or more processors. As another example, individual processor modules can include read-only-memory and / or local caches.

[0066] While subject matter of the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. Any statement made herein characterizing the invention is also to be considered illustrative or exemplary and not restrictive as the invention is defined by the claims. It will be understood that changes and modificationsmay be made, by those of ordinary skill in the art, within the scope of the following claims, which may include any combination of features from different embodiments described above.

[0067] The terms used in the claims should be construed to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article “a” or “the” in introducing an element should not be interpreted as being exclusive of a plurality of elements. Likewise, the recitation of “or” should be interpreted as being inclusive, such that the recitation of “A or B” is not exclusive of “A and B,” unless it is clear from the context or the foregoing description that only one of A and B is intended. Further, the recitation of “at least one of A, B and C” should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each of the listed elements A, B and C, regardless of whether A, B and C are related as categories or otherwise. Moreover, the recitation of “A, B and / or C” or “at least one of A, B or C” should be interpreted as including any singular entity from the listed elements, e.g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B and C.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient, the computer-implemented method comprising: extracting structured information from unstructured static data; transforming real-time data from a monitoring device into a homogeneous representation; selecting a portion of data provided by patients from a database based on a pre-defined time interval or pre-defined frequency; constructing a patient graph using the structured information, the homogenous representation of the real-time data, and the portion of data; determining whether there is a current risk based on analyzing the real-time data and one or more conditions associated with measured characteristics of the real-time data; determining whether there is a potential risk based on the patient graph and a classification method; and generating recommendation candidates for the patient to mitigate the medical condition based on one or more features of other patients meeting a similarity condition with the patient using the patient graph and a hierarchy search based on determining that the current risk and / or the potential risk exist.

2. The computer-implemented method according to claim 1, further comprising generating output with an explanation using a large language model (LLM) and the recommendation candidates.

3. The computer-implemented method according to claims 1 or 2, wherein the output with the explanation includes intermediate results of the hierarchy search.

4. The computer-implemented method according to claims 1 or 2, wherein extracting the structured information from the unstructured static data includes using an open information extraction model, and wherein the unstructured static data includes textual data.

5. The computer-implemented method according to any of the preceding claims, wherein transforming the real-time data into the homogenous representation includes using an extract, transform, and load (ETL) algorithm.

6. The computer-implemented method according to any of the preceding claims, wherein the database includes at least medical reports for the patients, self-assessments from the patients, laboratory test results from the patients, and time-series data from sensors associated with the patients, wherein the laboratory test results are structured data.

7. The computer-implemented method according to any of the preceding claims, further comprising: generating a risk alarm based on determining the current risk and / or the potential risk exist; and transmitting the risk alarm to a user associated with the real-time data and the static data.

8. The computer-implemented method according to any of the preceding claims, wherein determining whether there is the current risk is further based on a rule based analysis of the realtime data and whether a risk condition is met, and wherein determining whether there is the potential risk is further based on a classification method and extrapolating health status data determined periodically according to the pre-defined time interval or the pre-defined frequency.

9. The computer-implemented method according to any of the preceding claims, wherein constructing the patient graph includes using a certain subset of the structured information and the homogenous representation of the real-time data based on the pre-defined time interval or the pre-defined frequency.

10. The computer-implemented method according to any of the preceding claims, wherein the patient graph includes nodes each representing one of the patients, each node being associated with attributes including patient metadata, diagnostics / symptoms, treatment history, lifestyle data, and / or historical time-series data.

11. The computer-implemented method according to claim 9, further comprising: adding a health status evaluation attribute to the attributes of the patient graph based on the pre-defined time interval or the pre-defined frequency; and validating one or more features of the recommendation candidates using at least the historical time-series data.

12. The computer-implemented method according to any of the preceding claims, wherein the hierarchy search iteratively groups selected ones of the patients based on graph density with a decreasing threshold, wherein features of a first group of patients that are closest to the patient in the patient graph are used to provide personalized recommendations to the patient and features of a second group of patients that is larger than the first group of patients are used to provide general recommendations to the patient.

13. The computer-implemented method according to any of the preceding claims, wherein the similarity condition is based on a predetermined distance between the patients in the patient graph or sharing more attributes in common than other ones of the patients in the patient graph.

14. A computer system for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient, the computer system comprising one or more hardwareprocessors which, alone or in combination, are configured to provide for execution of the following steps: extracting structured information from unstructured static data; transforming real-time data from a monitoring device into a homogeneous representation; selecting a portion of data provided by patients from a database based on a pre-defined time interval or pre-defined frequency; constructing a patient graph using the structured information, the homogenous representation of the real-time data, and the portion of data; determining whether there is a current risk based on analyzing the real-time data and one or more conditions associated with measured characteristics of the real-time data; determining whether there is a potential risk based on the patient graph and a classification method; and generating recommendation candidates for the patient to mitigate the medical condition based on one or more features of other patients meeting a similarity condition with the patient using the patient graph and a hierarchy search based on determining that the current risk and / or the potential risk exist.

15. A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more processors, provide for generating explainable lifestyle recommendations for mitigation of a medical condition of a patient by execution of the following steps: extracting structured information from unstructured static data; transforming real-time data from a monitoring device into a homogeneous representation; selecting a portion of data provided by patients from a database based on a pre-defined time interval or pre-defined frequency; constructing a patient graph using the structured information, the homogenous representation of the real-time data, and the portion of data; determining whether there is a current risk based on analyzing the real-time data and one or more conditions associated with measured characteristics of the real-time data; determining whether there is a potential risk based on the patient graph and a classification method; and generating recommendation candidates for the patient to mitigate the medical condition based on one or more features of other patients meeting a similarity condition with the patientusing the patient graph and a hierarchy search based on determining that the current risk and / or the potential risk exist.

Images