Ai-powered system for personalized health management and prevention

Abstract

Aspects of the present invention relate to a computer-implemented method for generating a causal inference model, CIM, for assisting treatment of and optimal decision making for a medical condition. The CIM comprises: (i) an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, (ii) one or more processing layers connected to the input layer, and (iii) an output layer connected to the one or more processing layers, and configured to output an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval. The method comprises obtaining a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for a plurality of patients and a plurality of time intervals, wherein each labeled input data set comprises a label characterizing a presence of the symptom of the medical condition for a respective patient within a respective time interval; and generating the CIM for treatment of the symptom of the medical condition by training the CIM based on the obtained plurality of labeled input data sets. The method may also comprise machine learning (e.g. a large language model), configured for generating part of the labeled input data set. The present disclosure can be used in a variety of applications including, but not limited to, several anticipated use cases in medical device development, in medical diagnostics / applications and in healthcare.

Description

September 12, 2024 NEC Laboratories Europe GmbH N173008WO KAU / Bmn / Slj AI-powered system for personalized health management and prevention

[0001] 7KH^SUHVHQW^DSSOLFDWLRQ^FODLPV^EHQH¿W^RI^WKH^(XURSHDQ^3DWHQW^ $SSOLFDWLRQ^QR^^^^^^^^^^^^^^¿OHG^RQ^1RYHPEHU^^^^^^^^^^ZKLFK^LV^H[SUHVVO\^ incorporated herein in its entirety by reference. 5 Technical Field

[0002] The present disclosure is directed to artificial intelligence (AI) and machine learning (ML) technologies concerned with providing a causal model capable of identifying causal relationship between n nodes through node edges of a neural network. Application scenarios may include medical applications, such as the 0 prediction of particular symptoms of chronic diseases such as migraine, irritable bowel syndrome (IBS) or gastroesophageal reflux disease (GERD), as well as healthcare in general. Background

[0003] Neural networks used in AI systems are in principle known from the5 prior art. They usually comprise an input layer, which receives a dataset as input, one or more hidden layers, sometimes comprising convolutional layers, in which calculations are performed based on the input data, and an output layer which outputs a certain output. Such networks generally work based on a calculated correlation between the input data and the output data. As neural networks provide for a relatively0 simple and reliable method for predicting events based on a known dataset, it is desirable to put such networks to use in medical applications, for example for predicting the occurrence of symptoms of a chronic disease.

[0004] However, particularly if a neural network is to be used in a medical application, it is desirable to identify a clear causal relationship between specific input5 values and the occurrence of a symptom, rather than on a mere correlation between parameters. In other words, known neural networks oftentimes resemble a “black box” receiving an input dataset and outputting an output dataset which is calculated based on a correlation, but does not provide for a definitive causal relation between input andoutput parameters. However, because decisions and potential actions influencing the health status of a patient should generally be based on a clear causal picture of causes and effects, it is desirable to provide a neural network which is capable of identifying and displaying these causal relationships to allow a patient or medical professional to make an informed decision about a treatment or intervention. Summary

[0005] Thus, it is an object of the present invention to provide a causal inference model (CIM) for assisting treatment of a medical condition which mitigates the disadvantages of the known AI-powered solutions.

[0006] The invention is defined in the appended independent claims. Preferred embodiments are defined in the dependent claims.

[0007] In some aspects of the present disclosure, a computer-implemented method for generating a causal inference model (CIM) for assisting treatment of a medical condition is provided. The CIM comprises: (i) an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, (ii) one or more processing layers connected to the input layer, and (iii) an output layer connected to the one or more processing layers, and configured to output an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval, the method comprising: obtaining a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for a plurality of patients and a plurality of time intervals, wherein each labeled input data set comprises a label characterizing apresence of the symptom of the medical condition for a respective patient within a respective time interval; and generating the CIM for treatment of the symptom of the medical condition by training the CIM based on the obtained plurality of labeled input data sets.

[0008] In some aspects of the present disclosure, a method for generating a patient-specific causal inference model (CIM) for treatment of a symptom of a medical condition of a specific patient is provided. The patient-specific CIM comprises: an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, one or more processing layers connected to the input layer, and an output layer connected to the one or more processing layers and configured to output an output data structure comprising a prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval the method comprising: obtaining a CIM generated by the method of any of claims 1 to 6; and training the obtained CIM comprising: patient-specific feature preprocessing comprising: modifying the obtained CIM based on data characterizing patient-specific behavior of the specific patient; andfine-tuning the obtained CIM by retraining at least a part of the modified CIM based on a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for the specific patient and a plurality of time intervals.

[0009] In some aspects of the present disclosure, a method for generating a patient-specific behavior recommendation for treatment of a medical condition of a patient is provided. The method comprises: obtaining an input data set characterizing patient behavior within a specific time interval, and optionally, a presence of a symptom of the medical condition within the specific time interval, and, optionally, a presence of a therapeutic treatment within the specific time interval; generating the patient-specific behavior recommendation for treatment of the medical condition of the patient based on processing the input data set with a patient-specific CIM generated by a method for generating a patient-specific causal inference model(CIM) as disclosed herein.

[0010] In further aspects, the present disclosure also relates to computing devices and computer programs adapted to implement and / or carry out the various methods disclosed herein. Brief description of the drawings

[0011] Various aspects of the present disclosure are described in more detail in the following by reference to the accompanying figures.

[0012] Fig.1 illustrates a schematic view of data collection in accordance with aspects of the present disclosure.

[0013] Fig.2 illustrates a schematic view of a causal graph in accordance with aspects of the present disclosure.

[0014] Fig.3 illustrates a schematic view of the generation of personalized recommendations.

[0015] Fig.4 illustrates a flow diagram of a computer-implemented method in accordance with aspects of the present disclosure.

[0016] Fig.5 illustrates a flow diagram of a method in accordance with aspects of the present disclosure.

[0017] Fig.6 illustrates a flow diagram of a method in accordance with aspects of the present disclosure.

[0018] Fig.7 illustrates a schematic of a computing device in accordance with aspects of the present disclosure. Detailed Description of exemplary embodiments / implementations

[0019] In the following, some exemplary embodiments / implementations of the various aspects disclosed herein are described in more detail, with reference to the drawings. Naturally, the computing systems and apparatuses of the present disclosure may employ standard hardware components (e.g., a set of on-premises edge computing hardware and / or cloud-based computing resources connect to each other via conventional wired or wireless networking technology). In some implementations, application-specific hardware (e.g., circuitry for training neural networks and / or circuitry for executing a trained neural network) may also be employed. Further, such computing hardware may be configured to execute software instructions (e.g., retrieved from collocated or remote memory circuitry) to execute the computer-implemented methods discussed herein.

[0020] While specific feature combinations are described in the following paragraphs with respect to the exemplary embodiments of the present disclosure, it is to be understood that not all features of the discussed embodiments have to be present for realizing the disclosure, which is defined by the subject matter of the claims. The disclosed embodiments may be modified by combining certain features of one embodiment with one or more technically and functionally compatible features of other embodiments. Specifically, the skilled person will understand that features, components, processing steps and / or functional elements of one embodiment can be combined with technically compatible features, processing steps, components and / orfunctional elements of any other embodiment of the present disclosure as long as covered by the disclosure as specified by the appended claims.

[0021] Moreover, the various embodiments discussed herein can be implemented in hardware, software or a combination thereof. For instance, the various modules of the systems and apparatuses disclosed herein may be implemented via application specific hardware components such as application specific integrated circuits, ASICs, and / or field programmable gate arrays, FPGAs, and / or similar components and / or application specific software modules being executed on multi- purpose data and signal processing equipment such as CPUs, DSPs and / or systems on a chip, SOCs, or similar components or any combination thereof.

[0022] For instance, the various computing (sub)-systems discussed herein may be implemented, at least in part, on multi-purpose data processing equipment such as edge computing servers. Similarly, neural network training subsystems or processes discussed herein may be implemented, at least in part, on multi-purpose cloud-based data processing equipment such as a set of cloud-severs and similar technology.

[0023] Generally, neural networks are machine learning models that employ interconnected layers of nonlinear processing units to predict an output for a received input. Some neural networks include hidden layers in addition to an output layer. The output of each (hidden) layer is used as input to the next layer in the network, i.e., the next hidden layer or the output layer. Each layer of the network generates an output from a received input in accordance with current values of a respective set of network parameters (processing unit connection weights, activation function parameters, etc.).

[0024] In the following, the benefits of the methods disclosed herein are discussed in the context of predicting headaches of patients suffering from migraine. However, it will be understood that the methods are also applicable to the prediction of other chronic diseases, such as for example irritable bowel syndrome (IBS) or gastroesophageal reflux disease (GERD), as well as for diseases in general.

[0025] For example, migraine is the most prevalent neurological condition with multiple contributing factors. It manifests as recurring, incapacitating headacheepisodes, occurring either sporadically or persistently, with or without accompanying sensory disturbances, i.e., aura [1]. Migraine ranks as the second most significant contributor to global disability. It impacts approximately 15% of the population, primarily between the ages of 22 and 55, with a higher prevalence among women. This condition represents a substantial source of economic losses on a global scale [2].

[0026] Taking migraine medication before the onset of an attack, often referred to as a preemptive treatment, can be crucial for preventing the migraine from reaching its full intensity and duration while providing a faster relief. Preemptive treatment may involve non-pharmacological approaches such as stress management and dietary modifications. Previous research showed that early interventions reduce the patient’s need for excessive medication thus helps in avoiding medication overuse [3]. Forecasting migraine and identifying its triggers in different patients are challenging tasks. Therefore, patients often struggle to determine the right time to take their medications promptly. Consequently, they postpone treatment until the pain intensifies, thus potentially diminishing the effectiveness of the medication.

[0027] In recent years, electronic health diaries have gained popularity within the medical field as a means of monitoring and tracking symptoms, thereby aiding in the management and prevention of a wide range of health conditions and diseases, including migraine [3]. However, there has been a limited use of machine learning models to predict migraine attacks and pinpoint their triggers on a personalized level. For example, in [4], the researchers developed a multivariable prediction model to forecast headache attacks in the next 24-hour period based only on stress levels, while overlooking other important migraine trigger factors [3]. In another recent study [5], the scientists employed mobile phone-based headache diaries and wearable devices to automatically monitor indicators such as heart rate, peripheral skin temperature, and muscle tension. They aimed to predict migraine occurrences for the following day by using various conventional machine learning methods. However, this study had several limitations: (1) it was performed on only 18 patients; (2) the study spanned for only 295 days; (3) all machine learning models demonstrated limited generalization capability. Another similar study [9] leveraged sleep time features collected from 7 migraine patients using wrest-worn sensor to predict migraine attacks one night before it happens. To that end, two classifiers were used, Quadratic discriminant analysis (QDA)and linear discriminant analysis (LDA), and the reported balanced accuracy reached 84%, however their findings were not confirmed in a larger population.

[0028] Previous studies can either forecast future migraine attack or identify migraine triggers, however they are not able to answer questions about causal relationships or interventions, e.g., “What effect would taking drug A have on symptoms B given that they are caused by disease C?” nor counterfactual reasoning, e.g., “If I had taken drug B instead of drug A, would my symptoms caused by disease C be less severe?” [7]. Pearl and Mackenzie [8] highlighted that black-box machine learning methods such as neural networks are restricted to generating predictions and lack the capabilities of reasoning under causal representations.

[0029] Thus, according to aspects of the present disclosure, an AI-based system for personalized migraine management and prevention is provided. This allows, inter alia, for the following benefits: (i) prediction of future migraine attacks, (ii) identifying different migraine triggers for each patient and (iii) providing personalized recommendations to mitigate the risk of future headache attacks and / or reduce their intensity.

[0030] This can be achieved for example using causal inference models. Causal inference models are a variant of explainable AI or statistical methods that go beyond answering the question of What (i.e., making predictions) and aim to also answer the questions How and Why. In other words, causal inference models may help in determining not just what happens (i.e., correlations) but also why and how it happens (i.e., causations). For example, instead of just knowing that there is a correlation between hormonal imbalance and migraines, a causal inference model may explain why hormonal imbalance causes migraine by identifying the mechanisms and underlying factors. Therefore, such models are especially useful for domains that require transparency and explainability like the healthcare sector. While conventional black- box machine learning methods can be applied to static (unchanging) datasets, causal representations generally require the possibility to interact with the dataset, usually through interventions, which are changes or fixations to certain variables to reveal causal relationships. This makes causal inference models a good fit for the migraine management and prevention challenge, where preemptive treatments can be applied inreal-time. Causal interference models may utilize electronic headache diaries to provide patient-specific recommendations and to track patient’s outcomes after the interventions.

[0031] The disclosed CIM may process daily diaries collected from migraine patients in a structured (i.e., tabular) format as well as user-provided text notes in the form of electronic headache diaries.

[0032] The disclosed system may comprise two phases, namely an offline phase, and an online patient-specific phase. During the offline phase, migraine patient data may be collected and processed to train an initial causal inference model (see Fig.1). During the online phase, the disclosed system may initially perform customized feature preprocessing for each patient, then the causal inference model may be fine-tuned on each individual patient's headache records to generate personalized recommendations (or patient-specific interventions). Subsequent changes, based on the disclosed system’s personalized recommendations, may be monitored through an electronic diary collection guided by a Large Language Model (LLM), which may provide the causal inference model with the new feature values after the intervention for further recommendations (see Fig.3). The “online” and “offline” phases are explained in more detail elsewhere herein.

[0033] Fig.1 illustrates an exemplary schematic view 100 of data collection in accordance with aspects of the present disclosure.

[0034] According to Fig.1, data 110 may be collected. The data 110 may be structured (e.g. in tabular form) and may be collected by a variety of methods. Preferably, the structured data may be collected automatically, for example by wearable devices, such as a smartphone, smart watch, or augmented-reality smart glasses. These wearable devices can seamlessly track key indicators like heart rate and muscle tension, as demonstrated in previous research (refer e.g. to [6]). Furthermore, by integrating smart glasses—a novel addition to migraine management— data pertaining to various factors may be collected, including liquid and food intake, alcohol and caffeine consumption, smoking, sunlight exposure, physical activity, temperature and humidity fluctuations, noise levels, traveling / commuting distance, and more. Automating thisdata collection process may minimize the burden on patients, making their data gathering experience more effortless. Additionally or alternatively, a user may also input data manually. For example, patients may manually input information, including whether they experienced a headache on that day, the intensity of the headache, emotional and hormonal changes, sleep duration, fatigue, stress levels, medication, and more.

[0035] Moreover, the data 100 may comprise data from a database containing data of various users. For example, the database may be stored on a server, preferably in anonymous form. Moreover, the structured data may comprise datasets from literature, such as scientific publications and the like.

[0036] Furthermore, a user may maintain an electronic patient diary 140 for a medical condition. The electronic patient diary 140 may comprise data characterizing presence of the symptom of the medical condition within a time interval. For example, a migraine patient may maintain an electronic headache diary 140. In some aspects, the electronic patient diary 140 may be obtained from a computing device of a patient, such as a smartphone or personal computer. The electronic patient diary 140 may further comprise data characterizing a presence of a therapeutic treatment within the time interval, and, optionally, data characterizing patient behavior within the time interval.

[0037] In some aspects, the labeled input data set for a time interval may be generated based at least in part on data obtained from the computing device of the patient. In some aspects, the data obtained from the computing device, e.g. in form of the electronic patient diary 140, may be obtained in the form of text notes. In some aspects, the CIM may comprise a large language model (LLM) 150. The LLM 15o may be configured to generate part of the labeled input data set from the data obtained from the computing device. For example, LLM 150 may perform feature extraction 160 to extract relevant data from the electronic patient diary 140. In some aspects, the LLM may provide the extracted data to the CIM in structured form, e.g., in tabular form.

[0038] The data 100 and / or the data extracted by LLM 150 may serve as an input data set for the causal interference model (CIM).

[0039] Moreover, a label 120 for the data may be acquired. In some aspects, the label may be extracted from the data 100 and / or from the electronic patient diary. For example, the label may comprise information about a presence of a symptom of a medical condition for a respective patient. For example, the label may comprise information about the presence of headaches for a migraine patient. In some aspects, the labels may be binary, i.e., having only two forms, e.g. “headache” and “non- headache”.

[0040] Based on the data 100, the data extracted by the LLM 150 from the electronic patient diaries 140, and based on the labels 120, the CIM may be trained. In some aspects, this phase of training may be performed “offline”. “Offline” may refer to a phase during which the model is trained, validated, and evaluated using pre-collected data. During this phase, there may be no real-time interactions with users or live systems. The focus during the offline phase may be on developing and refining the model to ensure it performs well before deploying it in a live, user-interactive environment. The trained CIM model may provide a causal graph, as for example illustrated in Fig.2.

[0041] A specific example method of training the CIM is described elsewhere herein.

[0042] As discussed above, Fig.1 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.1.

[0043] Fig.2 illustrates an exemplary schematic view of a causal graph 200 in accordance with aspects of the present disclosure. The causal graph 200 may for example be generated and / or trained based on a CIM trained in accordance with Fig.1.

[0044] The causal graph 200 may comprise nodes 210, 220, 230 and edges 215. In a first node, a feature vector 210 may be represented in the graph. The feature vector may represent various features, such as weather properties (e.g. weather changes), consumption of substances (e.g., alcohol, nicotine, caffeine, etc.), medication, etc.. This feature vector may correspond to data 100 and / or data extracted from the electronic patient diaries, as described with respect to Fig.1.

[0045] The graph 200 may further comprise a node representing a binary label 230, characterizing the presence of a symptom. For example, the binary label may indicate whether a symptom was present on a particular day. With reference to the example pertaining to migraine, the binary label 230 may indicate whether it was a headache day or a headache-free day.

[0046] Furthermore, the graph 200 may comprise a node characterizing a treatment 220. For example, the treatment 200 may characterize whether a particular medication was taken by the patient on a particular day, or whether the patient performed a lifestyle intervention (e.g. sports activity, intake of a particular amount of water, dietary changes, stress reduction, etc.).

[0047] In this way, the causal interference model may identify and display causal relationships between these nodes 210, 220, 230 through edges 215. The edges may highlight potential correlations (undirected) and / or initial causal connections (directed) and may thus offer a comprehensive overview of the relationships between features and outcome within the causal interference model.

[0048] As discussed above, Fig.2 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.2.

[0049] Fig.3 illustrates an exemplary schematic view of the generation of personalized recommendations for a particular patient.

[0050] According to Fig.3, the causal interference model (e.g., the CIM as described with reference to Figs.1 and 2) may be further trained based on the data referring to an individual patient. The training procedure may be similar as the training procedures described elsewhere herein.

[0051] Thus, the input data for the CIM may also comprise data 310, labels (e.g. binary labels as described elsewhere herein) 320, electronic patient diaries 340, on which feature extraction 360 may be performed by an LLM 350.

[0052] Moreover, the method of Fig.3 may further comprise patient-specific feature pre-processing 315. For example, features (e.g., of the feature vector displayed in Fig.2) which are irrelevant for the specific patient may be deleted. For example, for a non-smoker, the feature of nicotine intake can be removed, or for a non-alcoholic drinker, the feature of alcohol consumption can be removed.

[0053] Moreover, all data inputted into the CIM for finetuning may be filtered by a specific timeframe, i.e., only a specific number of days before the current date. For example, only a time period of the previous 3-8 days may be analyzed, because he effect of features lying further back might only have very limited influence on the prediction of a symptom on a present day.

[0054] Then, the CIM may be finetuned based on the individual patient data comprising a time-series of multiple instances representing the patient’s daily headache records.

[0055] The causal inference model 330 provided with the patient-specific features may return various different outputs (or outcomes) 370, which may be used in the analysis. Firstly, the CIM may output a probability of whether a patient will experience an occurrence of a specific symptom (e.g. a migraine attack) based on the data collected during the analyzed time period 325 (e.g., during the last 3-8 days), similar to previous predictive models which forecast potential symptoms, such as migraine attacks.

[0056] A second output of the CIM may be the graph structure of the CIM which may allow for the identification of causal relationships leading to the symptom, e.g. the migraine attack. For example, patient-specific trigger factors for the observed symptom may be identified based thereon. Hence, the CIM may provide explanations similar to those provided by other “eXplainable Artificial Intelligence” (XAI) models.

[0057] A third output and primary advantage of causal inference models may be a set of variables the model may identify as candidates for either being held set to a certain value or experimented with. Such feature values can be dentified using the do- operator (refer to Equation 1 above) and serve as potential interventions that may befurther processed and filtered to output one or more recommendations for the patient to manage his / her health condition, e.g., his / her migraine.

[0058] Thus, the CIM may generate and refine a set of personalized recommendations (or patient-specific interventions) that are identified based on patient-specific trigger factors. In other words, to mitigate the risk of future migraine attacks, the CIM may automatically identify and suggest personalized interventive strategies (patient-specific recommendations) based on the inferred causal relationships between different features and the output label. For example, if missing a meal and high coffee intake cause a certain patient to develop a migraine attack, the CIM may recommend the patient not to miss a meal and to reduce coffee consumption.

[0059] Additionally, the CIM may filter out impractical or unsafe recommendations by utilizing predefined validity tests extracted from various publicly available health recommendations and guidelines databases 380. Thus, the safety of use of the system may be improved, as objectively unsafe or impractical recommendations can be filtered out.

[0060] The personalized recommendation 390 for the particular patient may be used as an input to the trained LLM model to direct the process of future patient diary collection (feedback loop). After a period of length ^ days, the system may analyze the effects of interventions on patient outcomes in terms of the number of days on which symptoms occurred and intensity of the symptoms before and after the intervention. The process of generating personalized recommendations can be repeated multiple times in a dynamic fashion to reduce the likelihood of occurrence for the respective symptom. Fig.3 may illustrate an “online” phase of the CIM system. The online phase may refer to the period during which the trained model is deployed in a live environment and interacts with real-time data and user inputs. In this phase, the model may be actively used to make predictions, provide recommendations, or perform other decision-making tasks based on newly, incoming data.

[0061] As discussed above, Fig.3 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.3.

[0062] Fig.4 illustrates a flow diagram of an example computer-implemented method in accordance with aspects of the present disclosure.

[0063] According to Fig.4, a computer-implemented method 400 for generating a causal interference model (CIM) is shown.

[0064] The CIM may comprise: (i) an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, (ii) one or more processing layers connected to the input layer, and (iii) an output layer connected to the one or more processing layers, and configured to output an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval.

[0065] For example, the input data set may comprise data 100 and data extracted from the electronic patient diary 140, as described with reference to Fig.1.

[0066] Data characterizing “a presence of a therapeutic treatment” may generally pertain to information whether a therapeutic treatment is present, i.e., the data may specify either that a therapeutic treatment was / is present or that no therapeutic treatment was / is present. Additionally or alternatively, the data may specify that it is unknown whether or not a therapeutic treatment was / is present. If a therapeutic treatment was / is present, the data may further specify which type of therapeutic treatment was / is present.

[0067] For example, the output data structure may comprise outputs 370 as described with reference to Fig.3. The recommendation for an interventive treatment may correspond to the recommendation 390 as described with reference to Fig.3.

[0068] In some aspects, the output data structure of the CIM may comprise a causal graph data structure comprising nodes and edges. The nodes may represent a patient behavior, a presence of the therapeutic treatment, and / or a binary outcome label indicating presence of the symptom of the medical condition. The edges may indicate potential correlations and / or directed causal relations between the nodes.

[0069] In some aspects, the CIM may further comprise a large language model (LLM) configured for generating part of the labeled input data set from the data obtained from the computing device. The LLM may comprise general publicly available open source models such as Llama2 (https: / / arxiv.org / abs / 2307.09288), or models such as Chat GPT-4 or Gemini.

[0070] The method 400 may comprise the step of obtaining (410) a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for a plurality of patients and a plurality of time intervals. Each labeled input data set may comprise a label characterizing a presence of the symptom of the medical condition for a respective patient within a respective time interval.

[0071] In some aspects, the obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals may comprise obtaining, from a sensor device of a patient, sensor data characterizing patient behavior within a time interval of the plurality of time intervals.

[0072] For example, the sensor data may be retrieved from a wearable device, such as a smartphone, a smart watch, or other biometric sensor devices.

[0073] This allows for a simple and effective way of acquiring data, which requires little to no effort of the patient. Moreover, data collected based on sensors may be more accurate than data entered manually.

[0074] In some aspects, the obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals may further comprise generating a labeled input data set for the time interval, based at least in part on the obtained sensor data.

[0075] In some aspects, obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals may comprise obtaining, from a computing device of a patient, and preferably in form of an electronic patient diary for the medical condition, data characterizing presence of the symptom of the medical condition within a time interval of the plurality of time intervals, and, optionally, presence of a therapeutic treatment within the time interval, and, optionally, data characterizing patient behavior within the time interval.

[0076] The electronic patient diary may correspond to the electronic patient diary 140 or 340 as discussed with reference to Figs.1 and 3, respectively.

[0077] In some aspects, obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals may further comprise generating a labeled input data set for the time interval, based at least in part on the data obtained from the computing device of the patient.

[0078] In some aspects, the computing device may correspond to or be part of a wearable device of the patient as discussed elsewhere herein.

[0079] In some aspects, the data characterizing patient behavior may comprise data characterizing one or more environmental or physiologic parameters associated with the patient behavior, wherein, optionally, the data characterizing the one or more environmental or physiologic parameters comprise one or more of: x data obtained from a wearable sensor device of the patient; x demographic data associated with the patient; x weather data for a location of the patient;x environmental temperature and / or humidity data; x environmental noise data; x sunlight exposure data for the patient; x movement data for the patient.

[0080] The demographic data associated with the patient may comprise data corresponding to sex, age, height, weight, ethnicity / race, residence type (urban / rural), disability status, etc. of the patient.

[0081] Weather data for a location of the patients may correspond to publicly available weather data associated with a location of the patient. The location may be retrieved e.g. based on a GPS sensor of a wearable device of the patient. Alternatively, the patient may provide his or her location manually. The weather data may be accessed via a database, e.g., over the world wide web. Additionally or alternatively, a patient may enter and / or modify the weather data manually.

[0082] Environmental temperature and / or humidity data may be measured by a wearable device of the user. Additionally or alternatively, the environmental temperature and / or humidity data may be entered and / or modified by the patient manually.

[0083] Similarly, the environmental noise data, the sunlight exposure data and the movement data can be measured by a sensor of the computing device and / or may be entered and / or modified by the patient manually.

[0084] According to Fig.4, the method 400 may further comprise generating (420) the CIM for treatment of the symptom of the medical condition by training the CIM based on the obtained plurality of labeled input data sets.

[0085] The method 400 of Fig.4, and particularly the use of causal machine learning methods, allows inherent explainability of the model, thus offering the user (ormedical practitioner) explanations about why a certain decision was brought by the invention.

[0086] Thus, the method 400 may provide a multi factorial AI-based system that leverages multiple machine learning components (e.g., causal interference modeling and LLMs) to investigate all possible triggers of a symptom and predict future occurrences thereof. To increase effectiveness, the proposed system is trained on pre- existing datasets then fine-tuned on patient-specific data, es described for example with respect to Figs.3 and 5.

[0087] Hence, it may be possible to investigate triggers for the occurrence of a particular symptom, e.g., a migraine attack without the need of extensive experimenting with the patient.

[0088] As discussed above, Fig.4 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.4.

[0089] Fig.5 illustrates a flow diagram of an example method 500 in accordance with aspects of the present disclosure.

[0090] According to Fig.5, a method 500 for generating a patient-specific causal inference model (CIM) for treatment of a symptom of a medical condition of a specific patient is described. The patient-specific CIM may comprise an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, one or more processing layers connected to the input layer, and an output layer connected to the one or more processing layers. The CIM may correspond to a CIM as described elsewhere herein.

[0091] According to Fig.5, the method 500 may comprise the step of obtaining 510 a CIM. The CIM may be generated based on a method described herein, e.g. a method as described with respect to Figs.1 and 4.

[0092] In some aspects, the obtained CIM may comprise a LLM as described elsewhere herein, for example the LLM as described with reference to Fig.5. In some aspects, the method 500 may further comprise using the LLM to extract patient- specific features from an electronic symptom diary of the specific patient. In some aspects, additionally or alternatively, the method 500 may further comprise using the LLM to generate a new entry in the electronic symptom diary based on an recommendation for an interventive treatment for the medical condition in a future time interval output by the patient-specific CIM.

[0093] The use of an LLM may provide for an easily intelligible way for the patient to enter date (i.e., in natural language) which can then be interpreted by the LLM to generate the data in a structured form (e.g., in a tabular form) such that it can be processed by a CIM. Hence, the use of an LLM provides for a more intuitive use of the disclosed CIM.

[0094] According to Fig.5, the method 500 may further comprise the step of training 520 the obtained CIM. The training the obtained CIM may comprise patient- specific feature pre-processing, for example as described with reference to Fig.3.

[0095] The patient-specific feature pre-processing may comprise modifying the obtained CIM based on data characterizing patient-specific behavior of the specific patient.

[0096] The method 500 may further comprise fine-tuning 530 the obtained CIM by retraining at least a part of the modified CIM based on a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for the specific patient and a plurality of time intervals. For re-training the CIM, any training method as described elsewhere herein may be utilized.

[0097] Thus, the model trained based on the data referring to multiple patients may be personalized to the patient currently using it. Hence, the model, its predictions, recommendations and the identified triggers may become more accurate.

[0098] In some aspects, fine-tuning the obtained CIM may comprise obtaining a labeled input data set characterizing patient behavior, presence of the symptom of the medical condition, and, presence of the therapeutic treatment for the specific patient in a specific time interval.

[0099] Data characterizing “a presence of a therapeutic treatment” may generally pertain to information whether a therapeutic treatment is present, i.e., the data may specify either that a therapeutic treatment was / is present or that no therapeutic treatment was / is present. Additionally or alternatively, the data may specify that it is unknown whether or not a therapeutic treatment was / is present. If a therapeutic treatment was / is present, the data may further specify which type of therapeutic treatment was / is present.

[0100] Thus, the method 500 of Fig.5, and particularly the use of causal machine learning methods, allows inherent explainability of the model, thus offering the user (or medical practitioner) explanations about why a certain decision was brought by the invention.

[0101] Thus, the method 500 may provide a multi factorial AI-based system that leverages multiple machine learning components (e.g., causal interference modeling and LLMs) to investigate all possible triggers of a symptom and predict future occurrences thereof. To increase effectiveness, the proposed system is trained on pre- existing datasets, as described e.g. with reference to Figs.1 and 4, and then fine-tuned on patient-specific data.

[0102] Hence, it may be possible to investigate triggers for the occurrence of a particular symptom, e.g., a migraine attack without the need of extensive experimenting with the patient.

[0103] Thus, the method described with reference to Fig.5 provides for an improved way of tracking a patient’s symptoms in the electronic patient diary guided by the LLM. Moreover, the feedback in the form of patient-specific recommendations allow gradually enhancing a patient’s health condition via an interventive approach. Thus, the overall health condition of a patient may be improved. Furthermore,necessary medications, such as pharmaceutical products, can be consumed by the patient early, as soon as the potential occurrence of a symptom is predicted, and even already before the occurrence of the symptom, such that the medication can be more effective, and in total less medication must be taken by the patient.

[0104] Particularly in the case of migraines, where patients must often take medications with strong side effects, such as triptan, it is especially desirable to lower the necessary dosage of intake. Moreover, typical migraine medications, such as triptan, are most effective if taken early on, and are most effective if taken before a migraine attack starts. Thus, the “early warning system” according to the present disclosure allows for a more effective use of migraine medication and may enhance the overall health status of migraine patients.

[0105] As discussed above, Fig.5 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.5.

[0106] Fig.6 illustrates a flow diagram of an example method 600 in accordance with aspects of the present disclosure.

[0107] According to Fig.6, an example method 600 for generating a patient- specific behavior recommendation for treatment of a medical condition of a patient is described.

[0108] According to Fig.6, the method 600 may comprise obtaining 610 an input data set characterizing patient behavior within a specific time interval, and optionally, a presence of a symptom of the medical condition within the specific time interval, and, optionally, a presence of a therapeutic treatment within the specific time interval.

[0109] According to Fig.6, the method 600 may further comprise generating 620 the patient-specific behavior recommendation for treatment of the medical condition of the patient based on processing the input data set with a patient-specific CIM generated by a method as described elsewhere herein, for example with reference to Fig.5.

[0110] In some aspects, the method may further comprise outputting an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval based on processing the input data set with a patient-specific CIM generated by a method described elsewhere herein, for example with reference to Fig.5. In some aspects, the method may further comprise outputting information of patient behavior correlated or causally associated with the predicted likelihood based on processing the input data set with a patient- specific CIM generated by a method described elsewhere herein, for example with reference to Fig.5.

[0111] In some aspects, generating the patient-specific behavior recommendation may comprise outputting the patient-specific behavior recommendation. In some aspects, generating the patient-specific behavior recommendation for the patient may comprise obtaining data characterizing general treatment guidelines for the medical condition from a database. In some aspects, generating the patient-specific behavior recommendation for the patient may comprise modifying, particularly filtering, the patient-specific behavior recommendation for the patient based on the data characterizing the general treatment guidelines for the medical condition. For example, the general treatment guidelines may correspond to or comprise the health recommendations and guidelines databases 380 as described with respect to Fig.3.

[0112] This may increase safety of the described system, as potentially unsafe recommendations can be filtered based on trustworthy, scientifically proven data.

[0113] In some aspects, method 600 may further comprise generating a patient data collection recommendation based on the patient-specific behavior recommendation for the patient.

[0114] In some aspects, the method 600 may further comprise generating a labeled input data set correlating patient behavior corresponding to the patient-specific behavior recommendation and a presence of the symptom of the medical condition within a time interval subsequent to the specific time interval.

[0115] Thus, the method described with reference to Fig.6 provides for an improved way of tracking a patient’s symptoms in the electronic patient diary guided by the LLM. Moreover, the feedback in the form of patient-specific recommendations allow gradually enhancing a patient’s health condition via an interventive approach. Thus, the overall health condition of a patient may be improved. Furthermore, necessary medications, such as pharmaceutical products, can be consumed by the patient early, as soon as the potential occurrence of a symptom is predicted, such that the medication can be more effective, and in total less medication must be taken by the patient.

[0116] As discussed above, Fig.6 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.6.

[0117] Fig.7 is a functional block diagram of a computing device 700 adapted for carrying out the methods disclosed herein. The computing device 700 may comprise processing circuitry 710 such as one or more processors operably connected to memory storing code comprising instructions for carrying out the methods disclosed herein, e.g., when these instructions are being executed by the processing circuitry 710. The computing device 700 may further comprise one or more communication interfaces 730 to obtain input data needed for executing the methods discussed herein (e.g., via accessing electronic data bases or by receiving sensor data). The communication interfaces 730 also allow to send control commands, triggers, warning messages, etc. to recipient devices for implementing some of the methods discussed above.

[0118] Thus, the computing device 700 may comprise one or more means for carrying out any method disclosed herein.

[0119] As discussed above, Fig.7 represents merely an example of aspects of the present disclosure. Other examples may depart from what is shown in Fig.7. Training procedure

[0120] An exemplary training procedure of the CIM according to embodiments disclosed herein will be described more closely in the following.

[0121] The training set may comprise datapoints which may include one or more of the following per patient and per time interval: x Tabular medical history x Environmental or physiologic parameters x Electronic symptom diary data in text form x Binary outcome label indicating presence of the symptom of the medical condition.

[0122] For each datapoint, all single features such as (gender, environmental temperature, binary outcome label, etc.) may be extracted. An LLM supports extracting patient-specific features from the electronic patient diary. Each feature may represent a node in the CIM.

[0123] The causal interference model may be based on estimating the causal effect of each feature, i.e., conditional average treatment effects (CATEs).

[0124] A migraine dataset ^ ൌ ^ {(^^ǡ ^^ǡ ^^)}^^ୀ^may be considered, where ^ may represent the total number of records, i.e., multiple patients and multiple records per patient, each entry(^^ǡ ^^ǡ ^^)א ^^, may represents a patient’s data where ^^is a feature vector, ^^is a treatment; either a drug or a lifestyle intervention or a combination of both, and ^^the outcome binary label, “symptom” or “no symptom”, e.g. “headache” or “no headache”, as shown in Fig.2. Here, the problem may be simplified to a single variable. The do-operator may be used, which serves as a mathematical representation of intervening in a system. Intervening, in this context, may refer to deliberately altering the data-generating process where certain features may be set to a certain value. The do-operator is denoted as ^^^^ ൌ ^^ which may represent the intervention of setting the variable ^ to a specific value ^. The CATE value ^^^ᇱǡ ^^ǡ ^^) may represent the expected change in the outcome for a migraine patient with features ^^, when treatment ^ᇱreplaces^^^. Using the do-operator, the causal effect of fixing a treatment variable ^ א ^ to a value ^ on outcome variable ^ א{0,1}may be defined as^[^^|^^^^^^ ൌ ^^^Ǥ This estimate may be different from this conditional expectation^[^^|^^^ ൌ ^^, because when using the do-operator, the term ^[^^|^^^^^^ ൌ ^^^^mayrepresent the effect of an external entity intervening on ^ by fixing it to a value ^Ǥ Thus, the CATE value is the difference between expected outcomes at different treatment values ^ᇱǡ ^ for a given patient feature vector ^.

[0125] To obtain the weighted, directed edges of the CIM, a conditional average treatment effects (CATEs) may be estimated as follows: Ʃ^^Ļ^^^, ^) = ^[^ | ^ = ^, ^^(^ = ^Ļ^@^í^^[^ | ^ = ^, ^^(^ = ^)]. (1)

[0126] Causal representations and interventions as described by the do- operator have a long history in the medical domain since they depict the underlining theoretical concept for randomized controlled trials (RCTs) as used in clinical trials. RCTs are used to overcome unknown biases in the participants of clinical trials. The participants are separated into two random groups (that should have similar characteristics) [X = x]. One group is given the treatment [do(T=t)], while the other group is given the placebo [do(T=t’)]. In theory, all other factors should be constant between the two groups, such that the RCT is able to measure the exact causal effect of the treatment (cleaned from other correlations and / or biases).

[0127] Thus, the do-operation may mirror the effect of a RCT in order to gain knowledge of specific causal relationships between nodes of the CIM.

[0128] The edges directed to the symptom of interest are the most interesting parameters (set Y equals the binary outcome label indicating presence of the symptom of the medical condition). However, any other edge (between any 2 nodes) may also be observed.

[0129] Prior work has offered approximations to learn CATE from data. For instance, Structured Intervention Networks

[0010] can be used. The training algorithm and the parameters of the loss-function may then be similar to the ones in Section 5.1 of

[0010] adjusted to the features of the CIM.

[0130] The outputs may be the values of the edges that indicate potential correlations and / or directed causal relations between nodes. These values may typically be set between zero (no edge) and one (perfect causal relation).

[0131] To obtain the prediction of a likelihood of a presence of a symptom of the medical condition Y with a fixed set of parameters X, one can calculate: ^[^ | ^ = ^]. (2)

[0132] Furthermore, the CIM can be explored automatically. Firstly, all nodes but the binary outcome label can be fixed. Then, one (or a small set of) nodes can be relaxed, i.e., “unfixed” and varied to observe ^[^ | ^ = ^, ^^(^ = ^Ļ^@^^ZKHUH^^Ļ^LV^WKH^variation and Y the outcome label. Finding a change of parameters T that obtain a small^[^ | ^ = ^, ^^(^ = ^Ļ^@^PD\^SURYLGH^WKH^UHFRPPHQGDWLRQ^^^(^ = ^Ļ^^IRU^DQ^LQWHUYHQWLYH^treatment for the medical condition.

[0133] The above training method is described as an example, and other training procedures for the CIM may be used without departing from the scope of the embodiments disclosed herein. References [1] Burstein, Rami, Rodrigo Noseda, and David Borsook. "Migraine: multiple processes, complex pathophysiology." Journal of Neuroscience 35.17 (2015): 6619-6629. [2] Ashina, Messoud, et al. "Migraine: epidemiology and systems of care." The Lancet 397.10283 (2021): 1485-1495. [3] Pavlovic, Jelena M., et al. "Trigger factors and premonitory features of migraine attacks: summary of studies." Headache: The Journal of Head and Face Pain 54.10 (2014): 1670-1679. [4] Houle, Timothy T., et al. "Forecasting individual headache attacks using perceived stress: development of a multivariable prediction model for persons with episodic migraine." Headache: The Journal of Head and Face Pain 57.7 (2017): 1041-1050.[5] Stubberud, Anker, et al. "Forecasting migraine with machine learning based on mobile phone diary and wearable data." Cephalalgia 43.5 (2023): 03331024231169244. [6] Park, Jeong-Wook, et al. "Analysis of trigger factors in episodic migraineurs using a smartphone headache diary applications." PloS one 11.2 (2016): e0149577. [7] Kitson, N.K., Constantinou, A.C., Guo, Z. et al. A survey of Bayesian Network structure learning. Artif Intell Rev 56, 8721–8814 (2023). [8] Pearl, J., & Mackenzie, D. (2018). The book of why: the new science of cause and effect. Basic books. [9] Siirtola, Pekka, et al. "Using sleep time data from wearable sensors for early detection of migraine attacks." Sensors 18.5 (2018): 1374.

[0010] Kaddour, Jean, et al. "Causal effect inference for structured treatments." Advances in Neuralnformation Processing Systems 34 (2021): 24841-24854.

Claims

1 September 12, 2024 NEC Laboratories Europe GmbH N173008WO KAU / Bmn / Slj Claims 1. Computer-implemented method for generating a causal inference model, CIM, for assisting treatment of a medical condition, wherein the CIM comprises: (i) an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, (ii) one or more processing layers connected to the input layer, and (iii) an output layer connected to the one or more processing layers, and configured to output an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval, the method comprising: obtaining a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for a plurality of patients and a plurality of time intervals, wherein each labeled input data set comprises a label characterizing a presence of the symptom of the medical condition for a respective patient within a respective time interval; and generating the CIM for treatment of the symptom of the medical condition by training the CIM based on the obtained plurality of labeled input data sets.

2. Method of claim 1, wherein obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals comprises: obtaining, from a sensor device of a patient, sensor data characterizing patient behavior within a time interval of the plurality of time intervals; and generating a labeled input data set for the time interval, based at least in part on the obtained sensor data.2 3. The method of claim 1 or claim 2, wherein obtaining the plurality of labeled input data sets for the plurality of patients and the plurality of time intervals comprises: obtaining, from a computing device of a patient, and preferably in form of an electronic patient diary for the medical condition, data characterizing presence of the symptom of the medical condition within a time interval of the plurality of time intervals, and, optionally, presence of a therapeutic treatment within the time interval, and, optionally, data characterizing patient behavior within the time interval; and generating a labeled input data set for the time interval, based at least in part on the data obtained from the computing device of the patient.

4. The method of any of claims 1 to 3, wherein the data characterizing patient behavior comprises data characterizing one or more environmental or physiologic parameters associated with the patient behavior, wherein, optionally, the data characterizing the one or more environmental or physiologic parameters comprise one or more of: data obtained from a wearable sensor device of the patient; demographic data associated with the patient; weather data for a location of the patient; environmental temperature and / or humidity data; environmental noise data; sunlight exposure data for the patient; movement data for the patient.

5. The method of any of claims 1 to 4, wherein the output data structure of the CIM comprises: a causal graph data structure comprising nodes and edges, wherein the nodes represent a patient behavior, a presence of the therapeutic treatment, and a binary outcome label indicating presence of the symptom of the medical condition, and wherein the edges indicate potential correlations and / or directed causal relations between the nodes.

6. The method of any of the preceding claims 3 to 5, wherein the CIM further comprises a large language model, LLM, configured for generating part of the labeled input data set from the data obtained from the computing device.3 7. A method for generating a patient-specific causal inference model, CIM, for treatment of a symptom of a medical condition of a specific patient, wherein the patient-specific CIM comprises: an input layer configured to receive an input data set characterizing patient behavior within a time interval associated with the medical condition, and a presence of a therapeutic treatment within the time interval, one or more processing layers connected to the input layer, and an output layer connected to the one or more processing layers and configured to output an output data structure comprising a prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval, information of patient behavior correlated or causally associated with the predicted likelihood, and a recommendation for an interventive treatment for the medical condition in the future time interval, the method comprising: obtaining a CIM generated by the method of any of claims 1 to 6; and training the obtained CIM comprising: patient-specific feature preprocessing comprising: modifying the obtained CIM based on data characterizing patient-specific behavior of the specific patient; and fine-tuning the obtained CIM by retraining at least a part of the modified CIM based on a plurality of labeled input data sets characterizing patient behavior, and presence of a therapeutic treatment for the specific patient and a plurality of time intervals.

8. The method of claim 7, wherein fine-tuning the obtained CIM comprises: obtaining a labeled input data set characterizing patient behavior, presence of the symptom of the medical condition, and, presence of the therapeutic treatment for the specific patient in a specific time interval.

9. The method of claim 7 or 8, wherein the obtained CIM comprises the LLM of claim 6 and the method further comprising: using the LLM to extract patient-specific features from an electronic symptom diary of the specific patient and / or4 using the LLM to generate a new entry in the electronic symptom diary based on a recommendation for an interventive treatment for the medical condition in a future time interval output by the patient-specific CIM.

10. A method for generating a patient-specific behavior recommendation for treatment of a medical condition of a patient, the method comprising: obtaining an input data set characterizing patient behavior within a specific time interval, and optionally, a presence of a symptom of the medical condition within the specific time interval, and, optionally, a presence of a therapeutic treatment within the specific time interval; generating the patient-specific behavior recommendation for treatment of the medical condition of the patient based on processing the input data set using a patient- specific CIM, wherein the patient-specific CIM is generated by the method of any of claims 7 to 9.

11. The method of claim 10, further comprising: outputting an output data structure comprising a patient-specific prediction of a likelihood of a presence of a symptom of the medical condition within a future time interval based on processing the input data set with a patient-specific CIM generated by the method of any of claims 7 to 9; and / or outputting information of patient behavior correlated or causally associated with the predicted likelihood based on processing the input data set with a patient- specific CIM generated by the method of any of claims 7 to 9.

12. The method of any of claims 10 to 11, wherein generating the patient-specific behavior recommendation for the patient comprises: outputting the patient-specific behavior recommendation; obtaining data characterizing general treatment guidelines for the medical condition from a database; and modifying, particularly filtering, the patient-specific behavior recommendation for the patient based on the data characterizing the general treatment guidelines for the medical condition.

13. The method of any of claims 10 to 12, further comprising5 generating a patient data collection recommendation based on the patient- specific behavior recommendation for the patient.

14. The method of any of claims 10 to 13, further comprising generating a labeled input data set correlating patient behavior corresponding to the patient-specific behavior recommendation and a presence of the symptom of the medical condition within a time interval subsequent to the specific time interval.

15. Computing device or system comprising means for carrying out the method of any of claims 1 to 6, any of claims 7 to 9, or any of claims 10 to 13.

16. Computer program comprising instructions for carrying out the method of any of claims 1 to 6, any of claims 7 to 9, or any of claims 10 to 14, when being executed on processing circuitry of a computing device or system.

Images