Deep Learning Method and Apparatus for Mental Disorder Prediction Using Expert Decision-Making Mechanism Modeling Based on Behavioral Data

Abstract

A deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation according to one feature of the present disclosure comprises: a step of generating a time-series behavioral data sequence by arranging behavioral data based on location movement information collected through a user device in chronological order; a step of converting behavioral changes at each time point into a time-series hidden vector reflecting contextual information through a bidirectional recurrent neural network using the time-series behavioral data sequence as input; a step of calculating a time-series concentration value by applying an importance concentration mechanism to the hidden vector to calculate weights so as to automatically identify and emphasize time intervals important for predicting mental illness; and a step of predicting the user's risk of mental illness based on a time-series representation reflecting the concentration value.

Description

Technology Field

[0001] The present invention relates to a deep learning technology for predicting mental illness based on behavioral data, and more specifically, to an artificial intelligence-based method and apparatus for predicting mental illness that simulates the decision-making mechanism of experts and provides explainability and interpretability of prediction results by analyzing movement-based digital biomarkers in a time-series form and applying a recurrent neural network and an importance concentration mechanism. Background Technology

[0003] Mental illness, particularly depression, is known to be a complex condition influenced by various behavioral, physiological, and environmental factors, and the demand for technologies to detect and predict it early is continuously increasing.

[0004] Recently, there have been active attempts to predict mental illnesses, including depression, by utilizing digital biomarkers such as movement information and usage patterns that are automatically collected from smartphones, wearable devices, etc.

[0005] Existing technologies predicted mental illness by applying statistical models or machine learning algorithms based on behavioral data such as users' GPS logs, call patterns, and app usage history.

[0006] However, these conventional technologies have limitations in that they cannot sufficiently reflect contextual information regarding the flow of time.

[0007] Furthermore, in conventional technology, the decision mechanism of the prediction model operates as a black box, which makes it difficult for medical professionals to interpret and trust the prediction results. This acts as a significant limitation, particularly in clinical decision-making environments where diagnostic responsibility is involved.

[0008] Furthermore, in conventional technologies, existing methods derived the relationship between behavioral data and mental illness based on static analysis, and there is a lack of quantitative and explainable analysis regarding how behavioral changes at critical points over time contribute to the prediction of mental illness.

[0009] Accordingly, there is a need for technology that can predict mental illness by analyzing behavior-based digital biomarkers in a time-series manner and identifying important behavioral segments in a manner similar to the judgment flow of mental health professionals, and provide the possibility of interpreting the results. Prior art literature

[0011] Korean Patent Publication No. 10-2344532 (Registration Date: Dec. 23, 2021) The problem to be solved

[0012] To solve the problems of the prior art described above, one embodiment of the present invention provides a deep learning method and apparatus for predicting mental illness based on behavioral data, which can analyze behavioral data such as user location movement information in a time series form and effectively reflect contextual information of behavioral changes over time.

[0013] One embodiment of the present invention provides a behavioral data-based deep learning method and apparatus for predicting mental illness, which can emphasize important points in time or behavioral patterns for predicting mental illness by applying an importance concentration mechanism (attention mechanism) to a deep learning-based prediction model, and provide in an explainable manner which time interval the prediction result is based on behavioral changes.

[0014] One embodiment of the present invention provides a deep learning method and apparatus for predicting mental illness based on behavioral data, which can provide a highly reliable artificial intelligence-based mental illness prediction system that can be practically utilized in the medical field by mimicking the decision-making mechanism of an expert, allowing the deep learning model to self-learn a judgment structure similar to the point in time that actual clinical experts consider important during the judgment process, and providing the judgment flow in the form of a digital profile.

[0015] One embodiment of the present invention provides a behavioral data-based deep learning method and apparatus for predicting mental illness that can automatically generate and provide a digital mental health profile that can be understood by both users and experts by visualizing not only the prediction results but also the characteristics of behavioral data by time interval that contributed to the prediction judgment.

[0016] However, the technical problems that the present disclosure aims to solve are not limited to the technical problems described above, and other technical problems may exist. means of solving the problem

[65535] One technical aspect of the present invention proposes a deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation. The method comprises the steps of: generating a time-series behavioral data sequence by arranging behavioral data based on location movement information collected through a user device in chronological order; converting behavioral changes at each time point into time-series hidden vectors reflecting contextual information using a recurrent neural network with the time-series behavioral data sequence as input; calculating time-series concentration values ​​by applying an importance concentration mechanism to the hidden vectors to calculate weights that automatically identify and emphasize time intervals important for predicting mental illness; and predicting the user's risk of mental illness based on a time-series representation reflecting the concentration values. The method includes the step of generating a user-customized digital mental health profile configured to enable visual interpretation of behavioral patterns that contributed to the prediction of mental illness risk by time intervals by utilizing the user's mental illness risk and the concentration value by time interval together. In one embodiment, the step of generating the user-customized digital mental health profile may include: configuring a two-dimensional matrix having a time axis and a behavioral feature axis using the concentration value by time interval and a plurality of behavioral features included in the time-series behavioral data sequence; placing a prediction contribution calculated based on the concentration value by time interval and the behavioral feature value of the corresponding time point as a cell value corresponding to each time point of the time axis and each behavioral feature of the behavioral feature axis in the two-dimensional matrix; generating visualization information of the digital mental health profile by visualizing the distribution of the prediction contribution according to the time axis and the behavioral feature axis in the form of a heatmap based on the two-dimensional matrix; and generating interpretation information for the distribution of the prediction contribution included in the digital mental health profile in time interval units separated into weekday intervals and weekend intervals, and including the interpretation information in the digital mental health profile.In one embodiment, the behavior data may include at least one of GPS-based travel distance, number of visited places, location diversity, travel start time, and travel end time. In one embodiment, the importance concentration mechanism may be configured to calculate an association score between a hidden vector at each time point and a global vector reflecting contextual information of the entire sequence, and to assign relative importance by time point by normalizing the concentration value calculated based on the association score using a softmax function. In one embodiment, the step of predicting the user's risk of mental illness may include a step of predicting the presence of depressive symptoms or a continuous score level of the user based on the CES-D score. In one embodiment, the time-series behavior data sequence may include behavior data for the past 14 days or more on a user-by-user basis. In one embodiment, the digital mental health profile may include data in the form of a heatmap that visually represents behavior data characteristics corresponding to concentration values ​​by time interval, and may include data that time-series explains the relationship between changes in individual activity characteristics and predicted depressive symptoms based on the difference in travel patterns between weekdays and weekends. Another technical aspect of the present invention proposes a behavior data-based deep learning device for predicting mental illness using simulation of an expert decision-making mechanism.The above device includes at least one processor and a memory for storing instructions, and when the instructions are executed by the processor, the instructions are configured to generate a time-series behavioral data sequence by arranging behavioral data based on location movement information collected through a user device in chronological order, convert the time-series behavioral data sequence into a time-series hidden vector reflecting contextual information using the time-series behavioral data sequence as input, calculate a time-series concentration value by applying an importance concentration mechanism, predict the user's mental illness risk based on a time-series representation reflecting the concentration value, and generate a user-customized digital mental health profile using the risk and the time-series concentration value. Effects of the invention

[0019] According to one embodiment of the present invention, by using time-series behavioral data based on location movement information collected through a user device to predict the risk of mental illness, and simultaneously providing the basis for the prediction results in a structured form, it has the effect of implementing an explainable-based mental health management system that goes beyond the simple calculation of risk scores. Furthermore, according to one embodiment of the present invention, a two-dimensional matrix having a time axis and a behavioral feature axis is constructed using concentration values ​​at each time point and a plurality of behavioral features included in the time-series behavioral data sequence, and by placing prediction contributions in cells corresponding to each time point and behavioral feature, the prediction results can be structured in units of combinations of time and behavioral characteristics. Accordingly, the user can intuitively understand which behavioral characteristics contributed to the risk prediction in which time interval, rather than simply checking the risk score. In addition, according to one embodiment of the present invention, by visualizing the distribution of prediction contributions in the form of a heatmap based on the two-dimensional matrix and further generating interpretation information in units of time intervals separated into weekday and weekend intervals and including it in a digital mental health profile, analysis results reflecting structural differences in the user's life rhythm can be provided. This configuration forms an explainable AI-based structure that supports a clear understanding of the timing and basis of predictive judgments, and the specification also describes the effect of supporting the understanding of the timing and basis of predictive judgments. Brief explanation of the drawing

[0020] FIG. 1 is a diagram illustrating a deep learning device for predicting mental illness according to a comparative example of the present disclosure. FIG. 2 is a flowchart illustrating a deep learning method for predicting mental illness based on behavioral data using simulation of an expert decision-making mechanism according to an embodiment of the present invention. FIG. 3 is a diagram illustrating the internal structure and sequential flow of a deep learning-based mental illness prediction model according to an embodiment of the present invention. FIG. 4 is a flowchart illustrating an application example of an attention mechanism according to an embodiment of the present invention. FIG. 5 is a diagram illustrating the components and statistical characteristics of behavior-based digital biomarkers utilized as input features for model training in an embodiment of the present invention. FIG. 6 is a table numerically comparing performance indicators between the present embodiment and a comparative embodiment of the deep learning device for predicting mental illness according to the present invention. FIG. 7 is a flowchart illustrating an example of a method for generating a digital mental health profile according to an embodiment of the present invention. FIG. 8 is a graph visualizing the difference in prediction contribution over time intervals for each user group (by level of depressive symptoms) based on attention weights calculated by the mental illness prediction deep learning device (100) of the present invention. FIG. 9 is a graph visually showing the difference in weekday and weekend behavior patterns between user groups by level of depressive symptoms, as part of a digital mental health profile generated by the mental illness prediction deep learning device (100) of the present invention for interpreting prediction results. Specific details for implementing the invention

[0021] Specific structural or functional descriptions of embodiments according to the concept of the present invention disclosed herein are provided merely for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and are not limited to the embodiments described herein.

[0022] Embodiments according to the concept of the present invention may be subject to various modifications and may take various forms; therefore, embodiments are illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes modifications, equivalents, or substitutions that fall within the spirit and scope of the present invention.

[0023] Terms such as "first" or "second" may be used to describe various components, but said components shall not be limited by said terms. For the sole purpose of distinguishing one component from another, for example, without departing from the scope of rights according to the concept of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0024] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. Conversely, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between. Expressions describing the relationships between components, such as "between," "exactly between," or "directly adjacent to," should be interpreted in the same way.

[0025] The terms used herein are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to specify the existence of the described features, numbers, steps, actions, components, parts, or combinations thereof, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0026] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification.

[0027] In this specification, the term "processor" may refer to hardware capable of performing functions and operations according to each name described in this specification, computer program code capable of performing specific functions and operations, or an electronic recording medium loaded with computer program code capable of performing specific functions and operations.

[0028] In other words, the term "processor" may refer to a functional and / or structural combination of hardware for carrying out the technical concept of the present invention and / or software for driving said hardware.

[0029] Hereinafter, embodiments will be described in detail with reference to the attached drawings. However, the scope of the patent application is not limited or restricted by these embodiments. Identical reference numerals in each drawing indicate identical components.

[0031] In the following, a deep learning method for predicting mental illness is proposed to generate prediction results for mental illnesses, such as depressive symptoms, using a recurrent neural network as a deep learning model, and to interpret them in a manner similar to the structure of an expert-level decision-making mechanism. In the following, the use of a bidirectional recurrent neural network (Bi-LSTM) as the recurrent neural network is explained as an example. However, it is not limited to this, and various recurrent neural networks, such as various recurrent neural networks (RNN) or long short-term memory networks (LSTM), can be applied.

[0032] In addition, the definitions of terms used in the present specification below are as follows.

[0033] In this specification, "attention mechanism" refers to an internal deep learning structure for quantitatively identifying and highlighting time intervals that contributed to a prediction result from a user's time-series behavior data. For example, the attention mechanism can calculate an association score based on the hidden vector and overall contextual information of each time point, and normalize this score with a softmax function to calculate an attention weight representing the relative importance of each time point.

[0034] These attention weights can be used to simulate the decision-making flow in which experts consider behavioral changes at specific time points to be more important during the diagnosis process. Additionally, they can be used as a core mechanism of explainable artificial intelligence to enable users or experts to intuitively understand the reasoning behind the judgments that contributed to the final prediction results. For example, among the time-series behavioral data of the last 14 days, high attention values ​​are assigned to specific dates where movement distance or location diversity has sharply decreased, thereby visually representing how important those dates are for the prediction judgment.

[0035] The "importance concentration mechanism" refers to an internal judgment structure designed to automatically highlight clinically significant time intervals within time-series behavioral data for the prediction of mental illness, and can calculate a concentration value that quantifies the relative importance of behavioral changes at each point in time. This concentration value is calculated by the deep learning model's internal attention algorithm and can quantitatively reflect the impact of user behavioral changes on the prediction judgment. Through this, not only the prediction results but also the behavioral characteristics by time interval that contributed to the judgment can be constructed into a digital mental health profile in visual forms such as heatmaps, allowing the judgment structure to be interpreted and explained in a manner similar to the diagnostic criteria of clinical experts.

[0036] The "focus value" or "importance focus value" is a time-series weight value calculated by the attention mechanism, which numerically represents the importance of the time interval that contributed to the prediction result. This focus value quantitatively expresses how much each time point in the time-series behavioral data influenced the prediction and can be utilized for subsequent visualization and the creation of digital mental health profiles.

[0037] "Mockup of expert decision-making mechanisms" refers to a structure designed to incorporate into a deep learning model the flow of judgment regarding key behavioral changes—such as timing, frequency, and magnitude—that clinical experts refer to when diagnosing mental illness. By simulating the expert decision-making characteristics, where behavioral changes at specific points in time are deemed more important for prediction, through the assignment of weights for each time point, it is possible to provide prediction results and explanations similar to expert-level judgment.

[0038] "Behavioral data" refers to data regarding a user's behavior, including time-series-based lifestyle habits and movement pattern data such as user location movement information, place visit history, whether the user stopped or moved by time period, distance traveled, and frequency of transition between places. In the present invention, such behavioral data is utilized to detect and predict abnormal signals of mental health.

[0039] A "digital mental health profile" refers to an output that visualizes prediction results and concentration values ​​calculated based on a user's time-series behavioral data, diagramming the main causes of abnormal behavior or predicted judgments that appeared during a specific time interval. In one embodiment of the present invention, it may be generated in the form of a heatmap or a timeline, and can be utilized by clinical experts as reference material for judgment or assist users in checking their own condition.

[0040] "Abnormal behavior patterns" are pattern data that include abrupt changes in behavior, such as abnormal reduction in movement, fixed location, and time-based imbalance, which deviate from the flow of normal daily habits, and are used as key indicators for predicting mental illness. The deep learning model of the present invention automatically detects these abnormal behavior patterns within a time series and uses them as a basis for prediction by reflecting them in concentration values.

[0041] A "time series heatmap" is a diagram that visually represents the concentration value at each point in time for behavioral data arranged in chronological order, and can distinguish and display important points in time that influenced the prediction results by color intensity. According to one embodiment of the present invention, the heatmap is provided as part of a digital mental health profile and can be utilized as a means of supporting expert judgment or user feedback.

[0043] FIG. 1 is a drawing illustrating a deep learning device for predicting mental illness according to one comparative example of the present disclosure.

[0044] A deep learning device (100) for predicting mental illness is a device that generates prediction results of mental illnesses, such as depressive symptoms, based on user behavior data provided from a user device (200), and enables the results to be interpreted in a manner similar to an expert-level decision-making mechanism structure, and is configured to include a processor (110) and a memory (120).

[0045] A deep learning device (100) for predicting mental illness calculates the risk of depression through a deep learning model including a recurrent neural network and an attention mechanism, and can express the structure of the prediction judgment in a manner similar to an expert decision-making mechanism.

[0046] A user device (200) can collect conscious or unconscious behavioral data, such as the user's movement path, location stay time, distance traveled, and activity patterns on weekdays and weekends, such as a smartphone or wearable device, and transmit it in real-time or periodically to a mental illness prediction deep learning device (100).

[0047] A deep learning device (100) for predicting mental illness can perform time-series-based behavioral characteristic analysis on individual users based on behavioral data received from a user device (200), generate a judgment basis structure similar to expert diagnostic criteria, and provide explainable prediction results and digital mental health profiles to users or experts.

[0048] The deep learning device (100) for predicting mental illness applies an attention-based importance concentration mechanism to automatically identify time points and behavioral patterns that are significant for predicting mental illness within the collected time-series behavioral data, thereby enabling advanced prediction that reflects the timing and context of behavioral changes, unlike traditional static analysis methods.

[0049] This allows for the precise determination of mental health status in near real-time from unconscious digital biomarkers, such as user movement data.

[0050] Here, a digital biomarker refers to an indicator that indirectly reflects a health or disease state based on personal behavioral, physiological, or psychological data collected automatically or manually through digital devices such as smartphones, wearable devices, and apps.

[0051] In addition, the deep learning device (100) for predicting mental illness can generate a digital mental health profile that is explainable to the user based not only on the prediction results but also on the relative contribution (concentration value) of the behavioral data at different points in time that influenced the prediction. This function can significantly improve the reliability and clinical applicability of the mental health prediction system by visualizing the judgment structure within the model in a form similar to the diagnostic mindset of an expert and providing AI-based feedback that is interpretable to both the user and medical professionals.

[0052] The deep learning device (100) for predicting mental illness can interpret and visualize the underlying structure of each judgment result based on data beyond simple prediction, so it can be usefully applied in various digital healthcare fields such as mental health monitoring, diagnostic assistance, or user feedback systems.

[0053] The deep learning device (100) for predicting mental illness performs a series of operations as described above in the processor (110), and the memory (120) performs the role of storing collected data and processing steps or storing operation programs.

[0054] The processor (110) can process data stored in memory (120). The processor (110) can execute computer-readable code (e.g., software) stored in memory (120) and instructions triggered by the processor (110).

[0055] The processor (110) may be a data processing device implemented in hardware having a circuit having a physical structure for executing desired operations. For example, the desired operations may include code or instructions included in a program.

[0056] For example, a data processing device implemented in hardware may include a microprocessor, a central processing unit, a processor core, a multi-core processor, a multiprocessor, an Application-Specific Integrated Circuit (ASIC), and a Field Programmable Gate Array (FPGA).

[0057] The memory (120) can store passive data and / or active data received from the user. For example, the memory (120) can store GPS coordinates, step count, sleep start and end times, call start and end times and / or the type of app being used and the usage time of the app received from the user's user device (102).

[0058] The memory (120) can store multiple instruction sets, such as deep learning learning model configuration information, behavior data processing logic, prediction result storage and digital profile generation logic, which the processor (110) can execute, and can also store intermediate result values ​​related to prediction results and digital mental health profiles.

[0059] The memory (120) can store instructions (or programs) executable by the processor (110). For example, the instructions may include instructions for executing the operation of the processor and / or the operation of each component of the processor.

[0060] The memory (120) may be implemented as a volatile memory device or a non-volatile memory device. The volatile memory device may be implemented as a DRAM (dynamic random access memory), SRAM (static random access memory), T-RAM (thyristor RAM), Z-RAM (zero capacitor RAM), or TTRAM (Twin Transistor RAM). Non-volatile memory devices can be implemented as EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, MRAM (Magnetic RAM), Spin-Transfer Torque (STT)-MRAM, Conductive Bridging RAM (CBRAM), FeRAM (Ferroelectric RAM), PRAM (Phase change RAM), Resistive RAM (RRAM), Nanotube RRAM, Polymer RAM (PoRAM), Nano Floating Gate Memory (NFGM), holographic memory, Molecular Electronic Memory Device, or Insulator Resistance Change Memory.

[0061] In the above description, the mental illness prediction deep learning device (100) is described as a type of computing device including a processor (110) and memory (120), but is not limited thereto and can be implemented as various electronic devices or components such as a printed circuit board (PCB) such as a motherboard, an integrated circuit (IC), or a system on chip (SoC).

[0062] Alternatively, the mental illness prediction deep learning device (100) and the user device (102) may be implemented in a PC (personal computer), a server, or a portable device. The portable device may be implemented as a laptop computer, a mobile phone, a smartphone, a tablet PC, a mobile internet device (MID), a PDA (personal digital assistant), an EDA (enterprise digital assistant), a digital still camera, a digital video camera, a PMP (portable multimedia player), a PND (personal navigation device or portable navigation device), a handheld game console, an e-book, or a smart device. The smart device may be implemented as a smart watch, a smart band, or a smart ring.

[0063] In one embodiment, the optimal behavioral data collection method performed in the mental illness prediction deep learning device (100) may be operated through a user device (200) in the form of an application or program. The application for mental illness prediction may be an application program running on a PC or a portable device.

[0065] FIG. 2 is a flowchart illustrating a deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation according to an embodiment of the present invention.

[0066] As shown in FIG. 2, the mental illness prediction deep learning device (100) can receive time-series behavior data of a user as input, predict the risk of mental illness, and perform a series of processing steps to convert the prediction result into a digital mental health profile in an explainable form.

[0067] The mental illness prediction deep learning device (100) can generate a sequence of time-series behavioral data at the user level by arranging behavioral data collected through the user device (200) in chronological order (S210).

[0068] Here, behavioral data refers to data that is automatically collected without user intervention or actively entered by the user, and includes, for example, GPS-based location movement information, distance traveled, number of visited places, variety of locations, start and end times of movement, time spent at locations, etc.

[0069] Behavioral data is a type of digital biomarker—that is, as a sub-concept—encompassing various behavioral indicators necessary for the quantitative analysis of mental health.

[0070] For example, a digital biomarker refers to a quantitative indicator that can be used to estimate or diagnose health conditions, including mental illness, based on user-related data collected automatically or semi-automatically through digital devices such as smartphones, wearable devices, and app usage logs.

[0071] These digital biomarkers can include various physiological, behavioral, and psychological factors such as heart rate, sleep duration, location information, behavioral patterns, and history of mood changes; compared to traditional biomarkers, they offer higher utility and scalability due to their non-invasive nature and the ability to be collected from everyday life environments.

[0072] In this invention, among digital biomarkers, behavioral data—specifically data related to user behavior, such as location movement information, distance traveled, place visit history, app usage time, and user responses—is utilized as a core input for predicting mental illness.

[0073] Such behavioral data can be classified into passive data and active data depending on the data collection method.

[0074] Passive data refers to data collected automatically without user intervention, such as GPS logs, distance traveled, steps taken, call logs, and app usage history. This data is based on unconscious behavioral habits and has the advantage of sensitively reflecting changes in one's mental state during daily life.

[0075] Active data refers to data collected through an active method in which users directly input information. This includes responses to daily surveys using the Experience Sampling Method (EMA), results from psychological tests such as the Big 4, and direct input of sleep duration or mood scores. Active data is based on self-reporting and provides high-level perceptual information regarding emotional states or subjective experiences.

[0076] Ultimately, behavioral data is a sub-concept of digital biomarkers and encompasses various behavioral indicators necessary for the quantitative analysis of mental health. The purpose of this invention is to structure such behavioral data into a time series and analyze it through a deep learning model to realize expert-level prediction of mental illness and explainable diagnostic assistance.

[0077] A user-unit time-series behavioral data sequence refers to a continuous data structure composed of behavioral data samples collected continuously from a single individual user over a certain period (e.g., the last 14 or 30 days) arranged in chronological order.

[0078] This sequence is at times t1, t2, … , t n Each corresponding point in time may include multiple behavioral features such as travel distance, location diversity, number of visited places, and start / end times of travel, and the behavioral state at each point in time may be expressed in a vector form and configured by formalizing the flow of the entire user's behavior into a fixed-length or variable-length time-series sequence.

[0079] By organizing behavior data into a time series on a user-by-user basis in this way, the mental illness prediction deep learning device (100) can learn the temporal changes, periodicity, and discontinuity of individual user behavior, and furthermore, it becomes possible to make customized predictions that reflect an expert decision-making mechanism structure specialized for individual users.

[0080] In one embodiment, a mental illness prediction deep learning device (100) analyzes indicators such as the user's average daily travel distance, location dwelling density, and diversity of place visits over the past 14 days on a daily basis (t1 to t 14 It can be organized into a [14 × N] time series vector sequence. This sequence is used as input for deep learning models, and the patterns of change and continuity at each time point are closely related to early signs of mental illness.

[0081] The deep learning device (100) for predicting mental illness can utilize the above-mentioned time-series behavior data sequence as input to a recurrent neural network and convert it into a hidden vector for each time point that includes the context of behavior change at each time point (S220).

[0082] Step S220 can effectively reflect time-flow information valid for diagnosing mental disorders, in accordance with the principle of expert diagnosis that user behavior should be analyzed within the context between previous and subsequent behaviors, rather than at a single point in time.

[0083] Here, a time-specific hidden vector is a vector representation that summarizes the state at a specific time point t based on the behavioral data input at that time point within the time series. The time-specific hidden vector has the characteristic of containing contextual information on the time axis that reflects information from both past and future time points.

[0084] As mentioned above, a bidirectional recurrent neural network (Bi-LSTM) can be applied as an example of a recurrent neural network. Unlike a unidirectional recurrent neural network, a bidirectional recurrent neural network (Bi-LSTM) processes the input sequence in both the past (t-1 → t) and future (t+1 → t) directions, so the hidden vector for each time step includes a more precise time step state representation that reflects the data flow of surrounding time steps.

[0085] These time-series hidden vectors reflect time-series correlations, pattern turning points, and periodicity within the behavioral data, and can subsequently be used as reference data to determine how important each time point is for predicting mental illness in the attention mechanism.

[0086] For example, in predicting depressive symptoms, hidden vectors at intervals where the amount of movement decreases sharply or where the diversity of positional movement decreases are given high importance when calculating concentration values.

[0087] The mental illness prediction deep learning device (100) can apply an attention mechanism to each of the above hidden vectors so as to automatically identify and highlight time intervals that contribute significantly to the prediction of mental illness (S230).

[0088] For example, a deep learning device (100) for predicting mental illness can calculate an association score between the hidden vector and the entire context information at each time point, and then calculate a concentration value representing the relative importance of each time point through a softmax function, etc.

[0089] These concentration values ​​serve to simulate the judgment method of actual experts who pay attention to behavioral changes over a specific period during the diagnostic process, and are a key mechanism that transforms judgment criteria within the model into a visualizable structure.

[0090] The mental illness prediction deep learning device (100) can predict the presence or severity level of a user's mental illness, particularly depressive symptoms, based on a time series representation that is emphasized by reflecting the concentration value (S240).

[0091] For example, predictions can be made based on the CES-D (Center for Epidemiologic Studies Depression Scale) score, either by determining whether it is above a threshold or by predicting a continuous score, thereby ensuring consistency with the actual clinical diagnostic flow.

[0092] The deep learning device (100) for predicting mental illness can generate and provide a user-customized digital mental health profile by utilizing the prediction results and the concentration values ​​together to visualize the behavioral patterns that contributed to the prediction of mental illness and their relative importance in an explainable form for each time interval (S240).

[0093] Unlike conventional simple numerical-based profiles, these user-customized digital mental health profiles are structured to enable users or experts to understand the relationship between behavioral characteristics and mental health status at each point in time, and can significantly improve model reliability and clinical applicability through the implementation of Explainable AI.

[0095] FIG. 3 is a diagram illustrating the internal structure and sequential flow of a deep learning-based mental illness prediction model according to an embodiment of the present invention.

[0096] In one embodiment illustrated in FIG. 3, a deep learning-based mental illness prediction model may be configured with a structure including a bidirectional recurrent neural network (Bi-LSTM) and an attention mechanism to predict CES-D or K-CESD-R scores using time-series behavior data of a user as input, and to reflect importance at each time point in the process.

[0097] Specifically, train feature data is input, and a visual example of the input dataset is displayed as the train feature data. In the train feature data, the rows represent the time axis, and the columns represent behavioral features for each time period.

[0098] For example, a total of 14 days' worth of data (rows) organized by day and 24 behavioral characteristics (N=24) measured on each day can be configured into a single sequence.

[0099] The Input node refers to the data tensor structure that the deep learning model receives as input, and can be configured in a three-dimensional form of {Batch Size, Number of Features (N), Number of Timesteps}.

[0100] A deep learning-based mental illness prediction model is composed of a bidirectional recurrent neural network (Bi-LSTM) and an attention mechanism.

[0101] The input time series data is first passed to a bidirectional LSTM (Bi-LSTM) layer, where the behavior data at each time point is converted into a hidden vector that considers the past and future behavior context together.

[0102] The output of Bi-LSTM is internally standardized through Batch Normalization, which has the effect of improving training stability and convergence speed.

[0103] Afterwards, by applying Leaky ReLU (Rectified Linear Unit) as a non-linear activation function, the loss of sensitive and fine patterns such as emotions or behavioral changes can be reduced by preserving negative information that may exist in the hidden representation at a certain rate.

[0104] According to an embodiment, the above-described structure can be iterated to form multiple layers, and the deeper the structure, the more complex behavioral patterns can be modeled.

[0105] Afterwards, attention processing is performed.

[0106] There are two paths from the final Bi-LSTM output: one path is passed directly to the Concat node, and the other is passed to the Attention algorithm to calculate attention weights that assign greater importance to the hidden vectors of time intervals contributing to the prediction. The output of the Attention algorithm is concatenated with the final Bi-LSTM output to construct the final time-series representation to be input to the prediction output layer. Through this process, the model can reflect the influence of behavioral patterns in specific time intervals on the prediction.

[0107] The merged vector is converted into a final predicted value through a dense layer. During this process, Batch Normalization and Leaky ReLU can be applied iteratively, and the output layer can finally calculate the user's CES-D or K-CESD-R score.

[0108] The CES-D or K-CESD-R score is a numerical prediction result indicating the degree of risk of depressive symptoms calculated based on the user's recent behavioral data, and it can be designed to be comparable to expert diagnostic criteria.

[0109] For example, the K-CESD-R score is a Korean version of the CESD-R score, which is a Korean translation and culturally adapted version of the Center for Epidemiologic Studies Depression Scale - Revised (CESD-R). In the present invention, the K-CESD-R score is used as the final output value of a deep learning device (100) for predicting mental illness. The K-CESD-R score can be directly predicted in a regression form based on the user's time-series behavior data, or can be classified in binary based on a specific threshold (e.g., 16 points or 20 points or more).

[0110] Through this, the present invention can secure quantitative consistency and explainability with diagnostic tools used by experts by predicting a depression scale with clinical validity based on behavioral data even in non-clinical environments.

[0112] FIG. 4 is a flowchart illustrating an application example of an attention mechanism according to one embodiment of the present invention.

[0113] The deep learning device (100) for predicting mental illness according to the present invention uses an attention mechanism to automatically identify and highlight important time intervals during the process of processing time-series behavior data of a user within the prediction model.

[0114] That is, the deep learning device (100) for predicting mental illness may include an importance concentration mechanism to automatically identify and highlight time intervals that are significant for predicting mental illness within the user's time-series behavior data.

[0115] This importance concentration mechanism can be implemented as an attention mechanism, for example, by calculating an association score between the hidden vector at each time point and a global vector reflecting the contextual information of the entire sequence, and normalizing the concentration value calculated based on the association score with a softmax function to assign relative importance to each time point.

[0116] Specifically, the attention mechanism operates by quantitatively evaluating, through learning, how much the model contributed to the prediction result for the hidden vectors generated at each time point, and assigning higher weights to relatively important times.

[0117] This attention mechanism enables the model to selectively highlight time steps within the input sequence that have a particularly significant impact on the prediction result, operating in a structure similar to the decision-making mechanism in which experts use specific behavioral change points as diagnostic grounds.

[0118] Below, each step of the attention mechanism is described with reference to Fig. 4.

[0119] First, the attention score can be calculated in the following way. The time-series hidden vector h calculated through Bi-LSTM t For (t = 1, 2, …, T), the attention score e at each time step t is calculated using the following mathematical formula 1 (S410).

[0120] [Mathematical Formula 1]

[0121]

[0122] Here, W h is the learnable weight matrix, b h is the bias vector (bias), and tanh is a non-linear activation function.

[0123] Subsequently, attention weight a representing the relative importance of each time point t It is normalized through the softmax function and calculated as shown in the following mathematical formula 1 (S420).

[0124] [Mathematical Formula 2]

[0125]

[0126] In other words, by processing the attention scores of all time points with softmax, the relative weight that each time point occupies in the entire sequence is expressed in the form of a probability distribution.

[0127] The attention weight a calculated in this way t is used to generate the final time series representation and is combined with the hidden vectors of the Bi-LSTM to produce a weighted sum (context vector) as shown in Equation 3 below (S430).

[0128] [Mathematical Formula 3]

[0129]

[0130] In the above formula, c is a context vector representing the entire sequence, and this vector c is passed to the model's dense layer to ultimately predict the K-CESD-R score.

[0132] FIG. 5 is a diagram illustrating the components and statistical characteristics of behavior-based digital biomarkers used as input features for model learning in one embodiment of the present invention.

[0133] Various behavioral data collected from users' smartphones and digital devices can be quantified into multidimensional features in terms of physical activity levels, location movement characteristics, spatial distribution, and temporal patterns.

[0134] Specifically, Figure 5 includes quantitative features such as total steps, distance traveled, number of visited places, location dispersion, and GPS-based entropy, with each item presented as the mean and standard deviation (SD) in user units. These items can contribute to detecting a decrease in daily activities, reduced spatial activity, and disruption of repetitive behavior patterns associated with mental illness, particularly depressive symptoms.

[0135] In addition, the right column of Figure 5 shows the within-pattern incoherence index calculated according to various movement distance thresholds. This index is a numerical value that quantitatively expresses how consistently a user's movement pattern is maintained at a specific distance level, and a higher value indicates that the irregularity of behavior and the degree of rhythm disruption have intensified.

[0136] The deep learning model of the present invention receives these features in a time-series format, emphasizes important time points and features for the prediction result through Bi-LSTM and an attention mechanism, and ultimately predicts the K-CESD-R score. As such, the features included in Fig. 5 constitute the input characteristics of the model while simultaneously providing key foundational information for generating explainable prediction results and digital mental health profiles.

[0138] Figure 6 is a table numerically comparing performance indicators between the present embodiment and the comparative embodiment of the deep learning device for predicting mental illness according to the present invention.

[0139] Figure 6(a) shows the results of a comparison between different deep learning model structures, and Figure 6(b) shows how performance varies depending on the combination of input features.

[0140] First, according to FIG. 6(a), the attention Bi-LSTM model used in the present invention has an RMSE of 12.90 and a coefficient of determination (R²). 2 With a ROC-AUC of 0.15 and 0.719, it showed the best performance among all comparison models. This represents an improvement not only compared to simple LSTM or CNN structures but also to Bi-LSTM without attention, confirming that the attention mechanism substantially contributes to the accuracy of mental illness prediction.

[0141] In addition, the results in Fig. 6(a) show that compared to existing unidirectional recurrent neural networks (Vanilla RNN) or feedforward structures, the Bi-LSTM-based structure effectively reflects the contextual information of time-series behavioral data, thereby improving the model's predictive power (R). 2 It demonstrates that it simultaneously improves ) and discrimination ability (ROC-AUC). In particular, the structure combining an attention mechanism with Bi-LSTM suggests that learning a more sophisticated decision mechanism is possible than simple time series processing.

[0142] Figure 6(b) shows how the prediction performance varies depending on the input feature configuration and exemplifies the performance of the attention Bi-LSTM model according to the present invention.

[0143] When combining all variables (Mobility + Demographics + Baseline Score), RMSE 8.64, R 2 It demonstrated the highest performance with an ROC-AUC of 0.62 and 0.90. This implies that a combination of various features, rather than a single feature group, enables reliable judgment in the prediction of mental illness, and supports the fact that the model of the present invention is designed to integrate and analyze multidimensional behavioral indicators and user background information.

[0145] FIG. 7 is a flowchart illustrating an example of a method for generating a digital mental health profile according to an embodiment of the present invention.

[0146] Referring to FIG. 7, the mental illness prediction deep learning device (100) can visually construct a user-customized digital mental health profile by utilizing the predicted result and attention information that contributed to the prediction judgment.

[0147] A deep learning device for predicting mental illness (100) receives as input a result of predicting mental illness—e.g., K-CESD-R score—performed based on a Bi-LSTM and an importance concentration mechanism (S710).

[0148] Predicted results can be presented as continuous numerical values ​​(regression results) or the presence or stage of depression (classification results), and are utilized as key supporting information for individual digital profiles.

[0149] The mental illness prediction deep learning device (100) collects attention weights calculated at the time of prediction execution (S720).

[0150] Attention weights are values ​​that quantitatively indicate how much each time point in time-series behavior data influenced the prediction result, and are used as a criterion for determining the order of importance in subsequent visualizations.

[0151] A deep learning device (100) for predicting mental illness combines attention weights and behavior feature data corresponding to the corresponding time interval to form a time × feature-based 2D matrix (S730).

[0152] Behavioral features may include, for example, travel distance, location variety, GPS entropy, dwell time, and number of visited places, and are sorted based on concentration values ​​to enable visually weighted interpretation.

[0153] Based on the combined time-feature matrix, the mental illness prediction deep learning device (100) generates a visual digital mental health profile in the form of a heatmap (S740).

[0154] For example, the horizontal axis represents the flow of time (e.g., Day 114 or month and day), the vertical axis represents behavioral features, and the cell colors reflect the prediction contribution (attention weight) of the corresponding features at each time point. This heatmap can provide explainability by visualizing the structure of prediction decisions.

[0155] A deep learning device (100) for predicting mental illness can separate and analyze the pattern of change in behavior in units of time intervals such as weekdays / weekends, day / night, and specific days of the week, and add interpretation information of time-based characteristics that affect the user's mental health state to the profile (S750).

[0156] For example, if movement patterns drop sharply and concentration values ​​become concentrated during the weekend, insights such as "weekend isolation tendency" can be automatically generated.

[0157] A digital mental health profile constructed in this manner supports users or professionals in clearly understanding the timing and basis of predictive judgments, and can provide effectiveness as an explainable artificial intelligence (XAI)-based mental health management system.

[0159] FIG. 8 is a graph visualizing the difference in prediction contribution over time intervals for each user group (by level of depression symptoms) based on attention weights calculated by the mental illness prediction deep learning device (100) of the present invention.

[0160] The graph demonstrates that the attention-based prediction model can not only output simple scores but also quantitatively interpret which points in the prediction decision-making process are important for each group.

[0161] The graph in Figure (a) shows the average attention weight for each day within the 14-day time series, and shows a tendency for the attention weight for specific days (e.g., Day 1, Day 14, etc.) to be higher in groups with more severe depressive symptoms. This suggests that it is significant for identifying sensitive points for diagnosing mental disorders, given that changes in behavioral rhythms are mainly concentrated at specific points in time.

[0162] The graphs in Figures (b) and (c) visualize the differences in the timing of prediction contribution by group, classified by day of the week (Monday–Sunday) and weekdays / weekends. Since the group with severe depressive symptoms shows higher concentration values ​​on weekends, while the non-depressive group shows higher prediction contribution on weekdays, it can be seen that the model of the present invention can significantly capture differences in time-based behavioral characteristics according to emotional state.

[0164] FIG. 9 is a graph that visually shows the difference in weekday and weekend behavioral patterns between user groups by level of depressive symptoms, as part of a digital mental health profile generated by the mental illness prediction deep learning device (100) of the present invention for interpreting prediction results.

[0165] Figure 9 illustrates a comparative example of how behavioral characteristics change over time according to the degree of depression, focusing on a total of three behavior-based digital biomarkers (distance traveled, time spent at home, and location entropy).

[0166] The first graph shows Total Moved Distance, indicating a tendency for the distance traveled during the weekend to increase significantly as depressive symptoms become more severe.

[0167] In particular, the 'Severely depressed' group showed a sharp increase in weekend travel distance, exhibiting a significant difference from the normal group (based on p-value).

[0168] This may mean a pattern where movement is restricted during the week, but unstructured or purposeless activities increase on weekends.

[0169] In the second graph, Home Stayed Time, the 'Not depressed' group shows a decrease in time spent at home on weekends, whereas the depressed group actually shows a trend of increasing time spent at home during the weekend.

[0170] This reflects the tendency of users with depressive symptoms to fall into social isolation or reduced activity on weekends, and is interpreted as a key behavioral signal supporting the interpretability of attention-based predictive models.

[0171] The third graph represents Location Normalized Entropy, a value indicating how diverse the locations visited by a user were and how unpredictably those visits were distributed.

[0172] The 'Not depressed' group shows an increase in entropy (increased activity diversity) on weekends, whereas the depressed group shows a tendency for location diversity to decrease sharply on weekends, suggesting that behavioral patterns become fixed and monotonous.

[0173] As such, these analysis results demonstrate that digital mental health profiles can go beyond simple prediction values ​​to enable the implementation of time-interval-based behavioral interpretation and the visualization of expert decision-making mechanisms.

[0174] In other words, the present invention can quantitatively analyze and visualize temporal changes in behavioral patterns and abnormal deviations as diagnostic clues for depression, and this can be usefully utilized to assist clinical judgment or induce the user to recognize their own condition.

[0176] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0177] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0178] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0179] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0180] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A deep learning-based method for predicting mental illness comprises: a step of generating a time-series behavioral data sequence by arranging behavioral data based on location movement information collected through a user device in chronological order; a step of converting behavioral changes at each time point into time-series hidden vectors reflecting contextual information using a recurrent neural network, with the time-series behavioral data sequence as input; a step of calculating time-series concentration values ​​by applying an importance concentration mechanism to the hidden vectors to calculate weights that can automatically identify and emphasize time intervals important for predicting mental illness; and a step of predicting the user's risk of mental illness based on a time-series representation reflecting the concentration values. The method includes the step of generating a user-customized digital mental health profile configured to enable visual interpretation of behavioral patterns that contributed to the prediction of mental illness risk by time interval by utilizing the user's mental illness risk and the concentration value at each time point together; wherein the step of generating the user-customized digital mental health profile comprises: configuring a two-dimensional matrix having a time axis and a behavioral feature axis using the concentration value at each time point and a plurality of behavioral features included in the time-series behavioral data sequence; placing a prediction contribution calculated based on the concentration value at each time point and the behavioral feature value at the corresponding time point as a cell value corresponding to each time point on the time axis and each behavioral feature on the behavioral feature axis in the two-dimensional matrix; and generating visualization information of the digital mental health profile by visualizing the distribution of the prediction contribution according to the time axis and the behavioral feature axis in the form of a heatmap based on the two-dimensional matrix. A deep learning method for predicting mental illness based on behavioral data using simulation of an expert decision-making mechanism, comprising the step of: generating interpretation information in time interval units separated into weekday intervals and weekend intervals for the distribution of the prediction contribution included in the digital mental health profile; and including the interpretation information in the digital mental health profile. Claim 2 delete Claim 3 A deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation, wherein the behavioral data comprises at least one of GPS-based travel distance, number of visited places, location diversity, start time of movement, and end time. Claim 4 In claim 1, the importance concentration mechanism calculates an association score between a hidden vector at each time point and a global vector reflecting contextual information of the entire sequence, and normalizes the concentration value calculated based on the association score using a softmax function to assign relative importance to each time point, thereby producing a deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation. Claim 5 A deep learning method for predicting mental illness based on behavioral data using simulation of an expert decision-making mechanism, wherein the step of predicting the risk of mental illness of the user according to claim 1 includes the step of predicting whether the user has depressive symptoms or a continuous score level based on the CES-D (Center for Epidemiologic Studies Depression Scale) score. Claim 6 A deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation, wherein the time-series behavioral data sequence includes behavioral data of the last 14 days or more on a user-by-user basis. Claim 7 A deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation, wherein the digital mental health profile includes data expressed in the form of a heatmap that visually represents behavioral data characteristics corresponding to concentration values ​​for each time interval. Claim 8 A deep learning method for predicting mental illness based on behavioral data using an expert decision-making mechanism simulation, wherein the digital mental health profile includes data that chronologically explains the relationship between changes in individual activity characteristics and predicted depressive symptoms based on differences in movement patterns between weekdays and weekends. Claim 9 As a deep learning device for predicting mental illness, at least one processor; The invention includes a memory for storing instructions, wherein when the instructions are executed individually or collectively by at least one processor, the processor: generates a time-series behavioral data sequence by arranging behavioral data based on location movement information collected through a user device in chronological order; takes the time-series behavioral data sequence as input and converts behavioral changes at each time point into time-series hidden vectors reflecting contextual information through a recurrent neural network; calculates time-series concentration values ​​by applying an importance concentration mechanism to the hidden vectors to calculate weights that can automatically identify and highlight time intervals important for mental illness prediction; predicts the user's mental illness risk based on a time-series representation reflecting the concentration values; and generates a user-customized digital mental health profile configured to allow visual interpretation of behavioral patterns that contributed to the prediction of mental illness risk by time interval by utilizing the user's mental illness risk and the time-series concentration values ​​together; wherein, to generate the user-customized digital mental health profile, the processor utilizes the time-series concentration values ​​and a plurality of behavioral features included in the time-series behavioral data sequence, a time axis and Constructing a two-dimensional matrix having a behavioral feature axis; placing a predicted contribution calculated based on the concentration value by time point and the behavioral feature value of the corresponding time point as a cell value corresponding to each time point on the time axis and each behavioral feature on the behavioral feature axis in the two-dimensional matrix; then, based on the two-dimensional matrix, visualizing the distribution of the predicted contribution according to the time axis and the behavioral feature axis in the form of a heatmap to generate visualization information of the digital mental health profile; generating interpretation information for the distribution of the predicted contribution included in the digital mental health profile in units of time intervals separated into weekday and weekend intervals; and including the interpretation information in the digital mental health profile.Deep learning device for predicting mental illness based on behavioral data using simulation of expert decision-making mechanisms. Claim 10 A computer-readable recording medium having a computer program stored therein for performing a deep learning method for predicting mental illness based on behavioral data using a simulation of the expert decision-making mechanism of any one of claims 1, 3 to 8, combined with hardware.

Images