Ai agent-based digital therapeutic system and method for controlling mood disorder

Abstract

The present invention relates to an artificial intelligence agent-based digital therapeutic system and method for controlling mood disorders. The system according to the present invention includes a data collection unit that collects data from a patient; an episode transition control agent that calculates control variables, including the direction and speed of episode transitions, and generates therapeutic intervention proposal information; a supervisor agent that determines a final therapeutic intervention policy; and an intervention execution unit that provides customized therapeutic intervention to the patient. According to an embodiment of the present invention, the point at which a patient's emotional state transitions from a stable state to depression or mania can be identified early and precisely, and the timeliness of therapeutic intervention can be improved by capturing the dynamic interval between episodes.

Description

Technology Field

[0001] The present invention relates to a method and system for digital treatment of mood disorders using artificial intelligence. Specifically, it relates to a system and method for digitally treating mood disorders by controlling the episode switching of depression or bipolar disorder based on an artificial intelligence agent. Background Technology

[0003] Bipolar disorder is a mood disorder characterized by the repeated alternation of depressive and manic episodes. The patient's severe mood swings lead to a decline in daily functioning and difficulties in social adaptation. Depression is a mood disorder in which persistent feelings of sadness and a loss of interest and motivation continue for a certain period, and can impair daily functioning along with changes in sleep, appetite, and concentration.

[0004] Traditional treatment for depression or bipolar disorder has primarily relied on medication and counseling based on clinical evaluation. This is based on a specialist's medical history and the patient's subjective reports.

[0005] Recently, attempts have been made to collect patients' biometric data using mobile devices or wearable devices. Technologies are being proposed to estimate the type or stage of disease progression through the collected data.

[0006] However, existing technologies remain static, classifying a patient's condition at a specific point in time or calculating risk scores. They have limitations in timely identifying dynamic transition periods that occur between episodes, such as in bipolar disorder.

[0007] Furthermore, the method of generating warnings based on fixed thresholds fails to reflect the varying conversion characteristics of individual patients. This leads to problems such as missing the appropriate timing for therapeutic intervention or the difficulty in expecting consistent treatment effects. Prior art literature

[0009] Korean Published Patent Application No. 10-2021-0058449 (Publication Date: May 24, 2021) The problem to be solved

[0010] The present invention aims to solve the aforementioned problems by providing a digital treatment method capable of precisely identifying the dynamic transition interval in which a patient's emotional state transitions from a stable state to depression or mania.

[0011] Furthermore, the present invention aims to provide a digital treatment method that can capture the momentum of a transition in a timely manner by analyzing the rate and acceleration of change in patient data, going beyond the limitations of statically classifying a patient's emotional state at a specific point in time.

[0012] In addition, the present invention aims to provide a digital therapy method that evaluates the reliability of data collected from multiple information channels and weights and fuses them to probabilistically estimate the direction of change in emotional state.

[0013] In addition, the present invention aims to provide a digital treatment method capable of predicting the expected duration of current symptoms and the remaining time until termination by performing a statistical survival analysis based on a history of past episodes.

[0014] In addition, the present invention aims to provide a digital therapeutic method capable of reviewing and mediating potential conflicts among multiple derived therapeutic intervention proposals from an integrated perspective.

[0015] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0017] According to an embodiment of the present invention, an artificial intelligence agent-based digital therapeutic system for controlling mood disorders comprises a data collection unit that collects data from a patient, an episode transition control agent that calculates control variables including episode transition direction and transition speed and generates therapeutic intervention proposal information, a supervisor agent that determines a final therapeutic intervention policy, and an intervention execution unit that provides customized therapeutic intervention to the patient.

[0018] The above data collection unit can collect data from a patient through a plurality of data channels including at least one of text, voice, behavioral patterns, and biosignals.

[0019] The illustration switching control agent may include at least one of a switching direction control agent, a switching speed control agent, an intensity and safety control agent, and a duration control agent.

[0020] The above-described transition direction control agent may be configured to calculate the reliability of each of the plurality of data channels in real time, assign optimal weights to individual channel information, fuse channel information reflecting the weights to generate a state change feature representation, and calculate a probability distribution for each illustration transition direction from the feature representation to estimate the direction of illustrations with a high probability of transition in a time series.

[0021] In addition, before assigning optimal weights to the individual channel information, the switching direction control agent may extract features by calculating at least one of emotional expression, change in thought flow, speech characteristics, change in daily rhythm, and bio-variability indicators for each of the plurality of data channels, and normalize the extracted features into relative deviations by comparing them with the individual patient's stable baseline.

[0022] The above-described switching speed control agent calculates the switching speed, which is the rate of change per unit time, from the collected time-series data, calculates the switching acceleration, which is the amount of change in the switching speed, and compares and analyzes the magnitudes of the switching speed and the switching acceleration with preset reference values ​​to identify the speed grade of the illustration switching and the actual switching start time.

[0023] The above intensity and safety control agent can compare real-time symptom data with symptom-specific risk thresholds, evaluate the risk that a specific therapeutic intervention has on symptoms according to intervention-symptom interaction rules, and generate a set of safety constraints that define the acceptable range of the therapeutic intervention based on the evaluation results.

[0024] The duration control agent above can calculate the expected duration of the current episode by executing a statistical survival analysis model using past episode history data as input, and can estimate the remaining time until the end of the episode based on the elapsed time of the current episode.

[0025] The supervisor agent can detect combinations that violate patient safety constraints or have mutually contradictory treatment directions as conflict pairs. The supervisor agent resolves the conflict pairs by applying predefined priority rules and constraints; however, if conflict resolution is not possible, the supervisor agent can execute interaction logic to request the relevant agent to reanalyze or modify the control variables.

[0026] The above-mentioned episode transition control agent and the above-mentioned supervisor agent may be configured to update control rules based on changes in patient response and episode transition characteristics after treatment intervention.

[0027] The above-mentioned personalized treatment intervention may include at least one of providing treatment content, controlling medication reminders, connecting emergency medical personnel, and transmitting signals to control the user's surrounding environment. The above-mentioned mood disorder may include depression and bipolar disorder.

[0028] According to another embodiment of the present invention, an artificial intelligence agent-based digital treatment method for controlling mood disorders comprises the steps of: a data collection unit collecting data from a patient through a plurality of data channels including at least one of behavioral patterns and biosignals; an episode transition control agent calculating control variables including an episode transition direction and a transition speed for an episode transition interval between a depressive episode, a manic episode, and a stable state, and generating treatment intervention proposal information; a supervisor agent detecting whether there is a conflict between the proposal information and adjusting it to determine a final treatment intervention policy; and an intervention execution unit providing a customized treatment intervention to the patient based on the final treatment intervention policy.

[0029] According to another embodiment of the present invention, a computer-readable storage medium is provided on which a program for executing the above-described artificial intelligence agent-based digital treatment method for controlling mood disorders is recorded. Effects of the invention

[0031] According to an embodiment of the present invention, the point at which a patient's emotional state transitions from a stable state to depression or mania can be identified early and precisely. In particular, by capturing the dynamic interval between episodes, the timeliness of therapeutic intervention can be dramatically improved. This can contribute to preventing the rapid deterioration or recurrence of mood disorders in advance.

[0032] Furthermore, according to an embodiment of the present invention, the momentum of mood changes can be quantitatively identified by analyzing not only the simple rate of change of data but also the acceleration. By detecting signs of transition more rapidly than conventional static classification methods, the response speed of the system can be increased. Through this, it is possible to take preemptive preventive measures before the patient's condition reaches a critical phase.

[0033] In addition, according to an embodiment of the present invention, the objectivity and reliability of data analysis can be enhanced by evaluating the reliability of each of the multiple information channels and assigning weights. The accuracy of diagnosis can be improved by intelligently excluding noise from a specific data channel and probabilistically deriving the direction of change in mood state. Through multidimensional data fusion, the risk of misdiagnosis occurring when relying on limited information can be minimized.

[0034] Furthermore, according to an embodiment of the present invention, the duration of symptoms based on a history of past episodes can be scientifically estimated by applying a statistical survival analysis model. By predicting the remaining time until termination, it is possible to support medical staff and patients in systematically establishing plans for treatment and return to daily life. This predictive information can be utilized as an indicator to enhance the efficiency of digital therapy and improve patient medication adherence.

[0035] Furthermore, according to an embodiment of the present invention, consistency in treatment policies can be ensured by intelligently mediating logical contradictions or conflicts that may arise between multiple proposals. By prioritizing clinical safety constraints, the risk of safety accidents resulting from inappropriate intervention can be eliminated. Consequently, a safe and integrated digital treatment experience can be provided to patients even in environments where various variables interact in a complex manner.

[0036] The effects obtainable from the present invention are not limited to those mentioned above, and should be understood to include all effects that can be inferred from the configuration of the invention described above. Brief explanation of the drawing

[0038] FIG. 1 is a schematic diagram showing the data processing flow between a user terminal and a server according to an embodiment of the present invention. FIG. 2 is a block diagram illustrating the detailed configuration of an artificial intelligence agent-based digital therapeutic system for illustration switching control according to an embodiment of the present invention. FIG. 3 is a block diagram showing the detailed configuration of an illustration switching control agent according to an embodiment of the present invention. FIG. 4 is a flowchart illustrating the collision adjustment logic of a supervisor agent according to one embodiment of the present invention. FIG. 5 is a diagram illustrating the detailed configuration of a database according to one embodiment of the present invention. FIG. 6 illustrates a flowchart of an artificial intelligence agent-based digital treatment method for controlling manic-depressive episode transitions according to an embodiment of the present invention. Specific details for implementing the invention

[0039] The above-mentioned objects, features, and advantages of the present invention are intended to be clarified through the following detailed description in conjunction with the drawings of this specification. However, various modifications may be made and various embodiments may be provided based on the following detailed description. In the following, specific embodiments are illustrated in the drawings and described in detail.

[0040] Throughout the present invention, identical reference numerals generally represent identical components. Additionally, functional components with identical appearance in the drawings of each embodiment are described using the same reference numeral, and redundant descriptions thereof are omitted.

[0041] In the following embodiments, terms such as "comprising" or "having" mean that the features or components described in the specification are present, and do not preclude the possibility that one or more other features or components may be added.

[0042] FIG. 1 is a schematic diagram showing the data processing flow between a user terminal and a server according to an embodiment of the present invention. Referring to FIG. 1, the system of the present invention may be configured so that a server (100) and a user terminal (200) communicate through a network.

[0043] The above network may be one or more of wired or wireless communication networks, and may include, for example, the Internet, intranet, local area network (LAN), wide area network (WAN), mobile communication network (3G, LTE, 5G, etc.), Wi-Fi, Bluetooth, near-field communication (NFC), satellite communication network, etc.

[0044] The above server and user terminal can perform data transmission and reception, command transmission, model updates, service requests and responses, etc. on a cloud infrastructure or edge computing environment.

[0045] This communication is based on the TCP / IP protocol, and various application layer protocols such as HTTP, HTTPS, MQTT, and WebSocket may be used as needed. The present invention is not limited to the form of the network or the communication protocol, and can utilize all electronic communication means including future next-generation communication technologies (e.g., 6G, IoT dedicated networks, etc.).

[0046] The server (100) may include one or more computing resources (e.g., cloud server, on-premises server) and may be responsible for centralized processing such as data storage and execution of high-performance AI models. The user terminal (200) may be any electronic device capable of network access and information processing, such as a smartphone, tablet, or wearable device.

[0047] The artificial intelligence agent-based digital treatment system (300) for controlling mood disorders described with reference to FIG. 2 is implemented on the server (100) in one embodiment and can perform data processing and core functions.

[0048] However, the present invention is not limited to the implementation of the server (100). In other embodiments, all or part of the functional modules of the system (300) may be implemented as dedicated applications (apps) or other possible forms of software modules installed on the user terminal (200). This may have the advantage of reducing communication delays with the server and providing an on-device processing environment for processing user personal data within the terminal.

[0049] In the present invention, “mood disorder” refers to a disorder characterized by an abnormality in an individual’s mood (affect) state, and includes, for example, depressive disorder (depression), bipolar disorder (manic-depressive disorder), etc. Specifically, it includes, but is not limited to, Persistent Depressive Disorder (PDD), Seasonal Affective Disorder (SAD), Premenstrual Dysphoric Disorder (PMDD), Disruptive Mood Dysregulation Disorder (DMDD), cyclothymia, and Substance / Medication-Induced Bipolar Disorder.

[0050] Referring to FIGS. 2 and 3, an artificial intelligence agent-based digital treatment system (300) for controlling mood disorders according to one embodiment of the present invention will be described in detail.

[0051] FIG. 2 is a block diagram illustrating the detailed configuration of an artificial intelligence agent-based digital treatment system (300) for controlling mood disorders according to an embodiment of the present invention. The digital treatment system (300) according to the present invention monitors the patient's condition in real time. The system precisely identifies the episode transition interval and performs optimal treatment intervention.

[0052] The digital treatment system (300) can operate in conjunction with a server (100) or a user terminal (200). The server (100) performs the function of processing large amounts of data and computing artificial intelligence models. The user terminal (200) is responsible for the interface function of collecting data from patients and outputting treatment content. Each component of the system (300) can be implemented by being physically separated or integrated within a single device.

[0053] The digital therapy system (300) includes a data collection unit (310), an illustration transition control agent (320), a supervisor agent (330), and an intervention execution unit (340). The system may further include a database (400) and a relearning unit (350). The control unit (360) performs the role of controlling the overall operation and data flow of the system.

[0054] The data collection unit (310) collects data from the patient through a plurality of data channels including behavioral patterns and biosignals. The plurality of data channels may further include text and voice. Behavioral patterns include sleep time, activity level, location changes, and app usage history. Biosignals may include heart rate variability (HRV), skin conductivity, and body temperature.

[0055] The data collection unit (310) preprocesses the collected data to generate feature vectors. In the text channel, word usage frequency or sentiment analysis indicators can be extracted. In the voice channel, the speed, pitch change, and intensity of speech are analyzed. The data collection unit (310) evaluates the quality of each data channel in real time.

[0056] The episode transition control agent (320) analyzes data received from the data collection unit (310). The agent identifies the episode transition intervals between depressive episodes, manic episodes, and stable states. The agent calculates control variables including the direction and speed of the episode transition. Based on the calculated control variables, it generates treatment intervention suggestion information.

[0057] Referring together with FIG. 3, the illustration switching control agent (320) may include a plurality of sub-agents. The sub-agents include a switching direction control agent (321), a switching speed control agent (322), an intensity and safety control agent (323), and a duration control agent (324). Each sub-agent independently performs its own analysis logic.

[0058] The illustration switching control agent (320), supervisor agent (330), and intervention execution unit (340), which are system components of the present invention, can be implemented based on Agentic AI technology. Below, the specific operating principles, learning process, and implementation form of the Agentic AI applied to the present invention will be described in detail.

[0059] Agentic AI refers to an intelligent system that autonomously reasons, formulates plans, and executes to achieve a given goal. This is distinguished from existing artificial intelligence models that simply generate outputs corresponding to inputs. The agentic AI in this invention may adopt a Large Language Model (LLM) as its core reasoning engine. The agentic AI possesses agency, identifying the premonitory intention of episode transitions from the patient's behavioral data and independently determining the optimal treatment path.

[0060] The operating principle of the aforementioned agentic AI can be based on the Chain-of-Thought and React frameworks. The agent processes complex episode transition analysis tasks by breaking them down into smaller sub-tasks. At each stage, the agent observes the current patient state, plans necessary analytical actions, and reflects the results back into the policy. Through this iterative loop, sophisticated back-inference regarding the patient's internal mood state becomes possible.

[0061] The agentic AI applied in the present invention can be implemented in the form of a multi-agent system. The system performs tasks by having multiple agents assigned different roles and expertise cooperate or coordinate. For example, a supervisor agent (330) can mutually review proposals generated by individual sub-agents within an illustration switching control agent (320) to eliminate bias. This cooperative structure suppresses the hallucination phenomenon that a single model may have and enhances the objectivity of reasoning.

[0062] The autonomous planning of the aforementioned agentic AI can utilize Hierarchical Planning. A higher-level agent sets a macroscopic goal of mood stabilization, and lower-level agents generate a detailed intervention sequence to achieve this. If the expected change in the patient's condition does not occur during execution, the agent undergoes a replanning process to modify the plan.

[0063] The training process for implementing the aforementioned agentic AI can be composed of multiple stages. First, a foundational model is prepared by pre-training it using a vast amount of plain text data. Subsequently, supervised fine-tuning is performed using specialized datasets for psychiatry and mood disorders. Through this process, the agent acquires therapeutic conversational techniques for mood disorders and clinical expertise.

[0064] Furthermore, human feedback-based reinforcement learning (RLHF) can align the agent's responses and behaviors with therapeutic guidelines. It is trained to select safe and effective intervention strategies by utilizing specialist feedback as a reward signal. In this invention, user feedback-based reinforcement learning (RLUF), which uses the patient's actual symptom improvement performance as a reward indicator, may be additionally applied. Through this, the system can continuously refine intervention methods optimized for the user's preferences.

[0065] The memory system of agentic AI can be managed by dividing it into short-term and long-term memory. Short-term memory stores the context of the current conversation and information regarding the user's immediate response. Long-term memory can be implemented using Vector Databases and Search Augmentation Generative (RAG) technology. The agent searches through a vast history of past episodes and successful intervention cases to derive the strategy most suitable for the current situation.

[0066] The aforementioned agentic AI may include the ability to use tools that interact with various external tools. The agent can invoke external functions, such as activity measurement APIs or sleep analysis modules, if necessary. The extracted quantitative data is then utilized as input for the agent's inference process. This capability helps the agent make decisions based on real-time lifestyle data, rather than becoming bogged down in text information.

[0067] The types of agents used in this invention can be classified according to their level of autonomy. A combination of fixed agents that operate only within specific rules and fully autonomous agents that freely change their paths to achieve goals may be used. Additionally, a persona agent responsible for user interaction and an analytical agent that verifies hypotheses in the backend may operate logically separately.

[0068] A guardrail mechanism may be applied to ensure the safety of the agentic AI. It checks in real time whether the intervention sequence generated by the agent violates medical ethics or safety constraints. If a violation is found, the supervisor agent (330) can exclude the relevant treatment intervention component or reconfigure it into a safe alternative.

[0069] The aforementioned components may operate as independent processes or exist as modules within a single integrated agentic system. Communication between agents can be performed using structured data formats such as JSON or XML. Additionally, intermediate results generated during the inference process may be stored in a separate log database for traceability. This design contributes to the retrospective analysis and improvement of the system's operation.

[0070] The description regarding the configuration, learning methods, operating principles, and tool utilization of the agentic AI described above is merely illustrative and the present disclosure is not limited thereto. Those skilled in the art may implement each configuration of the present invention by selecting various forms of agent architectures or learning algorithms according to their level of technology. Furthermore, technical names such as large-scale language models, reinforcement learning algorithms, and vector databases mentioned in the present disclosure are merely illustrative, and the scope of the present invention is not limited to specific model names or algorithms. Those skilled in the art may use other AI models or data structures that perform equivalent functions to achieve the objectives of the present invention.

[0071] Referring to FIG. 3, a plurality of sub-agents of the illustration transition control agent (320) are described in detail. The transition direction control agent (321) organically integrates a plurality of data channels to estimate the direction of the patient's emotional transition. The agent receives unstructured and structured data flowing in in real time in conjunction with the data collection unit (310). The collected data serves as the basis for determining which extreme illustration the patient's current mood state is moving toward.

[0072] The data channel can be composed of multiple channels including text, voice, biosignals, and behavioral patterns. The text channel analyzes the frequency of word usage within messages written by the patient or posts on social networking services. In particular, it quantifies changes in emotional indicators associated with depression or joy and utilizes this as data. The voice channel extracts speech speed, pitch variability, and tone intensity during a call to determine whether the patient is experiencing emotional ups or downs.

[0073] Features such as emotional expression, changes in thought flow, speech characteristics, changes in daily rhythms, or indicators of bio-variability are extracted from multiple collected data channels. The extracted features are compared to a baseline from the individual user's stable period or past reference state. Based on the comparison results, the agent normalizes the extracted feature data into relative deviations. This process is designed to mitigate classification distortion caused by differences in absolute values ​​between users. Through this, the system can perform precise analysis by focusing on the direction of change within the same user.

[0074] The switching direction control agent (321) calculates the reliability of each input data channel in real time. It evaluates the validity of individual information based on the noise level of data acquired from the sensor or the stability of the collection environment. It automatically reduces the influence of a channel whose reliability has rapidly deteriorated at a specific point in time to prevent errors in the overall analysis. Based on the evaluated reliability, it assigns weights optimized for individual channel information. It fuses the channel information with the weights to generate a feature representation indicating a state change.

[0075] From the generated feature representations, probability distributions for each transition direction, such as depressive or manic episodes, are calculated. An internal intelligent model receives feature vectors as input and computes the probability values ​​for transitioning to each episode state. The calculated distribution numerically indicates the direction in which the patient is shifting from a neutral, stable state. Through the analysis of the probability distribution, the dominant transition direction with the highest probability of transition is determined. The system monitors in real-time whether the probability toward a specific direction exceeds a threshold.

[0076] Based on the calculated probability distribution by transition direction, therapeutic intervention objectives suitable for the corresponding transition phase are mapped. If the probability of transitioning toward mania is dominant, stimulus inhibition and rhythm stabilization are set as intervention objectives. Conversely, if the probability of transitioning toward depression is dominant, behavioral activation and emotional support may be set as intervention objectives.

[0077] The transition direction control agent (321) generates and outputs proposal information including analysis results. The proposal information may be configured to include probability distributions by transition direction and confirmed dominant transition direction. Additionally, it conveys information that includes specific intervention objectives mapped by phase. The generated proposal information is transmitted to the supervisor agent (330) and used as supporting data for determining the final treatment policy.

[0078] The transition speed control agent (322) performs the role of precisely analyzing the physical aspects of emotional state changes. Referring to FIG. 3, the agent may be included as a sub-component of the illustration transition control agent (320). It tracks the trend of change in collected time-series data to quantify the momentum of state transitions. This serves as a key basis for determining the urgency of state changes, going beyond a simple diagnosis of the current state. The agent analyzes the temporal flow of data by converting it into physical variables.

[0079] The switching speed control agent (322) performs a process of comparing feature data with the user's individual stable baseline. It calculates the distance between the feature value extracted from the data collection unit (310) and the average data of past stable periods. Based on the calculated distance, the fluctuation of the current data is normalized into a relative deviation. The normalized data is used as an objective indicator that takes into account the individual's unique emotional fluctuation range.

[0080] The switching speed control agent (322) calculates the switching speed, which is the rate of change per unit time, from normalized time series data in real time. The switching speed refers to the speed at which a patient's mood moves from a stable state to a specific episode. Mathematically, it can be defined as the first derivative of the normalized deviation data. A higher switching speed indicates that the patient's emotional state is rapidly becoming unstable. The agent continuously monitors the fluctuations in speed by adjusting the window size of the time series data.

[0081] The transition speed control agent (322) may be configured to detect a point in time when the average level, variance, or periodicity of the time series data changes statistically significantly as a change point, and to identify said change point as the start point of the illustration transition. This allows distinguishing whether the transition is a simple variation or a structural change.

[0082] The transition speed control agent (322) calculates the transition acceleration, which is the amount of change in the calculated transition speed. Transition acceleration is an indicator that indicates the trend of the speed of state change increasing or decreasing. It can be understood as a physical quantity corresponding to the second derivative of time series data. If the acceleration increases in the positive direction, it means that entry into a specific episode is accelerating. Through acceleration analysis, it is possible to preemptively detect signs of minute changes before exceeding a threshold.

[0083] The agent compares and analyzes the calculated transition speed and transition acceleration magnitudes against preset thresholds. Based on the comparison results, the speed grade of episode transitions can be classified into multiple levels. For example, the grades can be subdivided into rapid cyclic transitions or gradual transitions. Through this classification, the intensity, immediacy, and scope of therapeutic interventions can be determined according to the patient's condition. The thresholds can be personalized and set based on the patient's past episode transition history.

[0084] The actual initiation point of the transition is identified by verifying whether the thresholds for transition speed and acceleration have been reached. The starting point is defined not simply as the moment when the emotion level exceeds a specific score, but as the moment when the momentum of change is established. This logic is designed to reduce false alarms and determine the initial phase of the true episode transition. The identified initiation point information serves as the criterion for determining the activation timing of the treatment policy. The agent generates and outputs proposal information containing the identified time point information, transition speed, acceleration, change point information, and corresponding intervention parameters.

[0085] The intensity and safety control agent (323) performs the role of ensuring the clinical safety of the digital therapeutic intervention provided to the patient. Referring to FIG. 3, the agent (323) is a sub-component of the episode transition control agent (320) and analyzes symptom data flowing in in real time. The collected data identifies potential risk factors that may affect the patient's health condition and determines the appropriate intensity of the intervention. This is a key component for minimizing side effects that may occur during the patient's extreme mood swings.

[0086] The intensity and safety control agent (323) compares and analyzes real-time symptom data received from the data collection unit (310) with pre-set symptom-specific risk thresholds. The risk thresholds can be set individually by combining general clinical indicators and the patient's past episode history.

[0087] The intensity and safety control agent (323) evaluates the risk that a specific therapeutic intervention will have on symptoms based on intervention-symptom interaction rules. Intervention-symptom interaction rules are a set of logic that defines the impact that specific content or activity recommendations will have on the current episode phase. For example, if there is a prodromal sign of a manic episode, it determines whether there is a risk that an intervention that overstimulates activity will worsen symptoms. Based on the evaluation results, a risk score for each type of therapeutic intervention can be calculated or a grade can be classified.

[0088] The agent (323) generates a set of safety constraints that defines the permissible range of treatment interventions based on the results of the risk assessment. The set of safety constraints consists of rules that specify limitations or essential recommendations on the treatment actions that the system can provide to the patient. This may include physical upper limits on the types of interventions that should not be performed in certain situations or the intensity of the interventions. The generated set of constraints is used as essential filtering data for the final treatment policy decision.

[0089] The intensity and safety control agent (323) outputs suggestion information including a risk level by symptom, a set of safety constraints, and whether intervention is allowed. The output information is integrated in the form of control variables and reflected in the system's final decision-making process.

[0090] The duration control agent (324) performs the function of predicting how long the currently occurring episode or transition phase will be maintained in time. Referring to FIG. 3, the agent may be an independent sub-agent constituting the episode transition control agent (320). The agent (324) analyzes the time-series characteristics of emotional fluctuations to quantify the temporal range of the corresponding symptoms. This serves as a key indicator for determining the duration of a treatment policy or controlling the frequency of medication reminders.

[0091] The agent (324) retrieves the patient's past episode history data as input from the user DB (401) of the database (400). The past history data includes information on the onset and end times of depressive or manic episodes and the total duration of each individual episode. The cycle and duration patterns of mood disorders, which vary from patient to patient, are statistically analyzed. The accumulated records of individual users can be utilized as important statistical parameters to improve the accuracy of the time prediction model.

[0092] The duration control agent (324) executes a statistical survival analysis model based on collected historical data. Survival analysis may be a technique that predicts the probability of occurrence by analyzing the time required for a specific event to occur. The agent calculates the expected duration by inputting the nature of the currently identified episode and the patient's clinical characteristics into the survival function. The expected duration refers to the total time range during which the state of the episode is statistically predicted to be maintained without ending.

[0093] The agent (324) estimates the remaining time until the end of the episode in real time based on the actual elapsed time of the current episode. The remaining time can be derived by subtracting the time elapsed so far from the calculated expected duration. The remaining time information is continuously adjusted by considering environmental variables that change as the symptoms progress and the patient's responsiveness. This can provide precise temporal guidelines to predict the time when symptoms stabilize for the patient or medical staff.

[0094] The calculated expected duration and remaining time information is transmitted to the supervisor agent (330) in the form of proposal information. The supervisor agent (330) utilizes the said time information when establishing long-term or short-term strategies for treatment intervention. For example, if the remaining time is predicted to be short, a policy of gradually discontinuing treatment intervention may be considered. On the other hand, if the expected duration is calculated to be long, a policy of prioritizing the placement of long-term emotional support content may be determined.

[0095] The supervisor agent (330) integrates individual proposals derived from multiple sub-agents to make final decisions for the system. Referring to FIG. 4, the supervisor agent (330) performs a "safety review" step on the proposal information generated by each sub-agent (321, 322, 323, 324). This step is a process for determining consistency between proposals with different analysis perspectives and excluding clinical risk factors. By verifying the results of each analysis node at the upper level, consistency of the treatment policy delivered to the patient is ensured.

[0096] The supervisor agent (330) detects combinations of mutually contradictory treatment directions among the received proposal information as conflict pairs. The supervisor agent (330) may be configured to resolve conflicts by applying predefined priority rules and constraints to the detected conflict pairs. As shown in FIG. 4, if the conflict is not resolved by priority rules alone, or if multiple conflicts occur in succession (No), a "request for proposal modification" step is performed. During this process, interaction logic is executed to request the relevant sub-agent to reanalyze or modify control variables. This is a key feedback function to converge fragmented information from individual agents into a single optimized treatment strategy.

[0097] On the other hand, if the proposed information satisfies safety constraints and consistency is ensured (e.g.), the final strategy is derived through the "intervention policy confirmation" stage. The confirmed policy is transmitted to the intervention execution unit (340) to control the provision of actual "strategic intervention" to the patient. The supervisor agent (330) performs intelligent intervention tasks by referring to priority rules stored in the control rule DB (402). In particular, safety constraints set by the safety control agent (323) are reflected as the highest priority when the policy is confirmed.

[0098] The following describes a specific example of conflict adjustment performed by the supervisor agent (330) through the logic of FIG. 4. For example, the transition speed control agent (322) may detect a sudden manic transition and propose a high-intensity activity suppression policy. At this time, the duration control agent (324) may predict that the state will end in a very short period based on the patient's history. The supervisor agent (330) identifies this as a conflict pair and adjusts the intensity of the intervention to reduce the patient's daily workload.

[0099] As another example, the transition direction control agent (321) may identify a transition to a depressive episode and suggest sending behavioral activation content. However, the safety control agent (323) may detect that the patient's current vital signs indicate an extreme state of anxiety and create safety constraints. The supervisor agent (330) determines through the review step of FIG. 4 that the suggestion violates the safety conditions. Accordingly, the policy is changed and finalized to meditation content that induces emotional stability instead of behavioral activation.

[0100] As another example, there is a risk that multiple agents may suggest interventions at different times, thereby disrupting the patient's sleep rhythm. The supervisor agent (330) manages the entire intervention schedule and reviews whether there is any temporal overlap between interventions. By setting the maintenance of the patient's stable daily rhythm as the top priority, the timing of interventions is repositioned or integrated. Through this intelligent coordination, the patient is provided with a consistent digital treatment experience that is not contradictory.

[0101] The conflict resolution and policy determination method according to the flowchart of FIG. 4 is merely an example to aid in understanding the invention. The supervisor agent (330) has a structure that allows it to self-evolve the adjustment rules themselves through the relearning unit (350). It learns patient response data after actual intervention to continuously refine the priority matrix or constraints. Through this closed-loop learning, the system provides more sophisticated and safe intervention services to specific users as time passes.

[0102] Referring to FIG. 2, the intervention execution unit (340) provides a customized treatment intervention to the patient based on the final treatment intervention policy. The customized treatment intervention includes providing treatment content or controlling medication reminders. It may include connecting to emergency medical personnel or transmitting signals to control the user's surrounding environment. The intervention execution unit (340) adjusts the intensity of the intervention according to the patient's condition.

[0103] Referring to FIG. 5, the relearning unit (350) collects changes in patient response and episode transition characteristics in real time after a customized treatment intervention is provided. The configuration may be implemented by including a feedback collector (351) and a relearning orchestrator (352). The feedback collector (351) records the patient's interaction indicators regarding the intervention content and whether the actual symptoms have improved as digital data. The collected feedback data serves as an indicator to determine how well the patient is adapting to the current treatment policy.

[0104] The retraining orchestrator (352) processes the collected data to optimize the control rules of the agents and the weights of the artificial intelligence model. Through this feedback process, the system forms a closed-loop learning structure that provides precision treatment specialized for specific users over time. By dynamically reflecting the unique mood change cycles of individual patients into the learning model, the accuracy of transition identification can be gradually improved. This plays a key role in preventing the recurrence of mood disorders and increasing treatment compliance in the long term.

[0105] The database (400) serves as an information hub for the system and manages the storage and input / output of various data required by each component. The database (400) may be configured to include a user DB (401), a control rule DB (402), and a model registry (403). Physically, it may be located in a secure storage area within the server (100) or implemented as a cloud-based distributed storage system. The database (400) accumulates and manages real-time time-series data generated during system computation processes.

[0106] The user DB (401) systematically stores the patient's stable baseline and all past episode history data. The duration control agent (324) retrieves time-series information recorded in the user DB (401) to perform statistical survival analysis. The control rule DB (402) manages priority rules and safety constraints that the supervisor agent (330) refers to when adjusting proposal information. The model registry (403) manages versions of various artificial intelligence algorithms used to estimate the direction or speed of episode transitions.

[0107] FIG. 6 illustrates the sequence of an artificial intelligence agent-based digital treatment method (600) for controlling manic-depressive episode transitions according to one embodiment of the present invention.

[0108] The digital treatment method (600) according to the present invention begins with a data collection step (S605). In the step (S605), the data collection unit (310) collects data including behavioral patterns and biosignals from a patient with bipolar disorder. Data collection can be performed in real-time or periodically through a plurality of data channels. Data on the patient's linguistic characteristics and speech characteristics are acquired together through text and voice channels.

[0109] Next, the episode transition control step (S610) is performed. An episode transition control agent (320) analyzes the collected data to calculate control variables for the episode transition interval. It generates proposal information including the transition direction and transition speed between depressive and manic episodes and a stable state. At this time, the data is normalized by comparing it with the individual patient's stable baseline, and the transition acceleration is calculated together. Additionally, the expected duration and remaining time are estimated through a statistical survival analysis model.

[0110] In the subsequent policy confirmation step (S615), the supervisor agent (330) determines the final treatment strategy. The supervisor agent (330) detects whether there is a conflict between the proposal information received from multiple sub-agents. Combinations that are mutually contradictory or violate safety constraints are identified as conflict pairs. The conflicts are adjusted according to predefined priority rules, and if necessary, the agent is requested to modify the proposal. Through the adjustment process, the final treatment intervention policy optimized for the patient is confirmed.

[0111] In the customized treatment provision stage (S620), the intervention execution unit (340) performs customized treatment intervention on the patient according to the established policy. The customized treatment intervention includes at least one of providing treatment content, controlling medication reminders, connecting emergency medical personnel, and transmitting signals to control the user's surrounding environment, and this intervention is provided customizedly at an optimal time through the patient's user terminal (200). The level of exposure of the content can be controlled according to the intensity and frequency specified in the treatment intervention policy.

[0112] Finally, a closed-loop learning step (S625) is performed. The relearning unit (350) collects data on changes in patient response and episode transition characteristics after treatment intervention. The collected feedback data is used to update the control rules of the episode transition control agent (320) and the supervisor agent (330). Through this relearning process, the system is continuously optimized for the individual patient's transition pattern.

[0113] The configuration and operation method of the aforementioned system are merely examples. The technical concept of the present invention is not limited to a specific hardware structure. Each module of the system may be implemented as a set of software instructions. Instructions may be recorded on a computer-readable storage medium and executed.

[0114] The features, structures, effects, etc. described in the embodiments above are included in at least one embodiment of the present invention and are not necessarily limited to only one embodiment. Furthermore, the features, structures, effects, etc. exemplified in each embodiment may be combined or modified and implemented in other embodiments by a person skilled in the art to which the embodiments belong. Accordingly, details regarding such combinations and modifications should be interpreted as being included within the scope of this disclosure. Explanation of the symbols

[0116] 100 servers 200 user terminals 300 AI Agent-Based Digital Therapy System for Mood Disorder Control 310: Data Collection Unit 320: Illustration Transition Control Agent 330: Supervisor Agent 340: Intervention Execution Unit 350: Re-learning Department 360: Control unit 321: Switching Direction Control Agent 322: Switching Rate Control Agent 323: Strength and Safety Control Agent 324: Duration Control Agent

Claims

Claim 1 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, comprising: a data collection unit that collects data from a patient; an episode transition control agent that calculates control variables including episode transition direction and transition speed and generates therapeutic intervention proposal information; a supervisor agent that determines a final therapeutic intervention policy; and an intervention execution unit that provides customized therapeutic intervention to the patient. Claim 2 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, wherein the data collection unit collects data from a patient through a plurality of data channels including at least one of text, voice, behavioral patterns, and biosignals. Claim 3 In paragraph 2, the illustration transition control agent comprises at least one of a transition direction control agent, a transition speed control agent, an intensity and safety control agent, and a duration control agent, in an artificial intelligence agent-based digital therapeutic system for controlling mood disorders. Claim 4 In paragraph 3, the switching direction control agent is configured to calculate the reliability of each of the plurality of data channels in real time to assign optimal weights to individual channel information, fuse the channel information reflecting the weights to generate a state change feature expression, and calculate a probability distribution for each episode switching direction from the feature expression to estimate the direction of episodes with a high probability of transition in a time series, thereby forming an artificial intelligence agent-based digital therapeutic system for controlling mood disorders. Claim 5 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, wherein, in paragraph 4, the switching direction control agent calculates at least one of emotional expression, thought flow change, speech characteristics, life rhythm change, and biovariability index for each of the plurality of data channels to extract features before assigning optimal weights to the individual channel information, and normalizes the extracted features by comparing them with the individual patient's stable baseline to form relative deviations. Claim 6 In paragraph 3, the switching speed control agent calculates a switching speed, which is a rate of change per unit time, from collected time-series data, calculates a switching acceleration, which is a change amount of the switching speed, and compares and analyzes the magnitudes of the switching speed and the switching acceleration with preset reference values ​​to identify the speed grade of the episode switching and the actual switching start time, an artificial intelligence agent-based digital therapeutic system for controlling mood disorders. Claim 7 An AI agent-based digital therapeutic system for controlling mood disorders, wherein the intensity and safety control agent compares real-time symptom data with symptom-specific risk thresholds, evaluates the risk that a specific therapeutic intervention has on symptoms according to intervention-symptom interaction rules, and generates a set of safety constraints that define the acceptable range of the therapeutic intervention based on the evaluation results. Claim 8 In paragraph 3, the duration control agent calculates the expected duration of the current episode by executing a statistical survival analysis model using past episode history data as input, and estimates the remaining time until the end of the episode based on the elapsed time of the current episode, in an artificial intelligence agent-based digital therapeutic system for controlling mood disorders. Claim 9 An AI agent-based digital therapeutic system for controlling mood disorders, wherein the supervisor agent detects combinations having mutually contradictory treatment directions that violate patient safety constraints as conflict pairs. Claim 10 An AI agent-based digital therapeutic system for controlling mood disorders, wherein, in claim 9, the supervisor agent resolves the conflict pair by applying predefined priority rules and constraints, and performs interaction logic requesting the agent to reanalyze or modify the control variable if conflict resolution is not possible. Claim 11 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, wherein, in claim 1, the episode transition control agent and the supervisor agent are configured to update control rules based on changes in patient response and episode transition characteristics after therapeutic intervention. Claim 12 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, wherein, in claim 1, the customized therapeutic intervention comprises at least one of providing therapeutic content, controlling medication reminders, connecting emergency medical personnel, and transmitting signals to control the user's surrounding environment. Claim 13 An artificial intelligence agent-based digital therapeutic system for controlling mood disorders, wherein, in claim 1, the mood disorders include depression and bipolar disorder. Claim 14 A method of operation of a system for controlling mood disorders, comprising: a data collection unit collecting data from a patient through a plurality of data channels including at least one of behavioral patterns and biosignals; an episode transition control agent calculating control variables including an episode transition direction and a transition speed for an episode transition interval between a depressive episode, a manic episode, and a stable state, and generating treatment intervention proposal information; a supervisor agent detecting whether there is a conflict between the proposal information and adjusting it to determine a final treatment intervention policy; and an intervention execution unit outputting customized treatment intervention information through a display based on the final treatment intervention policy. Claim 15 A method of operation of a system for controlling mood disorders, further comprising a closed-loop learning step for updating control rules of the agents based on changes in patient response and illustration transition characteristics after the output of customized information in claim 14. Claim 16 A computer-readable storage medium having a program recorded thereon for executing the method of paragraph 14 or 15 on a computer.

Images