System and method for calculating continuous mental health indicators based on personalized writing behavior

Abstract

The present invention relates to a system and method for monitoring mental health status by fusing microscopic behavioral data generated during a user's writing process and semantic information of text based on artificial intelligence. The present invention extracts behavioral features based on key input, editing, and input stop events collected from a user terminal, and extracts semantic and emotional features of text using a large language model. A fusion analysis unit performs an integrated analysis by synchronizing the behavioral and semantic features in a time-series manner, and an anomaly detection unit detects deviations from a normal state or signs of inconsistency between behavior and semantics by comparing them with a personal baseline that is learned and adaptively updated for each user. The mental health risk indicators calculated through this process are provided as continuous numerical and visualized information, and specific behavioral patterns or text segments that served as the basis for the risk calculation can be displayed together in an explainable form. According to the present invention, potential psychological state changes that are difficult to capture through surveys or single-modal analysis can be detected precisely and non-invasively in a personalized manner.

Description

Technology Field

[0001] The present invention relates to a system and method for calculating risk indicators relative to an individual baseline by analyzing behavioral data and text semantic information generated during a user's writing process using artificial intelligence, and more specifically, to a technology that detects abnormal signs related to the user's condition based on digital biomarkers that fuse writing behavior and text semantics, and visualizes and provides the findings. Background Technology

[0003] Recently, technologies have been proposed to non-invasively monitor mental health status by utilizing various digital behavioral data generated from smartphones, wearable devices, or user terminals to estimate a user's mental state or risk level. Some of these conventional technologies have a structure that collects digital activity data, such as app usage history, movement information, and call records, and inputs this into a machine learning model to calculate the amount of change relative to an individual's baseline or to provide an alert when a threshold is exceeded.

[0004] In addition, multimodal analysis techniques are known to analyze user-generated text to infer mental characteristics such as emotions, sentiments, or cognitive distortions, or to evaluate mental health status by combining linguistic data and digital behavioral data. Furthermore, digital biomarker-based health monitoring techniques are also being proposed to establish individual baselines based on biosignals or various sensor data, and to detect abnormal conditions by tracking long-term changes.

[0005] However, these conventional technologies are limited to utilizing macroscopic or indirect digital activity data, such as app usage, location information, and biosensor signals, and have failed to precisely measure and utilize the writing act itself that occurs during the process of a user actually expressing their thoughts in language. In particular, technologies for analyzing microscopic behavioral data—such as the temporal patterns of key inputs, editing and deletion actions, and input lag occurring during the writing process—and the semantic or emotional characteristics of text generated at the same point in time by synchronizing them in time have not been sufficiently disclosed in conventional technologies.

[0006] Furthermore, existing language-based mental health analysis technologies relied primarily on the surface meaning or sentiment scores of text, which had limitations in effectively capturing situations where there was a discrepancy between behavior and meaning—such as when a user verbally alleviates anxiety, repression, or cognitive burden despite actually feeling them. Moreover, there was a problem where the accuracy of personalized change detection was reduced because the technology failed to adequately reflect individual differences in writing habits and modes of expression, and evaluated the state based on fixed standards or group averages.

[0007] Accordingly, there is a growing demand for a new type of digital biomarker technology capable of detecting anomalies relative to individual user baselines by precisely collecting behavioral data—such as keystrokes, editing, and input lag—during the writing process, and integrating this with text semantic and sentiment analysis results using large language models within the same session. In particular, technologies that identify discrepancies between behavior and meaning, microscopic behavioral changes before and after the input of specific words, and provide these analysis results as continuous risk indicators, visualizations, and notifications are recognized as critical tasks that can complement the limitations of conventional technologies.

[0008] The present invention is proposed to solve the problems of the prior art. Prior art literature

[0010] Korean Published Patent KR1020230013793, Korean Registered Patent KR102650936, US Registered Patent US12220237B2, US Registered Patent US11810670B2 The problem to be solved

[0011] The present invention was developed to improve upon the aforementioned problems, and aims to resolve the limitations of conventional mental health analysis technology, which relies primarily on survey responses, single-text analysis, or general digital activity data, thereby failing to precisely capture the minute moments when changes in a user's actual psychological state occur. In particular, the invention aims to resolve the problem of failing to comprehensively consider behavioral signals inherent in the act of writing itself, which occurs during the process of a user expressing their thoughts in language, and the semantic and emotional characteristics of text generated at the same time.

[0012] Furthermore, conventional technology evaluates the state based on group averages or fixed standards, failing to adequately reflect individual writing habits, modes of expression, and cognitive characteristics; consequently, this leads to a problem of reduced accuracy in detecting personalized changes. The present invention aims to more precisely detect abnormal signs relative to an individual's unique normal state by comparing and analyzing changes in writing behavior and text semantics based on an individual baseline learned for each user.

[0013] Furthermore, conventional language-based analysis techniques have limitations in effectively identifying discrepancies between linguistically expressed emotions and actual cognitive burdens or states of psychological repression. The present invention aims to provide a technology capable of detecting potential risk signals, including discrepancies between behavior and meaning, by analyzing input delay, modification, and deletion behaviors in the writing process in temporal correlation with text meaning.

[0014] Furthermore, if the analysis results are limited to simple binary classification or qualitative judgment, there is a problem in that it is difficult to intuitively grasp the trends of state changes or utilize them for early response. Accordingly, the present invention aims to enable users or administrators to intuitively recognize state changes by calculating detected abnormal signs in the form of continuous risk indicators and providing them in conjunction with visualization and notification functions.

[0015] To solve these problems, the present invention aims to provide a system and method for calculating a continuous mental health risk indicator using a personal baseline-based writing behavior-semantic fusion digital biomarker, which detects abnormal signs related to mental health relative to a personal baseline and calculates and visualizes a continuous risk indicator based on a digital biomarker that fuses behavioral data generated during the writing process with semantic information of text. means of solving the problem

[0017] To solve the above-mentioned problem, the present invention provides a system and method for integrally processing behavioral data and semantic information of text generated during a user's writing process.

[0018] To this end, the system according to the present invention includes an event collection unit that collects writing events, such as key input, editing, and input cessation, input from a user terminal along with time information. The data collected by the event collection unit is converted into behavioral features including writing speed, modification frequency, and input pattern, and a behavioral feature extraction unit is provided to perform this.

[0019] Meanwhile, for text written by a user, a semantic and emotional feature extraction unit is provided to extract semantic and emotional features, including contextual semantic features and emotional features based on text expressions obtained during the process of generating the contextual semantic features, using a large language model. The behavioral features generated by the behavioral feature extraction unit and the semantic and emotional features generated by the semantic and emotional feature extraction unit are transmitted to a fusion analysis unit configured to combine them at the same writing session level to generate an integrated feature vector.

[0020] The anomaly detection unit according to the present invention is configured to analyze changes or deviations from a normal state and detect potential anomaly signs by comparing the integrated feature vector with a personal baseline, which is a reference pattern learned for each user. At this time, in order to reflect differences in individual writing habits and expression styles, the personal baseline is formed based on the user's accumulated writing data and can be updated by learning changes in input habits over time.

[0021] In addition, the indicator output unit according to the present invention is configured to calculate mental health-related risk indicators based on the anomaly detection results, provide them by visualizing them, or output them in conjunction with a notification function. Through this, a user or administrator can intuitively recognize the trend of changes in the state and, if necessary, take early action.

[0022] Meanwhile, in addition to being implemented through the system configuration described above, the present invention may also be implemented by a method comprising the steps of: collecting input events occurring during a user's writing process along with time information; extracting behavioral characteristics from said input events; extracting semantic and emotional characteristics of text using a large language model; deriving integrated psychological characteristics by fusing said behavioral characteristics and semantic and emotional characteristics; detecting abnormal signs through comparison with an individual baseline; and calculating and visualizing risk indicators based on the detected abnormal signs.

[0023] The present invention implements a system and method that detects abnormal signs related to mental health relative to an individual baseline and calculates and provides a continuous risk indicator based on a digital biomarker that fuses writing behavior and text meaning through the organic combination of such configurations. Effects of the invention

[0025] According to the present invention, by fusing and analyzing behavioral data such as key input, editing, and input delay occurring during a user's writing process with the semantic or emotional characteristics of the text at the same point in time, it is possible to precisely detect subtle signs of changes in mental health status that are difficult to capture with conventional survey-based or single-text analysis methods.

[0026] Furthermore, by comparing and analyzing changes in writing behavior and text semantics based on individual baselines learned for each user, it is possible to detect customized anomalies that reflect an individual's unique writing habits and cognitive characteristics, compared to conventional technologies that rely on group averages or fixed standards.

[0027] In addition, by considering the discrepancy between linguistically expressed emotional states and input instability or cognitive burden appearing in actual writing acts, the present invention can more reliably detect potential risk signals that are difficult to identify through surface text analysis alone.

[0028] Furthermore, by generating detected abnormal signs in the form of continuous risk indicators and providing them in conjunction with alerts or visualizing them, users or managers can intuitively recognize trends in mental health status changes and take early action. Brief explanation of the drawing

[0030] FIG. 1 is a block diagram illustrating the overall configuration and data processing flow of a writing behavior-based personalized mental health continuous indicator calculation system according to one embodiment of the present invention. FIG. 2 is a diagram illustrating a collection structure of input events occurring during a writing process according to an embodiment of the present invention. FIG. 3 is a conceptual diagram illustrating an analysis structure that fuses behavioral features and text semantic features according to one embodiment of the present invention. FIG. 4 is a diagram illustrating the flow of anomaly detection and risk signal generation relative to a personal baseline according to one embodiment of the present invention. FIG. 5 is a diagram illustrating an example of visualization and alarm provision of mental health risk indicators according to an embodiment of the present invention. FIG. 6 is a block diagram illustrating a detailed configuration in which a word-behavior association analysis engine is added within a fusion analysis unit according to one embodiment of the present invention. FIG. 7 is a block diagram illustrating the detailed configuration of an anomaly detection unit based on behavioral and semantic mismatch analysis according to one embodiment of the present invention. FIG. 8 is a block diagram illustrating the detailed configuration of a personal baseline management unit and an adaptive baseline update mechanism according to one embodiment of the present invention. FIG. 9 is a block diagram illustrating the detailed configuration of an alarm unit that provides alarm notifications based on user-customized thresholds and pattern analysis according to an embodiment of the present invention. FIG. 10 is a flowchart illustrating the overall processing procedure of a method for calculating a continuous mental health risk indicator according to one embodiment of the present invention. Specific details for implementing the invention

[0031] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings.

[0032] However, the present invention is not limited to the embodiments disclosed below but will be implemented in various different forms.

[0033] The embodiments described in this specification are provided to ensure that the disclosure of the invention is complete and to fully inform those skilled in the art of the scope of the invention.

[0034] And the present invention is defined only by the scope of the claims.

[0035] Accordingly, in some embodiments, well-known components, well-known operations, and well-known techniques are not specifically described to avoid the present invention being interpreted ambiguously.

[0036] Additionally, throughout the specification, the same reference numerals refer to the same components, and the terms used (mentioned) in this specification are for describing embodiments and are not intended to limit the invention.

[0037] In this specification, the singular form includes the plural form unless specifically stated otherwise in the text, and components and operations referred to as 'comprising (or comprising)' do not exclude the presence or addition of one or more other components and operations.

[0038] Unless otherwise defined, all terms used in this specification (including technical and scientific terms) may be used in a meaning that is commonly understood by those skilled in the art to which the present invention belongs.

[0039] Also, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless otherwise defined.

[0040] Hereinafter, preferred embodiments of the present invention will be described with reference to the attached drawings.

[0041] Referring to FIG. 1, a mental health continuous risk indicator calculation system (100) according to one embodiment of the present invention is configured to calculate a risk indicator related to a mental health state by integrally processing behavioral data and semantic information of text that occur during the user's writing process.

[0042] The above system (100) includes an event collection unit (120), a behavioral feature extraction unit (130), a semantic / emotional feature extraction unit (140), a fusion analysis unit (150), a personal baseline management unit (160), an anomaly detection unit (170), and an indicator output unit (180), and may further include an alarm unit (190).

[0043] The above system (100) is connected to an external user terminal (110) to transmit and receive data via a network, and is configured to operate in conjunction with a data storage unit (200) for storing and loading data required for analysis.

[0044] At this time, the event collection unit (120) collects key input, editing, and input stop (Idle) events input from the user terminal (110) along with time information, and the behavioral feature extraction unit (130) extracts behavioral features including writing speed, modification frequency, and input pattern based on the collected events.

[0045] Meanwhile, the semantic and emotional feature extraction unit (140) extracts semantic and emotional features from text written by a user using a large language model, and the fusion analysis unit (150) combines the behavioral features and semantic and emotional features to generate an integrated feature vector.

[0046] The anomaly detection unit (170) detects anomalies by comparing the integrated feature vector with the personal baseline, which is a standard pattern learned for each user, and the indicator output unit (180) calculates a mental health risk indicator based on the anomaly detection result and provides it by visualizing it.

[0048] Referring to FIG. 2, an event collection unit (120) according to one embodiment of the present invention includes a key input event collection module (121), an edit / delete event collection module (122), and an input stop event collection module (123) to precisely collect various input actions occurring during a user's writing process as time-series data.

[0049] The above event collection unit (120) logs each input event occurring during the user's writing process in the form of time-series data along with a timestamp, and stores metadata including not only the input characters themselves but also the time interval between inputs, the time of modification, and the input stop interval.

[0050] The above key input event collection module (121) is configured to detect each key input event occurring from the input interface of a user terminal and to collect key input data including the input character code, the timestamp of the time of input occurrence, and the time interval between consecutive inputs.

[0051] Specifically, the key input event collection module (121) can generate basic behavioral data indicating changes in the user's input speed, irregularity of the input rhythm, and instability of the input pattern by calculating and storing the time interval between inputs between consecutive key input events.

[0052] The above-mentioned edit / delete event collection module (122) is configured to detect modification, deletion, and re-entry actions performed on text already entered by a user, and to collect edit event data including the time of the edit, the length of the deleted character or string, the number of times the edit was repeated, and the edit location information.

[0053] In particular, the edit / delete event collection module (122) preserves deleted text that does not remain in the final result as log data, thereby enabling the user's self-censorship or thought correction process to be tracked at the level of behavioral data.

[0054] The above input stop event collection module (123) is configured to define a state in which no key input occurs for a certain period of time or longer as an input stop event, and to calculate and record the start time, end time, and duration of the input stop period.

[0055] At this time, the input stop event collection module (123) can use metadata to identify input gaps that persist for longer than a threshold time set based on the user's normal input pattern as input interruption due to cognitive hesitation or emotional burden.

[0056] In this way, the event collection unit (120) integrates and stores key input, edit / delete, and input stop events as timestamp-based time-series log data, and the collected event data is subsequently transmitted to the behavioral feature extraction unit (130) and used to quantify behavioral features such as writing speed, modification pattern, and input stability.

[0057] Through this, the present invention can precisely reflect the user's inherent psychological state, which is difficult to capture solely through text semantic analysis, using microscopic behavioral data from the writing process.

[0059] A behavioral feature extraction unit (130) according to one embodiment of the present invention is a module for quantifying behavioral characteristics that appear during a user's writing process by analyzing time-series input events transmitted from an event collection unit (120), and includes an input speed calculation module (131), a modification frequency calculation module (132), and an input pattern analysis module (133).

[0060] The above input speed calculation module (131) is configured to calculate the number of input characters per unit time, the time interval between average inputs, and the variance of the input speed based on the timestamp information of the key input event transmitted from the event collection unit (120).

[0061] Specifically, the input speed calculation module (131) can statistically process the time interval between consecutive key input events to calculate the user's average input speed and instantaneous speed change amount, thereby extracting a decrease in input speed or a sudden fluctuation in input speed that indicates instability of the input pattern as a behavioral feature.

[0062] The above modification frequency calculation module (132) is configured to analyze modification and deletion events transmitted from the edit / deletion event collection module (122) to calculate the number of modification occurrences per unit time, the average length of deleted characters or strings, and the repetitive modification rate.

[0063] In particular, the modification frequency calculation module (132) can generate quantitative behavioral indicators that reflect the user's tendency to self-censor, increased uncertainty, or frequency of thought reconstruction by identifying modification and deletion patterns that occur repeatedly in the same sentence or the same input section.

[0064] The above input pattern analysis module (133) is configured to analyze high-dimensional input pattern features, including input delay distribution, regularity of input rhythm, and input interruption and resumption patterns, by synthesizing the results calculated from the input speed calculation module (131) and the modification frequency calculation module (132).

[0065] For example, the input pattern analysis module (133) can detect potential psychological state changes that are not revealed in the final output of the text at the level of behavioral patterns by analyzing whether input stop events are concentrated before and after the input of a specific keyword, or whether a sudden change in input speed is associated with a specific context segment.

[0066] The behavioral features related to the input speed, correction frequency, and input pattern calculated in this way are composed of behavioral feature data in the form of vectors and transmitted to the fusion analysis unit (150), and combined with semantic features to be used in integrated analysis for calculating mental health risk indicators.

[0068] A semantic and emotional feature extraction unit (140) according to one embodiment of the present invention is configured to quantify contextual semantics and emotional states from text written by a user, and includes a large language model-based semantic analysis module (141) and an emotional feature extraction module (142).

[0069] The above-mentioned semantic and emotional feature extraction unit (140) is configured to convert sentence structure, semantic association, and emotional nuance into a quantified feature vector form, beyond the surface expression of the text, so that it can be processed in a subsequent analysis module.

[0070] The above semantic analysis module (141) is configured to convert text entered by a user into a high-dimensional vector in an embedding space using a Large Language Model.

[0071] Specifically, the semantic analysis module (141) breaks down the text into tokens and then generates semantic vectors that reflect the semantic similarity between sentences, context flow, and core topic distribution through the embedding layer or intermediate representation layer of the large language model.

[0072] Through this, contextual consistency, semantic emphasis sections, and topic transition patterns can be quantitatively expressed, rather than simply the presence or absence of keywords.

[0073] The above-mentioned emotional feature extraction module (142) is configured to quantify the emotional attributes inherent in the text based on the text expression processed by the semantic analysis module (141).

[0074] Specifically, the emotional feature extraction module (142) calculates the emotional polarity (positive / neutral / negative), emotional intensity, and emotional change trend of the text, and can calculate an emotional score on a sentence-by-sentence or token-by-token basis.

[0075] In particular, the emotion feature extraction module (142) can output subtle changes between the surface expression of the text and the actual emotional state in the form of a feature vector by analyzing the pattern of change in emotion scores within the same session.

[0076] The result generated by the above semantic analysis module (141) and emotional feature extraction module (142) is integrated into one or more semantic and emotional feature vectors, and the corresponding feature vector is transmitted to the fusion analysis unit (150) and used for combined analysis with behavioral features.

[0077] At this time, the above semantic and sentiment feature vectors are configured to be linked with time information or text segment information, thereby enabling time-series synchronization analysis with writing behavior data.

[0079] Referring to FIG. 3, a fusion analysis unit (150) according to one embodiment of the present invention is configured to combine behavioral features and semantic features generated during the writing process into the same analysis unit, and includes a session synchronization module (151) and an integrated feature vector generation module (152).

[0080] The above fusion analysis unit (150) aligns heterogeneous forms of data, such as behavior-based time series data and text-based semantic data, along the time axis and combines them in a quantitative feature space to generate an integrated feature vector suitable for subsequent anomaly detection.

[0081] The above session synchronization module (151) is configured to map time-series behavioral features transmitted from the behavioral feature extraction unit (130) and text-based semantic and sentimental features transmitted from the semantic feature extraction unit (140) to the same writing session unit.

[0082] Specifically, the session synchronization module (151) aligns the text generation time and behavioral characteristics that occurred before and after that time based on the timestamp information assigned to the input event, and temporally links behavioral characteristics corresponding to a specific text section.

[0083] Through this, it is configured to enable the analysis of not only the final result of the text but also the behavioral changes that occurred during the text generation process, along with semantic features.

[0084] The above integrated feature vector generation module (152) is configured to receive behavioral features and semantic features aligned by the session synchronization module (151) and to generate one or more integrated feature vectors.

[0085] Specifically, the integrated feature vector generation module (152) generates an integrated expression that simultaneously reflects the act of writing and the meaning of text by mapping the behavioral feature and the semantic feature to the same feature space, or by aligning them through a mapping function between different feature spaces, and then combining them by assigning weights to each feature.

[0086] In this case, the aforementioned weights can be dynamically adjusted according to semantic importance, behavioral variability, or trends of change within the session, thereby enabling a vector representation that sensitively reflects changes in psychological state.

[0087] The integrated feature vector generated by the integrated feature vector generation module (152) is transmitted to the anomaly detection unit (170) and compared and analyzed with the user's unique reference pattern stored in the personal baseline management unit (160).

[0088] Through this, complex psychological changes that are difficult to capture with a single behavioral feature or a single semantic feature alone are quantitatively expressed, and the accuracy of detecting signs of mental health risk is improved.

[0090] Referring to FIG. 4, a personal baseline management unit (160) according to one embodiment of the present invention is configured to form and maintain a long-term standard pattern of writing behavior and semantic characteristics for each user, and includes an adaptive baseline learning module (161) and a personal baseline update module (162).

[0091] The above personal baseline management unit (160) is configured to generate personalized standard patterns by taking into account differences in writing habits, input proficiency, and emotional expression methods for each user, and to distinguish between normal changes and abnormal signs over time.

[0092] The above adaptive baseline learning module (161) is configured to receive the user's initial or accumulated writing data and learn a standard pattern representing an individual normal state.

[0093] Specifically, the adaptive baseline learning module (161) calculates the mean and variance of the input speed distribution, correction frequency, input delay pattern, and semantic and sentimental features based on feature data transmitted from the behavioral feature extraction unit (130) and the semantic feature extraction unit (140), and defines these as the user's unique normal range.

[0094] At this time, the above learning is performed on multiple writing session data accumulated over a certain period, and may be configured to reflect only data satisfying statistical stability conditions in order to exclude the influence of short-term fluctuations or temporary outliers.

[0095] The above individual baseline update module (162) is configured to adaptively adjust the reference pattern formed by the adaptive baseline learning module (161) over time.

[0096] Specifically, the individual baseline update module (162) is configured to detect whether newly input writing data shows a gradual and consistent change trend compared to the existing baseline, and to update the baseline value only when the change satisfies a preset update condition.

[0097] Through this, normal pattern shifts, such as improvements in user input proficiency, changes in the usage environment, or changes in natural habits, are reflected in the baseline only under predefined stability conditions, while sudden or abnormal changes can be detected as signs of mental health risk by the anomaly detection unit (170).

[0098] The personal baseline formed and maintained by the personal baseline management unit (160) is provided to the anomaly detection unit (170) and used as a comparison standard with the integrated feature vector.

[0099] Thus, the present invention enables precise anomaly detection based on each user's unique long-term patterns, rather than uniform analysis relying on average values ​​among users.

[0101] Referring to FIG. 7, the anomaly detection unit (170) is configured to detect anomalies related to mental health status by comparing an integrated feature vector, in which behavioral features and semantic features are fused, with a user-specific standard pattern stored in the personal baseline management unit (160).

[0102] To this end, the above-mentioned anomaly detection unit (170) includes a behavioral / meaning discrepancy analysis module (171) and an anomaly detection module (172).

[0103] The above behavior-meaning discrepancy analysis module (171) is configured to analyze the integrated feature vector generated in the same writing session to evaluate whether there is a match between the surface semantic characteristics of the text and the writing behavior characteristics.

[0104] Specifically, the behavioral and semantic inconsistency analysis module (171) detects whether the behavioral features exhibit patterns suggesting cognitive load, such as increased input delay, increased frequency of modification and deletion, or increased input instability, even though the semantic features belong to a positive or neutral emotional category.

[0105] At this time, the behavior-meaning discrepancy analysis module (171) can quantitatively calculate the degree of discrepancy between behavior and meaning by analyzing the correlation, magnitude of deviation, or directionality between the behavior feature element and the meaning feature element within the integrated feature vector.

[0106] The above abnormal sign detection module (172) is configured to finally detect whether the discrepancy is an abnormal sign outside the normal range by comparing the discrepancy result calculated by the behavioral / semantic discrepancy analysis module (171) with the user's unique standard pattern stored in the personal baseline management unit (160).

[0107] Specifically, the anomaly detection module (172) can detect anomalies based on whether the integrated feature vector of the current session exceeds the statistical confidence interval of the individual baseline, and can be configured to consider not only the results of a single session but also repeatability or trend deterioration over multiple sessions. Through this, fluctuations caused by temporary fatigue, environmental changes, or accidental input errors can be effectively distinguished from persistent mental health risk signals.

[0108] The abnormal sign information detected by the above abnormal detection unit (170) is transmitted to the indicator output unit (180) and used as a key input value for calculating the mental health risk indicator, and, if necessary, can be used as a basis information for providing an alarm notification to a user or administrator through the alarm unit (190).

[0110] Referring to FIG. 5, the indicator output unit (180) includes a mental health risk indicator calculation module (181) and a risk indicator visualization module (182) to express the mental health risk state in a continuous numerical form.

[0111] The indicator output unit (180) receives the abnormal sign detection results transmitted from the abnormal detection unit (170), calculates a risk indicator so that a user or manager can intuitively recognize the change in state, and is configured to provide the indicator visually.

[0112] The above mental health risk indicator calculation module (181) is configured to calculate a mental health risk state as a continuous numeric indicator by synthesizing numeric indicators corresponding to the degree, persistence, and frequency of occurrence of abnormal signs detected by the abnormality detection unit (170).

[0113] Specifically, the risk indicator calculation module (181) can calculate a risk index for a single session or multiple sessions by weightedly combining the degree of discrepancy calculated by the behavioral / semantic discrepancy analysis module (171) and the detection results of the abnormal sign detection module (172).

[0114] In this case, the above risk indicator is defined as a continuous indicator that changes over time rather than a discrete "normal / abnormal" determination, and can be configured to reflect the relative amount of change compared to the user's individual baseline.

[0115] In addition, the risk indicator calculation module (181) can quantitatively express the tendency of the mental health condition to deteriorate or recover by considering the rate of increase, the degree of accumulation, or the pattern of change of the risk indicator.

[0116] The above risk indicator visualization module (182) is configured to visualize the risk indicator calculated by the mental health risk indicator calculation module (181) in the form of a graph, index, color change, or icon and provide it to the user or manager.

[0117] Specifically, the risk indicator visualization module (182) can display the trend of change of the risk indicator along the time axis or express it as stepwise or continuous visual elements so that the current risk level can be intuitively recognized.

[0118] In addition, the visualization module (182) can support the user in understanding the cause of the risk indicator by highlighting specific writing sections, input delay patterns, or sentiment keywords that served as the basis for the risk indicator calculation.

[0119] In one embodiment, when a user writes text for a certain period of time through a user terminal (110), the event collection unit (120) collects key input, editing, and input stop events along with timestamps.

[0120] The behavioral feature extraction unit (130) calculates a rapid decrease in input speed and an increase in the frequency of modification as behavioral features, and the semantic / emotional feature extraction unit (140) extracts an increase in emotional negativity within the text as semantic / emotional features.

[0121] The fusion analysis unit (150) generates an integrated feature vector combining the behavioral features and semantic features, and the anomaly detection unit (170) compares this with the user's personal baseline to detect that there is a significant deviation from the normal state.

[0122] The indicator output unit (180) provides the risk indicator calculated by the mental health risk indicator calculation module (181) based on the detection result by visualizing it in the form of a graph of change over time or an index by the risk indicator visualization module (182), thereby enabling the user to intuitively recognize changes in their mental health status.

[0124] Referring to FIG. 6, a system according to one embodiment of the present invention may further include a word-behavior association analysis engine (153) within a fusion analysis unit (150).

[0125] The above word-behavior association analysis engine (153) is configured to analyze the time of generation of specific meaningful words or sentiment keywords within the text written by the user, and the writing behavior data occurring before and after that time in a temporally linked manner.

[0126] To this end, the word-behavior association analysis engine (153) processes the time information of an input event collected by the event collection unit (120) and transmitted through the behavior feature extraction unit (130) and the session synchronization module (151), and synchronizes the time of creation of a specific word in the text identified by the semantic feature extraction unit (140).

[0127] Specifically, the word-behavior association analysis engine (153) sets the tokenization time of a specific semantic word or sentiment keyword extracted from the semantic feature extraction unit (140) as a reference time, and performs association analysis on key input, input pause (Idle), modification, and deletion events that occur within a preset time range centered on the reference time.

[0128] At this time, the word-behavior association analysis engine (153) identifies the behavioral change patterns that appear during the user's writing process by analyzing the input delay time, input density change, or input pause frequency that occurs before a specific word input, and the frequency and pattern of modification or deletion events that occur after a specific word input, respectively.

[0129] Additionally, the word-behavior association analysis engine (153) can be configured to compare the analysis results with a personal baseline learned for each user to derive an analysis result indicating whether the behavioral change occurring when a specific word is entered falls within a general range for that user. Through this, it is possible to distinguish and identify general changes caused by temporary typos or habitual input from significant behavioral patterns that suggest the possibility of psychological repression or self-censorship.

[0130] For example, one can assume a case where a user writes sentences that appear to contain positive content on the surface during the writing process, but experiences a significantly longer input delay than usual immediately before entering a specific emotional keyword, and observes repetitive deletion and re-entry behaviors immediately after entering the keyword.

[0131] In this case, the word-behavior association analysis engine (153) compares the behavioral characteristics before and after the input of the corresponding keyword with the user's personal baseline and identifies that the frequency of input delay and correction has significantly increased compared to when the user normally inputs neutral words. Accordingly, a behavioral pattern suggesting that the user may have been in a state of psychological repression or self-censorship during the process of expressing the content, contrary to the surface meaning of the text, can be detected as a potential danger signal.

[0132] In this way, the word-behavior association analysis engine (153) according to the present invention can detect mental health-related risk signs that are difficult to capture with conventional single text analysis or macroscopic digital activity analysis more precisely by analyzing the semantic information of the text and microscopic behavioral data in the writing process in temporal linkage.

[0134] Referring to FIG. 7, a system according to one embodiment of the present invention includes an anomaly detection unit (170), and the anomaly detection unit (170) includes a behavior-meaning mismatch analysis module (171).

[0135] The above behavioral / semantic discrepancy analysis module (171) is configured to determine risk signs related to mental health status by analyzing the relationship between the semantic / emotional features of the text produced by the semantic / emotional feature extraction unit (140) and the behavioral features, including input delay, correction frequency, and input pattern, produced by the behavioral feature extraction unit (130).

[0136] In addition, the analysis result produced by the above behavior / meaning discrepancy analysis module (171) is transmitted to the abnormal sign detection module (172), and the presence of an abnormal sign can be finally detected through comparison with the user's unique standard pattern stored in the personal baseline management unit (160).

[0137] Specifically, the behavioral and semantic inconsistency analysis module (171) detects whether, even when semantic features belong to a positive or neutral emotional category, behavioral features in the same writing session indicate an increase in input delay, an increase in the frequency of modification and deletion, or input instability above a threshold compared to the user's normal pattern stored in the personal baseline management unit (160).

[0138] At this time, the behavior-meaning discrepancy analysis module (171) can calculate the correlation between the meaning feature vector and the behavior feature vector, or calculate the discrepancy state based on whether the behavior feature of the session falls outside the statistical confidence interval of the individual baseline, and determine based on this.

[0139] For example, if a behavioral characteristic exceeds a pre-set confidence level relative to an individual baseline, this can be detected as a signal of incongruence suggesting cognitive load or a state of psychological repression.

[0140] In addition, the behavioral and semantic discrepancy analysis module (171) may be configured not to be limited to a single session, but to analyze multiple writing session data accumulated in the data storage unit (200) in a time-series manner, and to consider whether the discrepancy phenomenon is repeated or gradually intensifies, thereby distinguishing between temporary input errors and signs of persistent mental health risk.

[0141] In addition, the mental health risk indicator information identified by the behavior / meaning discrepancy analysis module (171) is transmitted to the indicator output unit (180) and used as an input value for calculating the mental health risk indicator, and, if necessary, can be used as a basis information for providing notifications to the user or manager through the alarm unit (190).

[0142] For example, when a user writes a sentence that has a superficially stable meaning, such as "it is okay" or "no problem," the meaning / emotion feature extraction unit (140) can classify the text as a positive or neutral emotional state.

[0143] However, if the behavioral feature extraction unit (130) detects a longer input pause time than usual, repetitive sentence modification and deletion, or irregular input speed change during the same writing process, the behavioral and semantic inconsistency analysis module (171) detects whether the behavioral feature deviates from the user's unique normal range stored in the personal baseline management unit (160).

[0144] As a result, when a stable expression is made in terms of meaning but instability suggesting cognitive burden or psychological repression is simultaneously observed in terms of behavior, the behavior-meaning discrepancy analysis module (171) can identify this as a sign of potential mental health risk due to the discrepancy between behavior and meaning.

[0145] In this way, the abnormality detection unit (170) according to the present invention can detect signs of mental health risk more precisely by complementarily analyzing the inherent psychological state, which is difficult to capture solely by the result of semantic analysis of the text, through comparison with writing behavior.

[0147] Referring to FIG. 8, a system according to one embodiment of the present invention includes a personal baseline management unit (160) for managing personal baselines for each user, and the personal baseline management unit (160) includes an adaptive baseline learning module (161) and a personal baseline update module (162).

[0148] The above adaptive baseline learning module (161) is configured to analyze behavioral characteristics and semantic and emotional characteristic data accumulated during the user's writing process to form a personal baseline representing the user's unique normal writing pattern.

[0149] The above personal baseline update module (162) is configured to update the personal baseline periodically or conditionally based on the analysis results of the above adaptive baseline learning module (161).

[0150] Specifically, the adaptive baseline learning module (161) statistically analyzes the input delay, correction frequency, input speed distribution, and text semantic and sentimental characteristics extracted from the user's multiple writing sessions to generate a standard pattern that reflects the user's usual writing behavior and expression characteristics.

[0151] At this time, the adaptive baseline learning module (161) may be configured to identify gradual pattern shifts that occur over time due to improvements in the user's input proficiency, changes in the device used, changes in age, etc.

[0152] If the above pattern movement is stably maintained within a preset range, the personal baseline update module (162) detects this as a normal change and updates the personal baseline to the latest state.

[0153] On the other hand, an increase in input delay occurring rapidly in a short period, a surge in the frequency of modifications, or behavioral and semantic inconsistencies are excluded from the individual baseline update and are configured to be identified as signs of mental health risk by the anomaly detection unit (170).

[0154] Through this, the present invention can clearly distinguish between natural habit changes and signs of mental health abnormalities, thereby minimizing false positives.

[0155] Additionally, the adaptive baseline learning module (161) may be configured to gradually update the individual baseline by applying a sliding window method or a similar technique that assigns relatively high weight to the data of the most recent n writing sessions.

[0156] For example, when a new user is registered in the system, the personal baseline management unit (160) collects the user's writing data during an initial period and forms an initial personal baseline based on the average input delay, correction frequency, and distribution of semantic and emotional characteristics during that period.

[0157] Subsequently, if the input speed gradually increases or the frequency of correction decreases as the user uses the system for a long period, the adaptive baseline learning module (161) detects this as a normal pattern shift due to improved input proficiency and updates the reference pattern step by step through the personal baseline update module (162).

[0158] On the other hand, if the frequency of input delay and correction increases rapidly at a specific point in time and these changes are repeatedly observed, the changes are excluded from the individual baseline update and are analyzed as signs of mental health risk by the anomaly detection unit (170).

[0159] In this way, the present invention can improve the reliability of calculating mental health risk indicators by accommodating user-specific long-term changes through adaptive baseline learning while effectively detecting short-term and abnormal changes.

[0161] A system according to one embodiment of the present invention can be implemented as a distributed processing structure in which the edge area of ​​a user terminal and a server or cloud area collaborate.

[0162] At this time, the event collection unit (120) and the behavior feature extraction unit (130) are operated in the edge area within the user terminal, and the fusion analysis unit (150) is configured to be deployed and operated in a server or cloud area.

[0163] In the above edge region, key input, editing, and input stop events occurring during the user's writing process are collected, and behavioral features including input delay, modification frequency, and input patterns are extracted from the events.

[0164] The behavioral features extracted in this way are transmitted to the server area in the form of anonymized feature vectors, and raw key input data can be configured not to be transmitted outside the terminal.

[0165] Specifically, the behavioral feature extraction unit (130) may be configured to minimize the possibility of backtracking of personal information by generating a feature vector by extracting only quantitative characteristics such as the statistical distribution of time intervals between inputs, the number of modifications, and the deletion rate, without including the input character sequence or original text.

[0166] Meanwhile, the semantic and sentiment feature extraction unit (140) is placed in a server or cloud area considering the computational load of the large language model and can be configured to produce semantic and sentiment features by analyzing text data received from a user or intermediate expressions generated from the text.

[0167] The above semantic and emotional feature extraction unit (140) is organically combined with the fusion analysis unit (150) located in the server area to perform integrated analysis of the behavioral feature vector and the semantic and emotional feature vector.

[0168] Through such an edge-cloud collaboration structure, the present invention can process sensitive input events requiring personal information protection within the terminal, while efficiently performing semantic analysis and fusion analysis requiring high-performance computation in the server area.

[0169] For example, when a user inputs text on a mobile terminal or personal computing device, key input events are collected in real time at the edge area of ​​the terminal, and behavioral features such as input latency, correction frequency, and input rhythm are extracted.

[0170] In this case, the original text of the input or the key input log is not transmitted outside the terminal, and only the anonymized behavioral feature vector is transmitted to the server.

[0171] In the server area, the semantic and emotional feature extraction unit (140) calculates semantic and emotional features of the text using a large language model, and the fusion analysis unit (150) combines the semantic and emotional features and behavioral features to generate an integrated feature vector.

[0172] Through this structure, the present invention can simultaneously secure personal information protection and computational efficiency, while performing precise analysis required for calculating mental health risk indicators.

[0174] Referring to FIG. 9, a system according to one embodiment of the present invention further includes an alarm unit (190) for providing an alarm notification based on the result of calculating a mental health risk indicator.

[0175] The above alarm unit (190) is configured to receive a mental health risk indicator calculated and visualized by the indicator output unit (180) and to provide alarm information regarding changes in mental health status to a user or manager.

[0176] At this time, the alarm unit (190) may be configured to operate in functional linkage with the indicator output unit (180), but to be placed as an independent component within the system to perform the functions of detecting risks and providing notifications.

[0177] Specifically, the alarm unit (190) may include a threshold detection module (191) and an alarm notification providing module (192).

[0178] The threshold detection module (191) may be configured to detect whether the calculated mental health risk indicator exceeds a customized threshold set for each user, or to comprehensively analyze whether the risk indicator shows an upward trend, rapid fluctuations, or an increase in the frequency of high-risk conditions over a certain period.

[0179] At this time, the threshold detection module (191) can set a dynamic threshold based on the user's normal fluctuation range provided by the personal baseline management unit (160), and the threshold can be configured to adaptively adjust according to the user's long-term state change.

[0180] The above alarm notification providing module (192) may be configured to provide an alarm notification based on the detection result through a notification interface of a user terminal, a monitoring screen for an administrator, or a communication means connected to an external system.

[0181] In addition, the above alarm notification may be provided at stepwise levels, such as caution and warning.

[0182] Meanwhile, the above-mentioned alarm notification may be linked with visualization information of the indicator output unit (180) and may be provided together with specific writing behavior characteristics or text semantic elements that served as the basis for the alarm generation, thereby improving the user's intuitive understanding and reliability.

[0183] For example, if the mental health risk indicator calculated based on the user's writing behavior and text semantic analysis results exceeds a user-customized threshold, the alarm unit (190) can detect this and provide the user with a warning or alert.

[0184] In addition, if the above risk indicators repeatedly rise within a short period or the frequency of occurrence of high-risk conditions increases rapidly compared to normal, the alarm unit (190) may be configured to detect this as a continuous risk condition and provide an alarm notification to an administrator or guardian.

[0185] In this way, the alarm unit (190) according to the present invention provides an alarm by considering not only the absolute value of the risk indicator but also the change pattern and frequency of occurrence, thereby enabling early recognition and preemptive response to changes in mental health status.

[0187] Referring to FIG. 10, a method for calculating a continuous mental health risk indicator according to one embodiment of the present invention is configured to calculate a risk indicator by processing behavioral data and semantic information of text generated during a user's writing process in stages.

[0188] The above method includes an input event collection step (a), a behavioral feature quantification step (b), a semantic feature extraction step (c), a fusion analysis step (d), an anomaly detection step (e), and a risk indicator calculation and provision step (f).

[0189] (a) Input event collection step

[0190] The above input event collection step (a) is performed by an event collection unit (120) and collects key input, edit, delete, and input stop events occurring through the input interface of a user terminal along with timestamp information.

[0191] The collected input events are stored in a time-series manner in a data storage unit (200) and can be used for analysis in subsequent stages.

[0192] (b) Step of quantifying behavioral characteristics

[0193] The above behavioral feature quantification step (b) is performed by the behavioral feature extraction unit (130), and quantifies behavioral features representing writing behavior, such as input delay time, frequency of modification and deletion, and change in input speed, from the input event.

[0194] At this time, the time-series intervals between logged key input events can be converted into statistical distribution values ​​and calculated as an indicator representing input consistency or instability.

[0195] (c) Semantic feature extraction step

[0196] The above semantic and emotional feature extraction step (c) is performed by a semantic and emotional feature extraction unit (140), and uses a large language model to extract semantic and emotional features including contextual semantic features and emotional features based on text expressions obtained during the process of producing the above contextual semantic features from text written by a user.

[0197] According to one embodiment, in the step, emotional polarity or semantic directionality can be quantified by converting text into a high-dimensional vector form through the embedding layer of a large language model.

[0198] (d) Fusion analysis step

[0199] The above fusion analysis step (d) is performed by the fusion analysis unit (150), and the behavioral characteristics and semantic characteristics are combined in a multimodal manner to derive an integrated feature vector representing integrated psychological characteristics.

[0200] In this case, behavioral data and text semantic information are linked and analyzed at the same writing session level, thereby allowing the temporal and contextual relationship between input behavior and semantic expression to be reflected.

[0201] (e) Abnormal sign detection step

[0202] The above abnormal sign detection step (e) is performed by the abnormal detection unit (170), which analyzes deviations or changes relative to the normal state by comparing the integrated psychological characteristics with a personal baseline set as a unique pattern for each user.

[0203] Through this, combinations of behavior and meaning that differ from an individual's usual writing habits can be detected as abnormal signs.

[0204] (f) Risk indicator calculation and provision step

[0205] The above risk indicator calculation and provision step (f) is performed by the indicator output unit (180), and the mental health risk indicator is calculated in a continuous numerical form based on the detected abnormal signs.

[0206] The calculated risk indicators are visualized and provided to users or administrators, enabling them to intuitively recognize trends in state changes over time.

[0207] For example, if a user used expressions that were semantically stable in a specific writing session but experienced significant input delays and repetitive correction behaviors compared to normal during the input process, the above method can detect abnormal signs by comparing the integrated psychological characteristics of the session with the individual baseline.

[0208] As a result, the relevant anomaly can be converted into a continuous risk value during the risk indicator calculation stage and presented visually.

[0209] In this way, the method according to the present invention can continuously track changes in a user's mental health state and provide them as risk indicators by collecting, transforming, fusing, and analyzing writing behavior and text meaning in stages.

[0211] A method for calculating a mental health risk indicator according to one embodiment of the present invention includes step (d), that is, a step of additionally analyzing a sequence of behaviors after the input stoppage that occurs during the writing process in the process of fusing behavioral characteristics and semantic characteristics in a multimodal manner.

[0212] The above step is configured to identify potential emotional instability signals that are not explicitly revealed in the final text output by analyzing sequential behavioral patterns in which the user inputs or deletes specific emotional vocabulary after pausing input for a certain period of time or longer.

[0213] Specifically, the behavior sequence analysis step is performed by the fusion analysis unit (150), and among the input events provided by the behavior feature extraction unit (130), an input idle time (Idle-time) greater than or equal to a threshold is set as a reference event.

[0214] At this time, the threshold of the input pause time is not a fixed value, but can be set based on the user's normal average key input interval (Inter-Keystroke Interval, IKI) and its variance or standard deviation, and can be defined as a variable threshold to which weights are dynamically applied according to changes in the user's current input speed.

[0215] The above-mentioned fusion analysis unit (150) defines word input, modification, or deletion events occurring within a certain time interval in front or behind the point in time when the input stop time occurs as a single action sequence, and detects whether the word included in the sequence has a specific emotional vocabulary or emotional meaning in conjunction with the extraction result of the meaning / emotion feature extraction unit (140).

[0216] In addition, the fusion analysis unit (150) can calculate a Psychological Load Index by combining the emotional polarity or intensity of the identified emotional vocabulary and the length of the input pause time, and if a specific emotional vocabulary is deleted after being input, it can reflect in the analysis by assigning a weight to the negative polarity of the vocabulary.

[0217] Furthermore, the aforementioned behavioral sequence is not limited to a single input stop event but can be defined to include the cumulative frequency, occurrence interval, or distribution characteristics of input stop-word input or deletion patterns that occur repeatedly within a certain period. Through this, behavioral patterns caused by temporary hesitation and continuous, repetitive psychological burden can be distinguished more precisely.

[0218] The results of the above behavioral sequence analysis are interpreted by comparing them with the user-specific normal writing behavior patterns stored in the personal baseline management unit (160).

[0219] In other words, it detects whether the input or deletion of emotional vocabulary after input cessation is a normal habit that occurs routinely for the user, or an abnormal pattern that increases sharply or repeats at a specific point in time, based on deviations from the individual baseline.

[0220] Through this, the above method effectively reduces the possibility of false positives by distinguishing between input pauses caused by temporary hesitation or typos and input pauses caused by psychological anxiety, avoidance, or suppression.

[0221] For example, assume a case where a user repeatedly exhibits behavior of typing emotional vocabulary such as "anxious" or "difficult" and then immediately deleting it after a significantly longer input pause compared to the usual average input interval while writing.

[0222] In this case, even if the final text output does not include the corresponding emotional vocabulary, the behavioral sequence analysis step analyzes the input pause time, the emotional polarity of the emotional vocabulary, and the deletion behavior in temporal correlation, and detects whether the pattern has a significant deviation from the individual baseline.

[0223] As a result, if the above behavioral sequence is observed continuously or repeatedly, the method according to the present invention can identify it as a danger signal suggesting a potential emotional instability or psychological avoidance state not explicitly expressed in the text.

[0224] As such, unlike conventional technology that utilizes only text analysis results, the present invention analyzes the continuous flow of hesitation and word selection and deletion during the writing process, thereby enabling more precise detection of the psychological state inherent in the user's thought process.

[0226] A method for calculating a mental health risk indicator according to one embodiment of the present invention includes step (d), that is, in the process of deriving integrated psychological characteristics by fusing behavioral characteristics and semantic characteristics, a step of distinguishing sections with different semantic importance within the text and assigning differential weights to behavioral characteristics.

[0227] The above step is configured to utilize semantic importance information of the text analyzed by a large language model, so that behavioral characteristics corresponding to psychologically significant sections of the user's writing behavior are reflected relatively more significantly in the integrated analysis process.

[0228] Specifically, in the semantic feature extraction step (c) above, the large language model breaks down the text written by the user into tokens or sentences, and calculates semantic importance or attention weight for each text segment.

[0229] In this case, the semantic importance information can be expressed as weight values ​​generated from the embedding layer or attention layer of a large language model, and quantitatively represents the emotional and semantic influence that the relevant text segment occupies within the overall context. Additionally, the semantic importance information may be generated through various natural language processing techniques, such as token contribution, sentence importance, and keyword extraction algorithms.

[0230] In the above fusion step (d), semantic importance information provided by the semantic and sentiment feature extraction unit (140) and behavioral features including input delay, modification frequency, and input pattern provided by the behavioral feature extraction unit (130) are temporally mapped.

[0231] In other words, if a specific text segment is assigned high semantic importance, relatively high weights are assigned to behavioral characteristics that occurred at the time the segment was input or modified, and these are reflected in the calculation of integrated psychological characteristics.

[0232] The above weights can be dynamically calculated by considering not only the emotional polarity of the text but also structural importance such as syntactic position within the sentence, the presence of self-referential expressions, and the inclusion of keywords.

[0233] Accordingly, the process of calculating integrated psychological characteristics can be structured so that behavioral changes arising from contextually key statements, rather than simple emotional words, are preferentially reflected.

[0234] For example, if behavioral characteristics such as increased input pause time, repeated editing, or deletion are observed in text sections containing emotional expressions, key statements, or self-referential sentences of high semantic importance, said behavioral characteristics may be reflected with greater weight in the process of calculating integrated psychological characteristics than behavioral characteristics in general narrative sections of low semantic importance.

[0235] Conversely, behavioral features occurring in text segments with low semantic importance are treated with relatively low weights, which can adjust their influence on the overall risk indicator calculation process.

[0236] The integrated psychological characteristics produced through the above weighted fusion are interpreted by comparing them with the user-specific normal patterns stored in the individual baseline management unit (160).

[0237] In other words, if behavioral characteristics in text segments of high semantic importance show a significant deviation from the user's personal baseline, this can be detected as a risk signal attributable to changes in psychological state rather than changes in general writing habits.

[0238] For example, if a user includes self-referential expressions such as "I find it overwhelming these days" along with everyday descriptive sentences during the writing process, a large language model can assign relatively high semantic importance to that section.

[0239] In this case, if a longer-than-usual input pause time or repetitive sentence modification is observed before or after the input of the aforementioned section, the behavioral characteristics are given a high weight and reflected in the calculation of integrated psychological characteristics.

[0240] As a result, the same level of behavioral change has a greater impact on mental health risk indicators than when it occurs in a segment that is semantically less important.

[0241] In this way, the present invention enables advanced multimodal analysis that organically combines text semantics and writing behavior by directly linking the internal importance detection results of a large language model to the behavioral data processing stage.

[0243] A method for calculating a mental health risk indicator according to one embodiment of the present invention further includes, in step (f), that is, the step of calculating and providing a risk indicator based on detected abnormal signs, the step of visually highlighting and providing the analysis element that served as the basis for calculating the risk indicator.

[0244] The above step is performed by an indicator output unit (180) and is configured to provide the user or manager with an element that substantially contributed to the calculation of the risk indicator among the behavioral characteristics and text semantic characteristics extracted during the user's writing process, so that the user or manager can intuitively recognize it.

[0245] Specifically, in step (f), the results of the behavior-meaning mismatch analysis module (171), the word-behavior association analysis engine (153), and the attention-based weighted analysis step (d) are combined to identify key analysis elements that contributed to the calculation of the risk indicator.

[0246] At this time, the indicator output unit (180) may further include logic for extracting inference grounds for normalizing the contribution or weight values ​​calculated from each analysis module and selecting the top N key elements with relatively high contributions to the risk indicator based on the result.

[0247] The above key element may include at least one of the following.

[0248] - Specific text sections evaluated as having high semantic importance

[0249] - Location where specific sentiment keywords were entered, modified, or deleted

[0250] - The point at which significant input lag, iterative modification, or deletion occurs relative to the individual baseline

[0251] The above logic for extracting inference grounds can be implemented by a rule-based method, a machine learning-based contribution analysis method, or a combination thereof.

[0252] The indicator output unit (180) is configured to visually highlight and display the data basis upon which the risk indicator was calculated by linking the key elements selected through the inference basis extraction logic with the numerical result of the risk indicator.

[0253] The above visual emphasis may be provided in at least one of text highlighting, icon display, color change, marking on a time axis, or point highlighting on a graph, and is not limited to a specific user interface implementation.

[0254] For example, if input delay occurs intensively in a specific sentence segment of high semantic importance, that sentence segment is highlighted, and the associated input delay pattern can be displayed together on the time axis.

[0255] The above visual emphasis information can also be utilized when an alarm notification is provided through the alarm unit (190).

[0256] That is, if an alarm is triggered because a risk indicator exceeds a user-customized threshold or the frequency of risk signals increases rapidly, the alarm notification may include summary information regarding the key text segment or behavioral pattern that served as the basis for calculating the risk indicator.

[0257] Accordingly, users or administrators can understand the reason for the alarm based on specific data grounds rather than simple warning messages, and can reduce unnecessary anxiety or alarm fatigue caused by false alarms.

[0258] For example, if a user writes a sentence containing the expression "it is okay," but a significant increase in input pause time and repetitive modifications relative to the individual baseline are observed before and after the input of that sentence, the sentence segment and the corresponding input delay pattern can be identified as key supporting elements in the risk indicator calculation stage.

[0259] In this case, the indicator output unit (180) can provide the user or manager with the cause of the risk indicator being calculated by visually highlighting the sentence section and displaying the input delay and correction pattern observed in the section together.

[0260] In this way, the present invention improves the reliability of analysis results and realizes explainable AI-based monitoring by transparently presenting not only the results of mental health risk indicators but also the calculation process and basis thereof.

[0262] A method for calculating a mental health risk indicator according to one embodiment of the present invention further includes, in step (f), that is, the step of calculating and providing a risk indicator, the step of transmitting a real-time alarm notification based on the change pattern of the calculated risk indicator and the frequency of occurrence of the risk signal.

[0263] The above steps are performed in conjunction with the indicator output unit (180) and the alarm unit (190), and are configured so that a user or manager can recognize and respond to a dangerous condition at an early stage.

[0264] Specifically, in step (f) above, an alarm notification may be triggered if at least one of the following conditions is satisfied.

[0265] - If the calculated mental health risk indicator exceeds the threshold set for the user

[0266] - When the upward trend of the risk indicator satisfies a predefined pattern condition

[0267] - Cases where the frequency of danger signals increases rapidly due to the analysis of behavioral-semantic discrepancies or high-risk behavioral sequences occurring repeatedly within a unit of time

[0268] At this time, the user-customized threshold is set based on the user's normal risk indicator fluctuation range stored in the personal baseline management unit (160), and may be a threshold that is dynamically adjusted according to the user's long-term state change.

[0269] In addition, the frequency of risk signal occurrences is aggregated chronologically over multiple writing sessions rather than a single session, and can be determined based on the number of occurrences or the rate of increase within a preset time window.

[0270] The above alarm trigger conditions can be implemented by fixed rule-based, statistical trend analysis-based, or machine learning-based prediction models, and are not limited to a specific judgment method.

[0271] If any of the above conditions are satisfied, the alarm unit (190) generates a real-time alarm notification and transmits it to a user terminal or an administrator system.

[0272] The above alarm notification may be provided in at least one of push notifications, messages, dashboard displays, or integration with external management systems, and is not limited to the implementation method.

[0273] In addition, the above alarm notification is linked with visual explanatory information,

[0274] - Key writing sections that contributed to the rise in risk indicators,

[0275] - Repeatedly observed input delay or modification patterns,

[0276] - Summary information on the increased frequency of danger signals

[0277] It can be provided together.

[0278] For example, if a specific user's mental health risk indicator did not exceed a threshold in a single session, but the frequency of risk signals increased rapidly due to the repeated occurrence of the same type of behavioral or semantic mismatch signals within a short period, step (f) can determine this as a potential risk state and send a real-time alert notification.

[0279] In this case, the alarm notification provides the recently occurring recurring input delay pattern along with the text section where the pattern was observed, allowing the user or administrator to intuitively understand the cause of the dangerous condition.

[0280] In this way, the present invention enables real-time alert transmission that considers the trend of change in risk indicators and the surge in the frequency of risk signal occurrences together, without relying on whether a simple threshold is exceeded, thereby distinguishing between temporary fluctuations and continuous risk conditions and supporting early response.

[0281] Accordingly, it provides a technical effect that enables more sensitive and stable detection of mental health risks, minimizes false alarms, and allows for practical preventive intervention.

[0283] The present invention can calculate changes in mental health-related risks relative to an individual baseline, reflecting individual writing habits, in the form of continuous indicators by fusing and analyzing behavioral data and semantic information of text generated during a user's writing process using artificial intelligence. Accordingly, potential changes in psychological state and signs of risk that are difficult to capture through surveys or single text analysis alone can be detected more precisely and non-invasively.

[0284] In addition, the present invention comprehensively considers the discrepancy between behavior and meaning, the relationship between input delay and emotional vocabulary, and the changing trends and frequency of occurrence of risk indicators, thereby enabling the visualization of risk states and the provision of real-time alerts, so that users or managers can intuitively recognize changes in state and respond early.

[0285] Meanwhile, the present invention can simultaneously secure personal information protection and computational efficiency through an edge-cloud collaboration structure, and can improve the reliability and usability of the system by providing explainable analysis results.

[0286] Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will understand that the invention can be implemented in various forms without changing the technical concept or essential components of the invention. Accordingly, the above embodiments are merely illustrative and should not be interpreted as limiting the scope of the invention.

[0287] The scope of the rights of the present invention is determined by the claims set forth below, rather than by the detailed description above, and all modifications or variations of embodiments falling within the wording of the claims, their equivalents, and equivalent concepts should be interpreted as being included within the scope of the rights of the present invention. Explanation of the symbols

[0289] 100: Mental Health Risk Indicator Calculation System 110: User terminal 120: Event Collection Section 121: Key input event collection module 122: Edit / Delete Event Collection Module 123: Input Stop (Idle) Event Collection Module 130: Behavioral Feature Extraction Unit 131: Input Speed ​​Calculation Module 132: Modification Frequency Calculation Module 133: Input Pattern Analysis Module 140: Semantic Feature Extraction Unit 141: Large Language Model-based Semantic Analysis Module 142: Emotion Feature Extraction Module 150: Fusion Analysis Department 151: Session Synchronization Module 152: Integrated Feature Vector Generation Module 153: Word-Behavior Association Analysis Engine 160: Personal Baseline Management Department 161: Adaptive Baseline Learning Module 162: Personal Baseline Update Module 170: Anomaly Detection Unit 171: Behavioral-Semantic Discrepancy Analysis Module 172: Anomaly Detection Module 180: Indicator Output Section 181: Mental Health Risk Indicator Calculation Module 182: Risk Indicator Visualization Module 190: Alarm Unit 191: Threshold detection module 192: Alarm Notification Provision Module 200: Data storage unit 500: Smartphone screen with a mental health monitoring application running 510: An area that visualizes the current mental health risk level using gauges and numbers. 520: Area showing a graph of the trend of weekly risk indicator changes compared to an individual baseline. 530: An area displaying a text list of specific key detection factors that increased the risk index. 540: Text area containing alert messages regarding detected risk patterns and recommendations for expert consultation.

Claims

Claim 1 A system for calculating risk indicators related to mental health status by fusing behavioral data and semantic information of text generated during a user's writing process, comprising: an event collection unit that collects key input, editing, and input pause (Idle) events input from a user terminal along with time information; a behavioral feature extraction unit that extracts behavioral features including writing speed, revision frequency, and input patterns based on the collected events; a semantic and emotional feature extraction unit that extracts semantic and emotional features including contextual semantic features and emotional features based on text expressions obtained during the calculation process of the contextual semantic features from text written by the user using a large language model; a fusion analysis unit that generates an integrated feature vector by combining the behavioral features and the semantic and emotional features; an anomaly detection unit that detects anomalies by comparing the integrated feature vector with a personal baseline, which is a reference pattern learned for each user, wherein the anomaly detection unit detects a discrepancy between the surface emotional state indicated by the semantic and emotional features and the input delay or instability of the input pattern indicated by the behavioral features, and determines mental health risk indicators by comparing the detected discrepancy with the personal baseline; and an indicator that calculates a mental health risk indicator based on the anomaly detection result and visualizes and provides the same. A writing behavior-based personalized mental health continuous indicator calculation system characterized by including an output unit. Claim 2 A writing behavior-based personalized mental health continuous indicator calculation system, characterized in that, in claim 1, the fusion analysis unit includes a word-behavior association analysis engine that synchronizes the input time of a specific semantic word within the text with the time information of the behavior data, and identifies patterns suggesting psychological repression or self-censorship behavior by associatively analyzing delay times or deletion events before and after the input of a specific emotional keyword. Claim 3 delete Claim 4 A writing behavior-based personalized mental health continuous indicator calculation system, wherein, in claim 1, the system further comprises a personal baseline management unit that forms and manages the personal baseline based on the user's accumulated writing data, and the personal baseline management unit comprises an adaptive baseline learning module that learns changes in the user's input habits over time and updates the personal baseline periodically or when preset conditions are met. Claim 5 A writing behavior-based personalized mental health continuous indicator calculation system according to claim 1, characterized in that the event collection unit and the behavior feature extraction unit operate in an edge area within a user terminal to generate a de-identified feature vector and do not transmit raw key input data externally, and the fusion analysis unit receives the feature vector and performs analysis, having an edge-cloud collaboration structure. Claim 6 A writing behavior-based personalized mental health continuous indicator calculation system, characterized in that, in claim 1, the system further includes an alarm unit that provides an alarm notification to a user or administrator when the calculated mental health risk indicator exceeds a pre-set user-customized threshold or when the change pattern of the risk indicator satisfies a pre-set condition. Claim 7 A method for calculating a mental health risk indicator based on writing behavior by fusing user's writing behavior and semantic information, comprising: (a) collecting input events occurring from an input interface of a user terminal along with timestamps; (b) quantifying behavioral features including input delay and modification patterns from the input events; (c) using a large language model to extract semantic and emotional features including contextual semantic features of text and emotional features based on text expressions obtained during the process of calculating the contextual semantic features; (d) fusing the behavioral features and the semantic and emotional features in a multimodal manner to derive an integrated feature vector representing integrated psychological characteristics, wherein the behavioral sequence in which specific emotional vocabulary is input or deleted immediately after an input pause time exceeding a threshold occurs is analyzed to identify potential emotional instability signals not explicitly revealed in the text; (e) detecting abnormal signs by comparing the integrated psychological characteristics with a personal baseline set as a unique pattern for each user; and (f) calculating a risk indicator based on the detected abnormal signs and providing it by visualization. Claim 8 delete Claim 9 A method for calculating a personalized continuous mental health index based on writing behavior, characterized in that, in claim 7, the above step (d) includes the step of generating the integrated psychological characteristic by utilizing semantic importance information of the large language model to assign weights to behavioral characteristics corresponding to text segments of high importance. Claim 10 A method for calculating a writing behavior-based personalized mental health continuous indicator, characterized in that, in claim 7, the above step (f) includes the step of visually highlighting a specific keyword range or input delay pattern that served as the basis for calculating the risk indicator and providing it to a user or manager. Claim 11 A method for calculating a writing behavior-based personalized mental health continuous indicator, characterized in that, in claim 7, the above step (f) includes the step of transmitting a real-time alarm notification when the change in the risk indicator exceeds a pre-set user-customized threshold or when the frequency of occurrence of the risk signal increases rapidly.

Images