High-precision addiction assessment method, device, and storage medium

Abstract

This invention discloses a high-precision addiction assessment method, device, and storage medium. The method acquires multimodal addiction-related data of the assessed subject; performs time synchronization, outlier processing, and missing value processing on the data; extracts basic addiction characteristic indicators and performs standardization to generate an addiction characteristic sequence; constructs an individualized baseline model based on the addiction characteristic sequence and generates an individualized baseline deviation score; generates credibility weights based on data quality factors, subjective-objective consistency factors, intermodal collaborative triggering factors, and historical prediction contribution factors; concatenates the weighted addiction characteristics with the individualized baseline deviation score to generate a fused addiction risk feature vector; generates an addiction severity score and relapse risk prediction results based on the fused addiction risk feature vector, and dynamically calibrates based on follow-up results. This invention can improve the accuracy, objectivity, and individual adaptability of addiction assessment.

Description

Technical Field

[0001] This invention relates to the field of mental health research technology, specifically to a high-precision addiction assessment method, device, and storage medium. Background Technology

[0002] Addictive behaviors include substance addiction and behavioral addiction, typically manifested as an intense craving for a specific substance or behavior, loss of control over the use or participation in that behavior, withdrawal symptoms after reduction or cessation, and the resulting impairment of physical and mental health, learning, work, and social functioning. Addiction is characterized by its chronicity, relapse, and significant individual variability. Accurately assessing the severity of addiction and the risk of relapse is crucial for subsequent intervention, follow-up management, and relapse early warning.

[0003] Current addiction assessment methods primarily rely on questionnaires, interviews, clinical experience, or single behavioral records. While these methods can reflect the subjective cravings, frequency of use, and withdrawal status of the assessed individuals to some extent, they still have significant limitations in practical application. On the one hand, assessed individuals may conceal, underestimate, or exaggerate addictive behaviors due to shame, defensiveness, cognitive biases, or self-interest, leading to discrepancies between subjective assessment results and the true state. On the other hand, single behavioral data or single physiological indicators are insufficient to fully reflect the addictive state, especially in simultaneously representing multidimensional information such as use intensity, craving impulses, withdrawal reactions, control failure, negative consequences, environmental triggers, and changes in cognitive control ability.

[0004] Furthermore, existing addiction assessment methods typically employ fixed thresholds, fixed weights, or static models, lacking dynamic baseline modeling mechanisms tailored to individual stable states. Different subjects vary significantly in their daily routines, mobile device usage, physiological indicators, psychological states, environmental exposures, and addiction types. Using uniform thresholds can easily lead to misjudgments. This is particularly true in relapse risk prediction scenarios, where relapse is often not caused by a single abnormality but by a combination of factors, including decreased sleep, increased cravings, proximity to high-risk environments, increased failure to control behaviors, physiological fluctuations, and historical relapse patterns. Existing methods struggle to continuously calibrate assessment models based on individual baseline deviations, modal data reliability, and follow-up outcomes, resulting in untimely identification of relapse warning signs and insufficient adaptability to different individuals and addiction types.

[0005] Therefore, it is necessary to provide a high-precision addiction assessment method and device that can integrate multimodal addiction-related data, establish individualized baseline models, dynamically adjust the credibility weights of each modality's data, and continuously calibrate the risk prediction model based on follow-up results, so as to improve the objectivity, accuracy, individual adaptability, and relapse risk prediction capabilities of addiction assessment. Summary of the Invention

[0006] The purpose of this invention is to provide a high-precision addiction assessment method, device, and storage medium, which solves the problem that existing addiction assessment methods rely too much on subjective questionnaires, single data, and fixed thresholds, making it difficult to accurately and dynamically assess the severity of addiction and the risk of relapse under conditions of multimodal data noise, conflict between subjective and objective information, and significant individual differences.

[0007] The first aspect of this invention provides a high-precision addiction assessment method, comprising:

[0008] Obtain multimodal addiction-related data of the evaluated subjects, wherein the multimodal addiction-related data includes at least two of the following: subjective assessment data, mobile terminal behavior data, wearable physiological data, environmental induced data, cognitive test data, and historical outcome data; The multimodal addiction-related data is processed for time synchronization, outlier handling, and missing value handling, and basic addiction characteristic indicators are extracted from the processed multimodal addiction-related data. The basic addiction characteristic indicators are standardized to generate an addiction characteristic sequence; Based on the addiction feature sequence, stable period identification, diurnal rhythm decomposition, and abnormal window removal are performed to construct an individualized baseline model for the evaluated object and generate an individualized baseline deviation score. The credibility weights of each modality data are generated based on the multidimensional credibility assessment index of the data. The multidimensional credibility assessment index includes at least two of the following: data quality factor, subjective and objective consistency factor, intermodal collaboration triggering factor, and historical prediction contribution factor. The addiction features corresponding to each modality are weighted based on the credibility weights, and the weighted addiction features are concatenated with the individualized baseline deviation score to generate a fused addiction risk feature vector. An addiction severity score is generated based on the fused addiction risk feature vector; The fused addiction risk feature vector, feature change trend, addiction severity score, individualized baseline deviation score, and risk confidence score are input into the risk prediction model to generate relapse risk prediction results. An interpretable addiction assessment report is generated based on the addiction severity score, the relapse risk prediction results, and the risk confidence level. The individualized baseline model, the confidence weights, and the risk prediction model are dynamically calibrated based on the follow-up results.

[0009] In a second aspect, the present invention provides a high-precision addiction assessment device, comprising: A multimodal data acquisition module is used to collect multimodal addiction-related data of the assessed subject. The multimodal addiction-related data includes at least two of the following: subjective assessment data, mobile terminal behavior data, wearable physiological data, environmental induced data, cognitive test data, and historical outcome data. The data preprocessing module is used to perform time synchronization processing, outlier handling, and missing value handling on the multimodal addiction-related data; The addiction feature extraction module is used to extract basic addiction feature indicators from the processed multimodal addiction-related data, standardize the basic addiction feature indicators, and generate an addiction feature sequence. The individualized baseline modeling module is used to identify the stable period, decompose the diurnal rhythm, and remove abnormal windows based on the addiction feature sequence, and construct an individualized baseline model. The baseline deviation scoring module is used to generate an individualized baseline deviation score based on the individualized baseline model. The multimodal credibility fusion module is used to generate credibility weights for each modality based on the multidimensional credibility assessment index of the data, and to generate a fused addiction risk feature vector based on the credibility weights and the individualized baseline deviation score. An addiction severity assessment module is used to generate an addiction severity score based on the fused addiction risk feature vector; The relapse risk prediction module is used to generate relapse risk prediction results based on the fused addiction risk feature vector, feature change trend, addiction severity score, individualized baseline deviation score and risk confidence level; An interpretable report generation module is used to generate an interpretable addiction assessment report based on the addiction severity score, the relapse risk prediction result, and the risk confidence level. The feedback calibration module is used to dynamically calibrate the individualized baseline modeling module, the multimodal reliability fusion module, and the relapse risk prediction module based on the follow-up results.

[0010] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned high-precision addiction assessment method. Compared with the prior art, the present invention has at least the following beneficial effects.

[0011] First, this invention acquires at least two types of data from the following sources: subjective assessment data, mobile terminal behavior data, wearable physiological data, environmentally induced data, cognitive test data, and historical outcome data of the evaluated subject. It then performs time synchronization processing, outlier handling, missing value processing, basic addiction feature extraction, and standardization on the aforementioned multimodal addiction-related data to form an addiction feature sequence. Furthermore, this invention generates credibility weights for each modality of data based on data quality factors, subjective-objective consistency factors, intermodal collaborative triggering factors, and historical prediction contribution factors. This reduces the impact of unreliable data on the assessment results, improving the objectivity and accuracy of addiction assessment results, even in situations where subjective and objective data conflict, some modal data is missing, or noise is high.

[0012] Second, this invention uses addiction feature sequences to identify stable periods, decompose circadian rhythms, and remove abnormal windows to construct an individualized baseline model for the assessed subject and generate an individualized baseline deviation score. Furthermore, the weighted multimodal addiction features are concatenated with the individualized baseline deviation score to generate a fused addiction risk feature vector. Based on this fused addiction risk feature vector, an addiction severity score and relapse risk prediction results are generated. By dynamically calibrating the individualized baseline model, confidence weights, and risk prediction model based on follow-up results, this invention can continuously adapt to individual changes in the assessed subject, improving the ability to identify changes in addiction status and relapse risk.

[0013] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the specific embodiments will be briefly described below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort.

[0015] Figure 1 A flowchart illustrating the high-precision addiction assessment method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the structure of the high-precision addiction assessment system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of multimodal data preprocessing and addiction feature extraction provided in an embodiment of the present invention; Figure 4This is a schematic diagram illustrating the construction of an individualized baseline model and the generation of baseline deviation scores provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the multimodal credibility dynamic fusion, addiction scoring, and relapse risk prediction provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] This invention provides a high-precision addiction assessment method and system based on multimodal behavioral representation and dynamic risk calibration, applicable to risk assessment scenarios of drug addiction, alcohol addiction, nicotine addiction, internet gaming addiction, short video addiction, social media addiction, gambling addiction, and other behavioral addictions or substance dependence.

[0018] like Figure 1 As shown, the high-precision addiction assessment method 100 provided in this embodiment of the invention includes: acquiring multimodal addiction-related data of the assessed subject S1; preprocessing the multimodal addiction-related data and extracting addiction characteristic indicators S2; constructing an individualized baseline model based on stable period identification, rhythm decomposition, and anomaly removal S3; dynamically fusing multimodal credibility based on data quality, modal consistency, and historical prediction contribution S4; generating an addiction severity score S5; generating relapse risk prediction results S6; generating an interpretable addiction assessment report S7; and performing dynamic calibration based on follow-up results S8.

[0019] The core concept of this invention is as follows: First, the multimodal addiction-related data of the assessed subjects are preprocessed and features are extracted to form a computable addiction feature sequence; then, based on the addiction feature sequence, a stable period window is selected, an individualized baseline model is constructed, and a baseline deviation score is generated; further, based on data quality, consistency between subjective and objective factors, intermodal synergistic triggering relationships, and historical prediction contributions, the credibility of the multimodal addiction features is dynamically fused to generate an addiction severity score, relapse risk prediction results, risk confidence level, and an interpretable assessment report, and dynamic calibration is performed based on follow-up results.

[0020] S1 Obtain multimodal addiction-related data of the evaluated subjects. like Figure 1 and Figure 3As shown, in step S1, multimodal addiction-related data 301 of the evaluated subject is acquired within a preset time range. The multimodal addiction-related data 301 includes at least two of the following: subjective assessment data 3011, mobile terminal behavior data 3012, wearable physiological data 3013, environmental induced data 3014, cognitive test data 3015, and historical outcome data 3016.

[0021] Among them, subjective assessment data 3011 includes one or more of the following: addiction craving score, withdrawal reaction score, impulse control score, emotional state score, stress level score, self-control ability score, self-reported data on usage frequency, self-reported data on usage scenarios, and withdrawal intention score.

[0022] Mobile terminal behavior data 3012 includes one or more of the following: application usage time, frequency of launching addiction-related applications, continuous usage time, proportion of nighttime use, application switching frequency, search behavior, records of enabling or disabling active restriction functions, payment or consumption behavior, and changes in call or social frequency.

[0023] Wearable physiological data 3013 includes one or more of the following: heart rate, heart rate variability, sleep duration, sleep quality, activity level, skin conductance, body temperature changes, exercise intensity, and resting state changes.

[0024] Environmental induced data 3014 includes one or more of the following: geographic location data, high-risk location proximity records, duration of stay at high-risk locations, high-risk contact contact records, environmental noise, time and scene, holiday information, and proximity information of past relapse scenarios.

[0025] Cognitive test data 3015 includes one or more of the following: reaction time, delayed gratification test results, inhibitory control test results, attentional bias test results, impulsive choice test results, and task completion stability data.

[0026] Historical outcome data 3016 includes one or more of the following: type of addiction, duration of addiction, frequency of use, relapse record, withdrawal record, treatment record, drug intervention record, psychological intervention record, manual review results, intervention response results, and family or social support status.

[0027] S2 preprocesses multimodal addiction-related data and extracts addiction feature indicators. like Figure 3 As shown, step S2 includes time synchronization processing 302, outlier processing 303, missing value processing 304, basic addiction feature index extraction 305, standardization processing 306, and addiction feature sequence generation 307.

[0028] It should be noted that in this embodiment, time synchronization processing 302, outlier processing 303, and missing value processing 304 are applied to multimodal addiction-related data 301 to obtain preprocessed data suitable for feature extraction; basic addiction feature index extraction 305 is applied to the preprocessed data to generate basic addiction feature indices; standardization processing 306 is applied to the basic addiction feature indices 305 to eliminate dimensional differences between different features; and addiction feature sequence generation 307 is applied to the feature values ​​after standardization processing 306 to form the addiction feature sequence required for subsequent individualized baseline modeling and risk prediction. Therefore, Figure 3 The process shown is as follows: Multimodal addiction-related data 301 undergoes time synchronization processing 302, outlier processing 303, and missing value processing 304 in sequence, and then enters the basic addiction feature index extraction 305; the extracted basic addiction feature index 305 is then standardized 306, and finally an addiction feature sequence 307 is generated.

[0029] S201, Time synchronization processing is performed on multimodal addiction-related data 301 302.

[0030] Data from different sources are aligned along a unified timeline to obtain multimodal time series data represented by a preset sampling time granularity. The preset sampling time granularity can be 1 minute, 5 minutes, 15 minutes, 1 hour, or 1 day.

[0031] For data with a sampling frequency higher than the preset sampling time granularity, resampling is performed using the mean, median, maximum, or cumulative value; for data with a sampling frequency lower than the preset sampling time granularity, forward padding, linear interpolation, or missing data markers are used for processing.

[0032] S202, perform outlier processing on multimodal addiction-related data 301 303.

[0033] For each type of data, set valid value ranges and change rate limits. When a data point exceeds the measurable range of the device, exceeds the reasonable range of human physiology, the positioning point drifts significantly, multiple consecutive sampling points are identical and exceed the preset time, or the change rate compared with the previous and subsequent sampling points exceeds the preset change rate threshold, the data point is marked as abnormal data.

[0034] For abnormal data, the system performs processing such as removal, interpolation, retention of missing markers, or reduction of data quality factor according to the type of abnormality.

[0035] S203, perform missing value processing on multimodal addiction-related data 301 304.

[0036] The proportion of missing data for each modality within each time window is calculated. When the missing proportion is below the first missing threshold, interpolation or neighboring window statistics are used to complete the data. When the missing proportion is above the first missing threshold but below the second missing threshold, the window is retained but the data quality factor of the corresponding modality is reduced. When the missing proportion is above the second missing threshold, the data for that modality within that time window is marked as unavailable.

[0037] The first missing threshold can be set to 20%, and the second missing threshold can be set to 60%. These thresholds can be adjusted according to the application scenario.

[0038] S204, Extract basic addiction characteristic indicators 305 from the preprocessed multimodal addiction-related data 301.

[0039] The basic addiction characteristic index 305 includes one or more of the following: use intensity characteristic 3051, craving impulse characteristic 3052, withdrawal reaction characteristic 3053, control failure characteristic 3054, negative consequences characteristic 3055, environmental inducement characteristic 3056, and cognitive control characteristic 3057.

[0040] Among them, the intensity of use characteristic 3051 includes one or more of the following: number of uses, duration of use, dosage, amount, continuous use time, degree of shortening of use interval, and proportion of nighttime use.

[0041] The craving impulse characteristic 3052 includes subjective craving rating, impulsive reaction time, degree of decline in delayed gratification ability, repeated opening of related applications in a short period of time, searching for related content or approaching related locations.

[0042] Withdrawal symptoms 3053 include anxiety, irritability, insomnia, increased heart rate, decreased activity, mood swings, and decreased attention after stopping or reducing use.

[0043] The failure control feature 3054 includes features such as multiple attempts to reduce but failing, exceeding the preset usage time, failure to execute the withdrawal plan, frequent removal of the active restriction function, and ignoring intervention reminders.

[0044] Negative consequences characteristic 3055 includes features such as decreased learning efficiency, decreased work efficiency, social withdrawal, impaired sleep, abnormal consumption, deterioration of health indicators, and family relationship conflicts.

[0045] Environmental triggering characteristics 3056 include proximity to high-risk locations, contact with high-risk contacts, being in high-stress periods, entering scenarios with a history of relapse or high-risk scenarios during holidays.

[0046] Cognitive control characteristics 3057 include one or more of the following: prolonged reaction time, decreased inhibitory control ability, decreased delayed gratification ability, increased attentional bias, and increased impulse selection ratio.

[0047] S205, standardize the basic addiction characteristic indicators 305 306.

[0048] To enable unified calculation of features with different dimensions, the basic addiction feature index 305 is standardized 306.

[0049] In one implementation, for the k-th feature in the m-th mode, the original feature value within the time window t is denoted as... The center value of the corresponding feature is calculated using historical stable samples or population samples. and scale value Then the standardized eigenvalues Calculate using the following formula:

[0050] in, To prevent smoothing constants with a denominator of zero.

[0051] In another implementation, robust standardization is performed using the median and interquartile range to reduce the impact of extreme values.

[0052] in, Let be the median of the k-th feature in the m-th mode. is the interquartile range of the k-th feature in the m-th mode.

[0053] S206, generating addiction feature sequence 307.

[0054] The standardized eigenvalues ​​(after standardization by 306) within the same time window are concatenated according to modality category and feature category to form an addiction feature vector. :

[0055] Arranging the addiction feature vectors corresponding to each time window in chronological order yields addiction feature sequence 307, i.e. This addiction feature sequence 307 serves as input for subsequent individualized baseline modeling, dynamic credibility fusion, addiction severity scoring, and relapse risk prediction.

[0056] S3 constructs an individualized baseline model based on stable period identification, rhythm decomposition, and anomaly removal. Based on the addiction feature sequence 307 generated in step S2, an individualized baseline model of the evaluated subject is constructed. For example... Figure 4 As shown, step S3 includes steps such as candidate time window division 401, state fluctuation index calculation 402, stable period window screening and abnormal window removal 403, diurnal rhythm decomposition 404, multi-level individualized baseline vector construction 405, multi-dimensional deviation calculation 406, and individualized baseline deviation score generation 407.

[0057] S401, divide the candidate time window.

[0058] The preset observation period is divided into multiple candidate time windows. The candidate time windows can be hourly windows, half-day windows, daily windows, or weekly windows, or they can be divided according to time periods such as early morning, daytime, evening, nighttime, and late night.

[0059] S402, calculate the state fluctuation index of the candidate time window.

[0060] For each candidate time window, a state fluctuation index is calculated based on the addiction feature sequence 307 obtained in step S2.

[0061] For the k-th feature within the candidate time window W, its fluctuation value Calculate using the following formula:

[0062] in, This represents the feature value of the k-th feature within the time window t. This represents the average value of the k-th feature within the candidate time window W. This represents the number of sampling points within the candidate time window W.

[0063] In another implementation, to improve robustness to outliers, the state fluctuation index 402 can be calculated using the mean absolute deviation:

[0064] in, This represents the median of the k-th feature within the candidate time window W.

[0065] Based on the fluctuation values ​​of each feature, a comprehensive state fluctuation value for the candidate time window W is generated. :

[0066] in, Contribute weights to the fluctuations corresponding to the k-th feature, satisfying .

[0067] S403, filter stable windows and remove abnormal windows.

[0068] Based on the comprehensive state fluctuation value, objective behavioral anomalies, data missing ratio, and historical outcome markers, stable period windows are selected from multiple candidate time windows, and abnormal windows are removed.

[0069] A candidate time window is designated as a stable period window when it meets at least some of the following conditions: the intensity of addiction-related behaviors does not exceed the preset fluctuation range; sleep indicators do not show a continuous abnormal decline; physiological indicators fluctuate within the preset range; the frequency of exposure to high-risk environments is lower than the preset threshold; the proportion of missing data is lower than the preset proportion; and the overall state fluctuation value is lower than the stable period fluctuation threshold.

[0070] A candidate time window is identified as an abnormal window and removed if any of the following conditions exist: confirmed or suspected relapse behavior occurs; addiction-related behaviors suddenly increase in a short period of time and exceed a preset intensity threshold; sleep duration or sleep quality continuously and abnormally declines; continuous approach to or prolonged stay in high-risk locations; the proportion of missing wearable devices, mobile terminals, or location data exceeds a preset threshold; or there are obvious errors in the collection of objective behavioral data.

[0071] It should be noted that when there is a conflict between subjective evaluation data and objective behavioral data, it is not directly used as an anomaly window removal condition, but is used as the subjective-objective consistency factor 502 in the subsequent step S4 to participate in the credibility weight calculation.

[0072] S404 performs diurnal rhythm decomposition on the stable period window.

[0073] The data within the selected stable period window are subjected to diurnal rhythm decomposition to obtain the stable state characteristics corresponding to different time periods.

[0074] In one implementation, the day is divided into multiple preset circadian rhythm periods. Different circadian rhythm periods include one or more of the following: early morning, morning, afternoon, evening, and sleep periods.

[0075] For the One characteristic during rhythmic periods The steady-state values ​​within the range are calculated as follows:

[0076] in: Indicates the first One characteristic during rhythmic periods The rhythm baseline value below; Indicates the first One feature in the time window Eigenvalues ​​within; Indicates a period of rhythm. The number of effective sampling windows.

[0077] In one implementation, to avoid outliers affecting the rhythm decomposition results, the data within the stable period window is preprocessed using a sliding median filter or quartile outlier removal method before calculating the rhythm baseline value.

[0078] In one implementation, for behavioral, physiological, or sleep characteristics with obvious periodicity, Fourier decomposition, wavelet decomposition, or cosine rhythm fitting methods can be used to extract diurnal periodic components, and these periodic components can be used as part of the rhythmic characteristics in subsequent baseline modeling.

[0079] S405, construct multi-layer individualized baseline vectors.

[0080] Based on the results of the stable period window and diurnal rhythm decomposition, a multi-layered individualized baseline vector is constructed for the evaluated objects.

[0081] The multi-layered individualized baseline vector includes one or more of the following: behavioral baseline vector 4051, physiological baseline vector 4052, emotional baseline vector 4053, environmental baseline vector 4054, cognitive baseline vector 4055, and executive baseline vector 4056.

[0082] In one implementation, the first The baseline vector is denoted as:

[0083] in: Indicates the first Baseline vector; Indicates the first One characteristic during rhythmic periods The rhythm baseline value below.

[0084] In one implementation, different categories of baseline vectors correspond to different feature sets: behavioral baseline vector 4051 is used to characterize mobile terminal usage behavior in a stable state; physiological baseline vector 4052 is used to characterize heart rate, heart rate variability, sleep quality, and exercise status in a stable state; emotional baseline vector 4053 is used to characterize the level of emotional fluctuation in a stable state; environmental baseline vector 4054 is used to characterize the frequency of exposure to high-risk locations in a stable state; cognitive baseline vector 4055 is used to characterize attention and impulse control ability in a stable state; and executive baseline vector 4056 is used to characterize task execution stability and plan completion ability in a stable state.

[0085] In one implementation, to prevent long-term behavioral changes from causing the baseline model to fail, various baseline vectors are dynamically updated using an exponential sliding update method:

[0086] in: Indicates the baseline update coefficient; Indicates the historical baseline value; This represents the updated baseline value.

[0087] S406, calculates the multidimensional deviation of the current state from the individualized baseline.

[0088] Within the current assessment period, the currently collected data is compared with the multi-layer individualized baseline vector for the corresponding time period to calculate the multidimensional deviation.

[0089] For the current time window Inner The deviation of each feature is calculated as follows:

[0090] in: Indicates the current time window Inner One eigenvalue; Indicates rhythmic periods Individualized baseline values ​​below; Indicates rhythmic periods The corresponding fluctuation scale; This represents the smoothing constant.

[0091] In one implementation, the fluctuation scale is calculated using the standard deviation within a stable window:

[0092] Subsequently, the deviations of features within the same category are aggregated to obtain the deviations of the corresponding dimensions:

[0093] in: Indicates the first Dimensional deviation; This indicates the dimensional contribution weight of the corresponding feature.

[0094] S407 generates an individualized baseline deviation score.

[0095] A personalized baseline deviation score is generated based on the multidimensional deviation. In one implementation, the personalized baseline deviation score 407 is calculated as follows:

[0096] in: Indicates the individualized baseline deviation score; Indicates the first The contribution weight of dimensional deviation; This represents the offset parameter.

[0097] The individualized baseline deviation score is used to characterize the degree of abnormality of the current state of the assessed object relative to its own stable state, and serves as an important input for subsequent fusion of addiction risk feature vectors and relapse risk prediction results.

[0098] S4 dynamically fuses multimodal credibility based on data quality, modal consistency, and historical prediction contributions.

[0099] Addiction characteristic indicators and individualized baseline deviation scores are dynamically fused using multimodal confidence metrics to generate a fused addiction risk feature vector and risk confidence level. For example... Figure 5 As shown, step S4 includes: S501, calculate the data quality factor for each modality of data.

[0100] Calculate the data quality factor for each type of modal data. The data quality factors include one or more of the following: data integrity, sampling continuity, noise level, device wearability effectiveness, positioning accuracy, time synchronization error, data freshness, and user authorization integrity.

[0101] Specifically, the system first performs quality checks on each type of modal data. For mobile terminal behavior data 3012, the system checks whether the application usage records are continuous, whether background data collection is interrupted, whether the timestamps are complete, and whether there are any abnormal jumps in data. For wearable physiological data 3013, the system checks the effectiveness of device wearing, the stability of sampling frequency, the sensor contact status, and the signal-noise level. For environmentally induced data 3014, the system checks the positioning accuracy, GPS drift, and trajectory continuity. For cognitive test data 3015, the system checks the test completion rate, abnormal response time, and the proportion of invalid clicks. For subjective evaluation data 3011, the system checks the completeness of questionnaire completion, abnormal completion time, and the occurrence of continuous repeated responses.

[0102] In one implementation, the system calculates multiple sub-quality indicators, including data integrity factor, sampling continuity factor, noise suppression factor, device effectiveness factor, and time synchronization factor. For example, data integrity reflects the ratio between the actual amount of data collected and the theoretically required amount of data collected within the current time window; sampling continuity reflects whether there are long-term interruptions in data acquisition; noise level reflects whether sensor data fluctuations are abnormal; device wearability determines whether the wearable device is actually being worn; and time synchronization error reflects the time alignment error between different modal data.

[0103] For example, for wearable physiological data 3013, the system can determine whether the device is worn effectively based on the continuity of the heart rate signal, skin contact impedance, and the proportion of motion artifacts. When the system detects that there is no physiological fluctuation for a long time, the heart rate signal is fixed or missing, the device effectiveness score is reduced.

[0104] For environmentally induced data 3014, the system can assess the quality of positioning data based on the GPS positioning error radius, positioning refresh frequency, and trajectory continuity. When positioning drift is severe, the trajectory is broken, or the positioning update frequency is lower than a preset threshold, the environmental data quality score is reduced.

[0105] Subsequently, the system normalizes each sub-quality indicator and performs weighted fusion according to preset weights to generate the data quality factor for the corresponding modality within the current time window. In one implementation, the data quality factor can be calculated as follows:

[0106] in: This represents the data quality factor of the m-th mode within the time window t; Indicates the data completeness factor; Indicates the sampling continuity factor; Indicates the noise suppression factor; Indicates the equipment effectiveness factor; Indicates the time synchronization factor; This represents the contribution weight of each sub-quality indicator, and the sum of the weights is 1.

[0107] This indicates that the overall quality of the modality data is high, therefore, it will receive a higher weight in the subsequent credibility fusion process.

[0108] S502, calculate the consistency factor between subjective data and objective data.

[0109] Based on the degree of consistency between subjective assessment data 3011 and mobile terminal behavior data 3012, wearable physiological data 3013, environmentally induced data 3014, and cognitive test data 3015, a subjective-objective consistency factor is calculated. The subjective-objective consistency factor is used to reflect the degree of matching between the subjective self-report and the actual behavior of the evaluated object, and serves as an important basis for subsequent credibility weight calculation.

[0110] Specifically, the system first generates a subjective risk score based on subjective assessment data 3011. The subjective assessment data 3011 includes one or more of the following: subjective craving score, anxiety score, self-control score, withdrawal distress score, self-reported usage frequency, and self-reported usage duration. After standardizing each subjective characteristic, the system performs weighted aggregation based on the contribution weight of each characteristic to generate the subjective risk score for the current time window. A higher subjective risk score indicates a higher perceived risk of addiction for the assessed individual.

[0111] Subsequently, the system generates an objective risk score based on mobile terminal behavior data 3012, wearable physiological data 3013, environmentally induced data 3014, and cognitive test data 3015. The objective risk score is calculated based on one or more of the following characteristics: actual usage time, number of times the application is opened, proportion of nighttime use, continuous usage time, degree of sleep decline, changes in heart rate variability, time spent in high-risk locations, and impulse control error rate. After standardizing the objective risk characteristics, the system performs weighted aggregation according to the contribution weight of each objective characteristic to generate the objective risk score for the current time window.

[0112] In obtaining subjective risk scores and objective risk score The system then calculates the difference between the two. If the subjective risk score and the objective risk score are close, it indicates that the subjective perception of the assessed object is highly consistent with the objective state; if the difference is large, it indicates that there is subjective concealment, subjective underestimation, or subjective exaggeration.

[0113] In one implementation, the system generates a subjective-objective consistency factor using an exponential decay method; that is, the greater the difference between subjective and objective risks, the lower the subjective-objective consistency factor; the smaller the difference between subjective and objective risks, the closer the subjective-objective consistency factor is to 1. Through this method, the system can dynamically assess the credibility of subjective data and automatically reduce the impact of low-consistency subjective data on the overall risk assessment results during subsequent multimodal credibility fusion.

[0114] For example, if the person being evaluated subjectively reports that "they hardly used short video apps that day," but the system detects that their actual short video usage time has increased significantly, their continuous nighttime usage time has increased, and their sleep quality has decreased, then their subjective risk score is significantly lower than their objective risk score. In this case, the system will reduce the subjective consistency factor and correspondingly reduce the weight of subjective data in the credibility fusion, thereby avoiding the underestimation of addiction risk due to subjective concealment.

[0115] S503, calculates the intermodal co-triggering factor.

[0116] Calculate the intermodal collaborative triggering factor based on whether different modal characteristics form a preset risk combination within the same time window. .

[0117] In one implementation, the preset risk combination includes one or more of the following: increased nighttime use and decreased sleep quality; increased time spent in high-risk locations and increased subjective craving scores; decreased heart rate variability and decreased impulse control; continuous opening of addiction-related applications and active restriction functions being turned off.

[0118] When multiple risk characteristics simultaneously meet preset triggering conditions within the same time window, the collaborative triggering factor of the corresponding modality is increased.

[0119] In one implementation, the co-triggering factor is calculated as follows:

[0120] in: This indicates the number of risk combinations that have been triggered within the current time window; This represents the synergistic enhancement coefficient.

[0121] S504, calculate the historical prediction contribution factor.

[0122] Based on the assessed subjects' historical relapse events, withdrawal failure events, withdrawal success events, effective intervention events, and manual review results, the contribution of each modal characteristic to the historical prediction results is calculated, resulting in the historical prediction contribution factor. .

[0123] The historical prediction contribution factor is used to reflect the effectiveness, stability, and discriminative ability of different modal data in historical relapse prediction, and serves as an important component of the subsequent credibility weight. The higher the historical prediction contribution factor, the more predictive value the corresponding modal feature has in historical risk prediction; the lower the historical prediction contribution factor, the weaker the contribution of the corresponding modality to risk prediction or the poorer the stability.

[0124] Specifically, the system first establishes a historical event sample set. This historical event sample includes one or more of the following: historical relapse events, failed withdrawal events, successful withdrawal events, effective intervention events, and results of manual review. For each historical event sample, the system retains multimodal feature data within a preset time window prior to the event, along with the corresponding true outcome label. The true outcome label is used to characterize whether relapse, uncontrolled use, or successful withdrawal occurred within the corresponding prediction time window.

[0125] In one implementation, the system first establishes a complete prediction model based on all modal features and then calculates the prediction loss of the complete model on historical samples. The prediction loss may be one or more of the following: cross-entropy loss, mean squared error loss, F1 loss, or AUC error.

[0126] Subsequently, the system employs a "modal ablation" approach to evaluate the contribution of each modality to the prediction results. Specifically, the system sequentially removes features of a certain modality, retains only the remaining modal features, re-performs historical predictions, and calculates the prediction loss after removing that modality. If removing a certain mode significantly increases the prediction loss, it indicates that the mode contributed significantly to the historical prediction results; if the change in prediction loss is small, it indicates that the mode contributed less.

[0127] In one implementation, the historical prediction contribution factor corresponding to the m-th mode is... Calculate as follows:

[0128] in: This represents the historical prediction contribution factor of the m-th mode within the time window t; This represents the prediction loss given complete feature input; This represents the prediction loss after removing the m-th mode; Indicates the contribution amplification factor; This represents the smoothing constant.

[0129] Furthermore, in one implementation, the system can also adjust the historical prediction contribution factor based on the results of manual review. For example, when manual review confirms that a certain mode has long-term abnormal noise, false triggering, or equipment drift problems, the historical prediction contribution factor of the corresponding mode can be reduced; when a certain mode shows abnormal changes in advance of multiple reabsorption events, the historical prediction contribution factor of the corresponding mode can be increased.

[0130] Furthermore, to prevent older data from continuously influencing the current model, the system can also perform time decay processing on historical contributions. That is, the more recent a historical event is, the higher its contribution weight; the more distant a historical event is, the lower its contribution weight gradually becomes, thereby improving the model's ability to adapt to recent behavioral changes.

[0131] S505 generates the credibility weights for each modality of data.

[0132] Based on data quality factors Consistency factor between subjective and objective factors Intermodal cooperative triggering factors Historical prediction contribution factors and time decay factor Generate the credibility weights for each modality of data.

[0133] In one implementation, the unnormalized confidence weights are first calculated:

[0134] in, This represents the initial empirical weights for the m-th mode. This represents the time decay factor.

[0135] Subsequently, the unnormalized confidence weights for each modality are normalized to obtain the final confidence weights:

[0136] The credibility weights of each modality data are adjusted according to the addiction type.

[0137] S506, generate a fusion addiction risk feature vector.

[0138] like Figure 5 As shown, the credibility weights and those from Figure 4 The individualized baseline deviation scores are used as inputs to the fused addiction risk feature vector. In other words, the fused addiction risk feature vector is not generated solely by the confidence weights, but is formed by the cascading of each modality feature adjusted for confidence weights and the individualized baseline deviation scores.

[0139] In one implementation, the modal feature vector of the m-th mode within the current time window t is denoted as Then, the addiction risk feature vector is fused. Calculate using the following formula:

[0140] in, This represents the individualized baseline deviation score, and concat represents the feature cascading operation.

[0141] S510 generates risk confidence level and insufficient data prompts.

[0142] Risk confidence is generated based on the number of data modalities involved in the fusion, the data quality factor of each modality, the subjective and objective consistency factor, the proportion of missing data, and the time synchronization error.

[0143] In one implementation, the risk confidence level is calculated as follows:

[0144] in: Indicates the level of confidence in the risk; Indicates the number of data modalities participating in the fusion; This indicates the percentage of missing data in the current time window.

[0145] In one implementation, when the risk confidence level is lower than a preset threshold, a data insufficiency warning is generated, and the output level of the relapse risk prediction result corresponding to the current time window is reduced.

[0146] S5 generates an addiction severity score. like Figure 5 As shown, addiction severity scores are generated based on fused addiction risk feature vectors.

[0147] In one implementation, an addiction severity score is used. Calculate using the following formula:

[0148] in, For bias terms, To fuse addiction risk feature vectors The corresponding rating coefficient.

[0149] S6 generates relapse risk prediction results. like Figure 5 As shown, the risk prediction result of relapse is generated based on the changing trend of the fusion addiction risk feature vector within the most recent preset time window.

[0150] This embodiment uses a time-window logistic regression risk model to predict the risk of relapse within a preset time window. Specifically, let the current time window be t, and the future prediction time window be... , The duration can be 24 hours, 72 hours, 7 days, or 30 days. The system is based on the current fusion addiction risk feature vector. The change in the fusion addiction risk feature vector over the past K time windows Addiction severity score Individualized baseline deviation score and risk confidence level Calculate the future Risk of relapse within the time window :

[0151] in, For bias terms, For model parameters, This indicates the changing trend of risk characteristics within the most recent K time windows.

[0152] In this way, the relapse risk prediction result is the prediction result output by the time window logistic regression risk model; dynamic calibration is applied to the model parameters of the time window logistic regression risk model, rather than directly to the relapse risk prediction result itself.

[0153] S7 generates an interpretable addiction assessment report. like Figure 5 As shown, an interpretable addiction assessment report is generated based on addiction severity scores, relapse risk prediction results, risk confidence levels, and feature contribution information formed during the fusion process.

[0154] The interpretable addiction assessment report includes one or more of the following: current addiction severity score, current addiction risk level, predicted relapse risk within a future preset time window, risk confidence level, main risk contributing factors, trend of change compared to historical assessment results, deviation compared to individualized baseline, insufficient data indication, suggested additional data items, suggested items for manual review, and suggested intervention level.

[0155] S8 performs dynamic calibration based on follow-up results. like Figure 1 and Figure 5 As shown, the follow-up results of the evaluated subjects are obtained after the evaluation, and dynamic calibration is performed based on the follow-up results.

[0156] Specifically, dynamic calibration applies at least to the individualized baseline model, confidence weights, and time-window logistic regression risk model. In one implementation, dynamic calibration may further apply to historical prediction contribution factors to update the contribution of each modality feature to historical prediction results based on follow-up outcomes.

[0157] The follow-up results include one or more of the following: whether relapse occurred, whether the abstinence plan was completed, whether intervention was received, whether the score decreased after intervention, manual review results, treatment records, medication use records, psychological counseling records, and user feedback results.

[0158] In one implementation, for the time-window logistic regression risk model 508 in step S6, labels are generated based on the follow-up results. When reabsorption or runaway use occurs within the predicted time window, ;otherwise The system uses the cross-entropy loss function to update the model parameters:

[0159] Once the number of new follow-up samples reaches the preset number, the system updates the model parameters in batches; or it performs small updates using incremental learning after each follow-up result is obtained.

[0160] Through the aforementioned dynamic calibration 512, the evaluation model can gradually adapt to the individual characteristics of the evaluated object, thereby improving the long-term evaluation accuracy.

[0161] System Implementation Examples like Figure 2As shown, the present invention also provides a high-precision addiction assessment system 200. The high-precision addiction assessment system 200 includes a multimodal data acquisition module 201, a data preprocessing module 202, an addiction feature extraction module 203, an individualized baseline modeling module 204, a baseline deviation scoring module 205, a multimodal credibility fusion module 206, an addiction severity assessment module 207, a relapse risk prediction module 208, an interpretable report generation module 209, a feedback calibration module 210, and a privacy and access control module 211.

[0162] The multimodal data acquisition module 201 is used to collect subjective assessment data 3011, mobile terminal behavior data 3012, wearable physiological data 3013, environmental induced data 3014, cognitive test data 3015, and historical outcome data 3016 of the evaluated object.

[0163] The data preprocessing module 202 is used to clean, remove outliers, fill in missing values, convert formats, synchronize time, filter noise and standardize the multimodal addiction-related data 301.

[0164] The addiction feature extraction module 203 is used to extract usage intensity features 3051, craving impulse features 3052, withdrawal reaction features 3053, control failure features 3054, negative consequences features 3055, environmental inducement features 3056 and cognitive control features 3057 from the preprocessed multimodal addiction-related data 301, and generate an addiction feature sequence 307.

[0165] The individualized baseline modeling module 204 is used to divide candidate time windows 401 based on the addiction feature sequence 307, calculate state fluctuation index 402, screen stable period windows, remove abnormal windows, and perform diurnal rhythm decomposition 404 on the stable period windows to construct multi-layer individualized baseline vectors 405.

[0166] The baseline deviation scoring module 205 is used to compare the data in the current assessment period with the multi-layer individualized baseline vector 405 of the corresponding time period, calculate the multidimensional deviation 406, and generate an individualized baseline deviation score 407.

[0167] The multimodal reliability fusion module 206 is used to calculate the data quality factor of each modality of data. Consistency factor between subjective and objective factors Intermodal cooperative triggering factors Historical prediction contribution factors and time decay factor And generate the credibility weights for each modality of data. Further combined with individualized baseline deviation scores Generate a fusion addiction risk feature vector and risk confidence .

[0168] The addiction severity assessment module 207 is used for assessment based on fused addiction risk feature vectors. Generate an addiction severity score And addiction risk level.

[0169] The relapse risk prediction module 208 is used to predict relapse risk based on fusion addiction risk feature vectors. Characteristics and trends Individualized baseline deviation score and risk confidence The 508 time-window logistic regression risk model generates predictions of relapse risk within a preset future time window. .

[0170] Interpretable report generation module 209 is used to generate reports that include an addiction severity score. Relapse risk prediction results Risk confidence level Explainable addiction assessment report 511, including key contributing factors, individualized baseline deviations, data insufficiency warnings, and recommended intervention levels.

[0171] The feedback calibration module 210 is used to dynamically calibrate the individualized baseline modeling module 204, the multimodal credibility fusion module 206, and the relapse risk prediction module 208 based on follow-up results, actual relapse records, completion status of the abstinence plan, intervention response status, and manual review results.

[0172] The privacy and access management module 211 is used to control access permissions, encrypt and store data, de-identify data, grant hierarchical authorization, and manage access records for the data of the evaluated object.

[0173] Any aspects not covered in this invention are applicable to existing technologies.

[0174] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A high-precision addiction assessment method, characterized in that, include: Obtain multimodal addiction-related data of the evaluated subjects, wherein the multimodal addiction-related data includes at least two of the following: subjective assessment data, mobile terminal behavior data, wearable physiological data, environmental induced data, cognitive test data, and historical outcome data; The multimodal addiction-related data is processed for time synchronization, outlier handling, and missing value handling, and basic addiction characteristic indicators are extracted from the processed multimodal addiction-related data. The basic addiction characteristic indicators are standardized to generate an addiction characteristic sequence; Based on the addiction feature sequence, stable period identification, diurnal rhythm decomposition, and abnormal window removal are performed to construct an individualized baseline model for the evaluated object and generate an individualized baseline deviation score. The credibility weights of each modality data are generated based on the multidimensional credibility assessment index of the data. The multidimensional credibility assessment index includes at least two of the following: data quality factor, subjective and objective consistency factor, intermodal collaboration triggering factor, and historical prediction contribution factor. The addiction features corresponding to each modality are weighted based on the credibility weights, and the weighted addiction features are concatenated with the individualized baseline deviation score to generate a fused addiction risk feature vector. An addiction severity score is generated based on the fused addiction risk feature vector; The fused addiction risk feature vector, feature change trend, addiction severity score, individualized baseline deviation score, and risk confidence score are input into the risk prediction model to generate relapse risk prediction results. An interpretable addiction assessment report is generated based on the addiction severity score, the relapse risk prediction results, and the risk confidence level. The individualized baseline model, the confidence weights, and the risk prediction model are dynamically calibrated based on the follow-up results.

2. The high-precision addiction assessment method according to claim 1, characterized in that, The extraction of basic addiction characteristic indicators from the processed multimodal addiction-related data includes: Extract at least one of the following characteristics: intensity of use, craving impulse, withdrawal reaction, control failure, negative consequences, environmental triggering, and cognitive control. The usage intensity characteristics include at least one of the following: number of uses, usage duration, dosage, amount, continuous usage time, degree of shortening of usage intervals, and proportion of nighttime use; The characteristics of the craving impulse include at least one of the following: subjective craving score, impulse reaction time, degree of decline in delayed gratification ability, repeated opening of relevant applications in a short period of time, searching for relevant content or approaching relevant locations; The control failure characteristics include at least one of the following: exceeding the preset usage time, failure to execute the abstinence plan, removal of the active restriction function, and ignoring the intervention reminder.

3. The high-precision addiction assessment method according to claim 1, characterized in that, The standardization process for the basic addiction characteristic indicators to generate an addiction characteristic sequence includes: For the k-th feature in the m-th mode, the original feature value within the time window t is denoted as . The center value of the corresponding feature is calculated using historical stable samples or population samples. and scale value Calculate the standardized eigenvalues ​​using the following formula. : in, It is a smoothing constant; The standardized feature values ​​within the same time window are concatenated according to modality category and feature category to form an addiction feature vector. The addiction feature vectors corresponding to each time window are arranged in chronological order to obtain the addiction feature sequence.

4. The high-precision addiction assessment method according to claim 1, characterized in that, The process of identifying the stable period, decomposing the diurnal rhythm, and removing abnormal windows based on the addiction feature sequence, and constructing an individualized baseline model for the evaluated subject, includes: The preset observation period is divided into multiple candidate time windows; Calculate the state fluctuation index for each candidate time window based on the addiction feature sequence; Based on the comprehensive state fluctuation value, objective behavioral anomalies, data missing ratio, and historical outcome markers, stable period windows are selected from multiple candidate time windows, and abnormal windows are removed. Perform diurnal rhythm decomposition on the aforementioned stable period window; Multi-layered individualized baseline vectors are constructed based on the results of diurnal rhythm decomposition.

5. The high-precision addiction assessment method according to claim 4, characterized in that, The state fluctuation index is calculated as follows: For the k-th feature within the candidate time window W, its fluctuation value Calculate using the following formula: in, This represents the feature value of the k-th feature within the time window t. This represents the average value of the k-th feature within the candidate time window W. This represents the number of sampling points within the candidate time window W; Based on the fluctuation values ​​of each feature, the comprehensive state fluctuation value of the candidate time window W is generated according to the following formula. : in, Contribute weights to the fluctuations corresponding to the k-th feature, and .

6. The high-precision addiction assessment method according to claim 4, characterized in that, The process of filtering stable windows and removing abnormal windows includes: A candidate time window is defined as a stable period window when it meets at least some of the following conditions: the intensity of addiction-related behaviors does not exceed the preset fluctuation range, sleep indicators do not show a continuous abnormal decline, physiological indicators fluctuate within the preset range, the frequency of exposure to high-risk environments is lower than the preset threshold, the proportion of missing data is lower than the preset proportion, and the overall state fluctuation value is lower than the stable period fluctuation threshold. When a candidate time window contains any of the following situations: confirmed or suspected relapse behavior, a sudden increase in addiction-related behavior exceeding a preset intensity threshold, a continuous abnormal decline in sleep duration or sleep quality, continuous approach to or prolonged stay at high-risk locations, a data missing ratio exceeding a preset threshold, or errors in the collection of objective behavioral data, it will be identified as an abnormal window and removed. When there is a conflict between subjective evaluation data and objective behavioral data, it is not used as an anomaly window removal condition, but is used as a subjective-objective consistency factor in the calculation of the credibility weight.

7. The high-precision addiction assessment method according to claim 4, characterized in that, When the number of stable period windows is lower than the preset minimum number, an initial baseline model is established by jointly using the population prior baseline and the collected individual data. The initial baseline value is... Calculate using the following formula: in, Let be the prior baseline value of the k-th feature in the population sample. The available statistical value of the k-th feature in the collected data of the evaluated object. For group prior weights, The value range is from 0 to 1.

8. The high-precision addiction assessment method according to claim 1, characterized in that, The generation of individualized baseline deviation scores includes: For the k-th feature in the current time window t, the baseline deviation is calculated using the following formula. : in, This represents the value of the k-th feature within the current time window t. This represents the individualized baseline value of the k-th feature within the rhythmic time period r to which the current time window t belongs. This indicates the corresponding baseline fluctuation scale. It is a smoothing constant; The baseline deviation is aggregated according to the feature category to obtain the multidimensional deviation. An individualized baseline deviation score is generated based on the multidimensional deviation using the following formula. : in, This represents the deviation in the d-th dimension. This represents the contribution weight of the deviation in the d-th dimension. This indicates a deviation from the scoring bias. .

9. The high-precision addiction assessment method according to claim 1, characterized in that, The data-based multidimensional credibility assessment index generates credibility weights for each modality of data, including: Calculate the data quality factor of the m-th mode within the time window t. ; The subjective-objective consistency factor is calculated based on the difference between subjective assessment data and objective risk characteristics. ,in: The standardized value corresponding to the subjective desire rating. Aggregated values ​​representing objective risk characteristics; Calculate the intermodal collaborative triggering factor based on whether different modal characteristics form a preset risk combination within the same time window. ; Prediction loss based on complete feature input from historical labeled samples and the prediction loss after removing the m-th modality features. Calculate historical prediction contribution factors ; Based on the data quality factor The subjective-objective consistency factor The intermodal cooperative triggering factor The historical prediction contribution factor and time decay factor Calculate the unnormalized confidence weight using the following formula. : The final credibility weight is obtained by normalizing according to the following formula. : in, This represents the initial empirical weights for the m-th modality.

10. The high-precision addiction assessment method according to claim 9, characterized in that, It also includes the step of adjusting the credibility weights based on the type of addiction: Pre-set addiction type adjustment coefficient The final credibility weight Revised to: The revised credibility weights are then normalized a second time to generate credibility weights adjusted for addiction type. : 。 11. The high-precision addiction assessment method according to claim 1, characterized in that, The risk prediction model is a time-window logistic regression risk model; The process of inputting the fused addiction risk feature vector, feature change trend, addiction severity score, individualized baseline deviation score, and risk confidence score into the risk prediction model to generate relapse risk prediction results includes: Based on the current feature vector of fusion addiction risk The change in the fusion addiction risk feature vector over the past K time windows Addiction severity score Individualized baseline deviation score and risk confidence level Calculate the future using the following formula Risk of relapse within the time window : in, For bias terms, These are the model parameters.

12. The high-precision addiction assessment method according to claim 1, characterized in that, The dynamic calibration includes: Labels were created based on follow-up results. When relapse or runaway use occurs within the predicted time window, set Otherwise set ; The model parameters of the risk prediction model are dynamically updated using the following cross-entropy loss function: The historical prediction contribution factor will be updated based on the follow-up results. The final confidence weight is adjusted based on the updated historical prediction contribution factors. .

13. A high-precision addiction assessment device, characterized in that, include: A multimodal data acquisition module is used to collect multimodal addiction-related data of the assessed subject. The multimodal addiction-related data includes at least two of the following: subjective assessment data, mobile terminal behavior data, wearable physiological data, environmental induced data, cognitive test data, and historical outcome data. The data preprocessing module is used to perform time synchronization processing, outlier handling, and missing value handling on the multimodal addiction-related data; The addiction feature extraction module is used to extract basic addiction feature indicators from the processed multimodal addiction-related data, standardize the basic addiction feature indicators, and generate an addiction feature sequence. The individualized baseline modeling module is used to identify the stable period, decompose the diurnal rhythm, and remove abnormal windows based on the addiction feature sequence, and construct an individualized baseline model. The baseline deviation scoring module is used to generate an individualized baseline deviation score based on the individualized baseline model. The multimodal credibility fusion module is used to generate credibility weights for each modality based on the multidimensional credibility assessment index of the data, and to generate a fused addiction risk feature vector based on the credibility weights and the individualized baseline deviation score. An addiction severity assessment module is used to generate an addiction severity score based on the fused addiction risk feature vector; The relapse risk prediction module is used to generate relapse risk prediction results based on the fused addiction risk feature vector, feature change trend, addiction severity score, individualized baseline deviation score and risk confidence level; An interpretable report generation module is used to generate an interpretable addiction assessment report based on the addiction severity score, the relapse risk prediction result, and the risk confidence level. The feedback calibration module is used to dynamically calibrate the individualized baseline modeling module, the multimodal reliability fusion module, and the relapse risk prediction module based on the follow-up results.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the high-precision addiction assessment method as described in any one of claims 1 to 12.

Images