An auxiliary method and system for depression risk and cause analysis by fusing a bracelet and a large language model

Abstract

The application discloses an auxiliary method and system for depression risk and cause analysis by fusing a bracelet and a large language model, and belongs to the technical field of signal recognition and processing. The method comprises the following steps: acquiring bracelet signal data; preprocessing the bracelet signal data to obtain a plurality of frequency band data; extracting features of the plurality of frequency band data; inputting the features into a trained support vector machine (SVM) and a large language model; and performing weighted average on depression classification probabilities obtained by the SVM and the large language model to obtain a depression risk assessment result. Meanwhile, the large language model performs cause analysis on features such as illumination intensity, exercise amount, life regularity index and circadian rhythm to obtain possible causes of depression, and the risk assessment result and the possible causes of depression are provided to doctors for auxiliary diagnosis. The application provides an effective solution to the problems in the prior art, such as limited auxiliary assessment effect, low reliability and incapability of providing pathological cause analysis.

Description

Technical Field

[0001] This invention relates to the field of signal recognition and processing technology, and more specifically to an auxiliary method and system for analyzing the risk and causes of depression by integrating a wristband and a large language model. Background Technology

[0002] Depression is the most common depressive disorder, characterized by a significant and persistent low mood, and is a major type of mood disorder.

[0003] Currently, the main methods for identifying and diagnosing clinical depression include: (1) Traditional depression screening and diagnosis, which involves questionnaires and interviews, requires the professional experience of doctors and the cooperation of patients, is highly subjective, and is prone to misjudgment. (2) Collecting several electrocardiogram (ECG) signals and comparing each ECG signal with a pre-set normal value. Based on the comparison results, clinicians assess the patient's depression level. However, the above judgment process often uses sensors such as EEG and ECG, which are difficult to integrate into wearable devices, to monitor and collect the user's vital signs and activities. It is difficult to collect the user's vital sign information for a long time, and the comprehensiveness is low.

[0004] Therefore, how to provide an auxiliary method and system for analyzing the risk and causes of depression by integrating a wristband and a large language model that can solve the above problems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an auxiliary method and system for analyzing the risk and causes of depression by integrating a wristband and a large language model. It has the characteristics of being objective, convenient, comprehensive in assessment, and low in cost, and solves the technical problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model includes the following steps: Acquire wristband signal data; The wristband signal data is preprocessed to obtain multiple frequency band data; Extract features from multiple frequency band data; The features are then input into the trained Support Vector Machine (SVM) and the large language model, respectively. The depression classification probabilities obtained from the Support Vector Machine (SVM) and the Large Language Model are weighted and averaged to obtain the depression risk assessment result.

[0007] Optionally, acquiring wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

[0008] Optionally, features of multiple frequency band data are extracted, including: Extract frequency band energy features, sample entropy features, spectral entropy features, Hjorth features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from the frequency band data. The frequency band data is analyzed to obtain the life pattern index and circadian rhythm.

[0009] Optionally, obtain the daily routine index and circadian rhythm: The expression for the regularity index is: ; The 24-hour regularity index R24(x) is: ; The 12-hour regularity index R12(x) is: ; The expression for circadian rhythm is: ; In the formula, psd(x) represents the power spectral density at frequency x.

[0010] Optionally, the large language model also performs causal analysis through lifestyle regularity index and circadian rhythm to obtain prompt words, wherein the prompt words include: system prompt words and user data prompt words.

[0011] An auxiliary system integrating a wristband and a large language model for analyzing the risk and causes of depression includes: Acquisition module: Acquires wristband signal data; Preprocessing module: preprocesses the wristband signal data to obtain multiple frequency band data; Feature extraction module: Extracts features from multiple frequency band data; Feature input module: Inputs the features into the trained Support Vector Machine (SVM) and the large language model respectively; Risk assessment module: The depression classification probabilities obtained from support vector machine (SVM) and large language model are weighted and averaged to obtain the depression risk assessment result.

[0012] Optionally, acquiring wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

[0013] Optionally, features of multiple frequency band data are extracted, including: Extract frequency band energy features, sample entropy features, Hjorth Activity features, Mobility features, Complexity features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from the frequency band data. The frequency band data is analyzed to obtain the life pattern index and circadian rhythm.

[0014] Optionally, obtain the daily routine index and circadian rhythm: The expression for the regularity index is: ; The 24-hour regularity index R24(x) is: ; The 12-hour regularity index R12(x) is: ; The expression for circadian rhythm is: ; In the formula, psd(x) represents the power spectral density at frequency x.

[0015] Optionally, the large language model also performs causal analysis through lifestyle regularity index and circadian rhythm to obtain prompt words, wherein the prompt words include: system prompt words and user data prompt words.

[0016] As can be seen from the above technical solution, compared with the prior art, this invention discloses an auxiliary method and system for analyzing the risk and causes of depression by integrating a wristband and a large language model. It acquires data from the wristband every minute, recording the user's acceleration in three directions (x, y, z), ambient light intensity, ambient sound intensity, and the number of Bluetooth devices in the user's vicinity. The user wears the wristband for seven days, constituting one cycle. Data features are then extracted and combined with machine learning methods and a large language model to improve the recognition rate of depression detection. This method is objective, convenient, comprehensive, and low-cost. Simultaneously, the large language model analyzes the causes of depression based on features such as light intensity, exercise volume, lifestyle regularity index, and circadian rhythm, obtaining possible causes of depression. These results, along with the risk assessment, are provided to doctors for auxiliary diagnosis. This invention provides an effective solution to the problems of limited auxiliary assessment effectiveness, low reliability, and inability to provide pathological cause analysis in existing technologies. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the process for the auxiliary detection method for depression risk of the present invention; Figure 2 This is a hardware block diagram of a wristband system provided in an embodiment of the present invention; Figure 3 This is a processing logic diagram of a wristband system provided in an embodiment of the present invention; Figure 4(a) shows the original hourly ambient sound amplitude and corresponding spectrum of participants with high circadian rhythm values ​​provided in the embodiment of the present invention; Figure 4(b) shows the original hourly ambient sound amplitude and corresponding spectrum of participants with low circadian rhythm values ​​provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the depression risk auxiliary detection system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention discloses an auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model, comprising the following steps: Acquire wristband signal data; The wristband signal data is preprocessed to obtain data in multiple frequency bands; Extract features from data across multiple frequency bands; The features are input into the trained Support Vector Machine (SVM) and the large language model; The depression classification probabilities obtained from the support vector machine (SVM) and the large language model are weighted and averaged to obtain the depression risk assessment results.

[0021] In one specific embodiment, acquiring wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

[0022] Specifically, the wristband collects data every minute, recording the user's acceleration in three directions (x, y, z), ambient light intensity, ambient sound intensity, and the number of Bluetooth devices around the user. The user wears the wristband for seven days, which constitutes one cycle. Data features are extracted and combined with machine learning methods and large language models to analyze and diagnose the user's risk of depression and analyze the possible causes of depression.

[0023] For details, see Figure 2 The diagram shown is a hardware block diagram of the wristband system provided in this embodiment. The wristband uses an ESP32-C3-Super Mini as the main controller. In this embodiment, the development environment is Espressif's ESP-IDF combined with VSCODE's ESP-IDF extension. The main controller, along with the ADXL3453 axial accelerometer (I2C), INMP441 digital microphone (I2S), light acquisition circuit, LDO voltage regulator circuit, and lithium battery charging circuit, constitute the entire system. The system is small, lightweight, wearable, and has low power consumption. It can collect the wearer's motion information (X, Y, Z axis acceleration), ambient light intensity, ambient noise level, and the number of Bluetooth devices in the surrounding environment. The wristband is powered by an 800mAh 3.7V lithium battery. In the power consumption estimation, it is estimated that a full charge of the wristband can last for about 12 days, ensuring that a week's worth of data will not be lost. The wristband's data is stored in FLASH memory; data will be cleared upon power failure. This is to ensure the wearer's personal privacy and security, as well as the portability of the device.

[0024] Specifically, the wristband includes: a power management module for low-power management; an ESP32-C3 central processing module; an ADXL345 accelerometer module to collect the user's acceleration in the X, Y, and Z directions; a recording module to collect the sound intensity of the user's environment; a lighting module to collect the light intensity of the user's environment; and a Bluetooth module to collect the number of Bluetooth devices that can be scanned around the user. A storage module handles data storage, and a Wi-Fi module handles network communication. These modules connect to the central processing module, sending the collected data to it and writing it into the memory for storage.

[0025] The system will start counting from the set time. Every time the timestamp changes by 60 seconds (i.e., 1 minute), it will wake up to collect X, Y, and Z acceleration, light intensity, and volume data, as well as the number of Bluetooth devices in the vicinity. In order to ensure the system's power consumption, extend the usage time, and prevent data loss due to power failure, the scanning time has been changed to 30 minutes. The scanning cycle can be modified.

[0026] The exported data format is: Timestamp|X|Y|Z|Bluetooth Quantity|Volume|. The timestamp is a time_t variable, X, Y, and Z are short variables, illumination is an unsigned short variable, Bluetooth quantity is an int variable, and volume is a float variable.

[0027] The light intensity was converted from voltage to a digital value ranging from 0 to 4095 using an AD converter. To distinguish the Bluetooth quantity from a true 0 and records where scanning was not enabled, -1 was used as a marker for when scanning was not enabled. The I2S initialization set the sampling rate to 16kHz, the data bit width of the sampling point to 16 bits, and used mono recording. See the processing logic below. Figure 3 As shown.

[0028] Specifically, a massive dataset will be built using motion data, ambient light, Bluetooth connectivity, and volume data collected by the wristband, providing data support and detection methods for early-stage mental illnesses such as depression.

[0029] In one specific embodiment, features of multiple frequency band data are extracted, including: Extract frequency band energy features, sample entropy features, spectral entropy features, Hjorth features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from frequency band data; By analyzing the frequency band data, we can obtain the life pattern index and circadian rhythm.

[0030] Specifically, the various collected signal sequences are denoted as x(n), which can be any of the following signals: user acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth number. The statistical characteristics of these data are described below.

[0031] In one specific embodiment, feature extraction includes: The data signal is x(n) (representing any one of four signals: acceleration, light, sound, and Bluetooth). Since one data point is sampled per minute, there are 24 data points per day. The sampling time is 60 minutes, so the sampling frequency is fs = 1440 (times / day). Each user is sampled for 7 days, so each user's acceleration, illumination, sound, and Bluetooth data are all collected over 7 days. A time series consisting of 1440 points.

[0032] Specifically, regarding frequency band energy characteristics: Existing technologies indicate that circadian rhythms are crucial for human sleep and emotional states, and the corresponding frequency band energy can characterize emotional features [29-31]. Therefore, frequency band filtering is applied to any signal x(n), dividing the frequency bands into B1~B5: [0, 1 / 24fs], [1 / 24fs, 1 / 12fs], [1 / 12fs, 1 / 6fs], [1 / 6fs, 1 / 3fs], [1 / 3fs, 1 / 2fs], and the entire frequency band B0, i.e., [0, 1 / 2fs]. This divides the signal filtering into a total of 6 frequency band signals B0~B5. The sum of squares of each frequency band signal is calculated to obtain the energy of the 6 frequency bands. Since there are a total of four types of signals: acceleration, light, sound, and Bluetooth, each signal has 6 energy characteristics, so there are a total of 4 6 = 24 features.

[0033] Specifically, for the sample entropy features: the sample entropy features of each signal sample in each frequency band B0~B5 of each channel are calculated using the following formula, for a total of 6 features; where the sequence x0, x2…x5 is the data of signal x(n) divided into the above 6 frequency bands. Each signal has 6 sample entropy features, for a total of 4 6 = 24 features.

[0034] Specifically, sample entropy represents the degree of disorder between the beginning and end of a time series. The higher the sample entropy, the more disordered the data and the more complex the pattern. In wristband data, higher sample entropy indicates a more disordered lifestyle and a more chaotic and complex biological clock. Let there be a sequence of length... Original time series u represents the six sub-frequency band signals of x(n), where u = xi (i = 1~7). The specific algorithm is as follows: Using sequences structure 3D vector (1) Among them, 1 .

[0035] Define any two The distance between the two vectors is , (2) In the formula, 0 1 1 , .

[0036] Given distance tolerance ( ),when When two vectors are considered similar, then each vector is statistically analyzed. Calculated distance The number, and using This indicates that the ratio is calculated. : (3) right To calculate the mean: (4) dimension Increase by 1, that is, for By repeating the steps of formulas (1) to (4) for the dimensional vector, we can obtain the following results. ; Therefore, we can obtain the data length The sample entropy of the original sequence under finite values ​​is: (5) Specifically, regarding the spectral entropy characteristic: Suppose that for each channel signal x(n), the corresponding Fourier transform result S(n) is obtained. Then the spectral probability distribution of the signal is: (m=1, 2…n) (6) The spectral entropy of the signal is then: (7) Spectral entropy can characterize the degree of disorder in the spectral distribution of data; the larger the value, the more irregular the spectral distribution of the data.

[0037] Specifically, regarding Hjorth features: Hjorth parameters are widely used in the feature analysis and processing of EEG signals due to their ability to characterize the statistical features of EEG signals in the time domain and their relatively low computational cost. Activity reflects the average power of the EEG signal by calculating the variance of the signal; Mobility represents the ratio of the root mean square of the slope to the root mean square of the amplitude of the EEG signal, which is a measure of the average frequency of the EEG signal and is used to estimate the average frequency. The Mobility parameter is proportional to the standard deviation of the power spectrum, and the Complexity parameter provides an estimate of the signal bandwidth. Complexity calculates the root mean square of the ratio of the change in the signal slope to the ideal curve, used to measure the bandwidth of the EEG signal, and can indicate the similarity between the EEG signal and a pure sine wave. Hjorth parameters can effectively represent the power distribution characteristics of time-varying signals in the time-frequency domain. Since the calculation of Hjorth parameters is based on variance, it has a lower computational cost compared to wavelet transform and short-time Fourier transform, making it suitable for real-time tasks. After the signal x(n) is decomposed into 6 frequency band signals B0~B5 using a filter bank, the three parameters of Hjorth are calculated according to the following formulas. The specific calculations are as follows.

[0038] It is the variance of x(n) (8) in (9) (10) Specifically, regarding the Teager-Kaiser energy operator (TKO): The Teager-Kaiser Operator (TKO) is a nonlinear energy operator capable of determining the instantaneous energy of non-stationary signals. Besides energy, TKO can also track instantaneous amplitude and signal frequency. Local amplitude and frequency fluctuations can be effectively separated using TKO, and it offers better tracking capabilities for non-stationary signals.

[0039] For a continuous signal x(t), TKO is defined as: (11) in, Let represent the first and second derivatives of the signal x(t), respectively. For the discretized signal x(n) corresponding to x(t), the TKO operator is calculated as follows: (12) In the specific calculation, x(n) is first subjected to the following Discrete Wavelet Transform (DWT), as shown in Equation (13). After obtaining the corresponding wavelet transform coefficients, the wavelet transform coefficients of each layer are processed according to the Discrete TKO operator above, and finally the corresponding mean and variance are calculated.

[0040] (13)

[0041] In one specific embodiment, the expression for calculating the mean is: (14); The expression for calculating the maximum value: (15); The expression for calculating the median: (16); The expression for calculating kurtosis: (17); The expression for calculating skewness: (18); Specifically, the indicators for obtaining a regularity index include: Biological rhythms have a significant impact on a person's emotional state. This embodiment defines a person with a regular lifestyle as someone whose total intensity of physical activity, light exposure, and ambient sound intensity between 7 AM (morning) and 11 PM (evening) is greater than the total intensity between 11 PM (evening) and 7 AM. A higher ratio indicates greater daytime energy and a calmer nighttime, reflecting a stronger biological clock and a higher life regularity index. Therefore, the Life Regularity Index (LRI) is defined as: ; (19) Meanwhile, the 24-hour regularity index R24(x) is defined as: (20); The 12-hour regularity index R12(x) is: (twenty one); Through the above technical solution, the life regularity index is obtained, B1~B5: [0, 1 / 24fs], [1 / 24fs, 1 / 12fs], [1 / 12fs, 1 / 6fs], [1 / 6fs, 1 / 3fs], [1 / 3fs, 1 / 2fs], and the entire frequency band B0, i.e., [0, 1 / 2fs], fs=1440 times / day; X(i) is the data sampled by the wristband in the i-th minute, such as the light intensity, exercise, etc. in the i-th minute.

[0042] Specifically, circadian rhythm (CR) measures the intensity of a user's activity rhythm over a 24-hour period.

[0043] Humans have a biological clock that optimizes an organism's physiology and behavior, as well as hormone secretion and mood. However, people's circadian rhythms vary. For example, previous research has shown that individual differences in morning and evening rhythms are related to personality. This embodiment uses spectral analysis to calculate the circadian rhythm of hourly behavioral data throughout the study period. Specifically, this embodiment first uses least-squares spectral analysis to convert the behavioral data (i.e., motion, light, and ambient sound) from the time domain to the frequency domain psd(x). Then, this embodiment calculates the CR of energy in the 24 ± 0.5 h period (corresponding to 2π / (24 ± 0.5) = (0.2565, 0.2674)) and the total spectral energy in the 24 ± 12 h period (corresponding to 2π / (24 ± 12) = (0.1745, 0.5236)): (twenty two); Where psd(x) represents the power spectral density at frequency x, which is the frequency spectrum and can be understood as the power intensity after Fourier transform (power is the square of amplitude).

[0044] Before performing spectral analysis, the raw hourly signals were first aligned to zero mean. Figures 4(a)-(b) show an example of calculating the CR for ambient sound. They display the raw hourly ambient sound amplitudes and corresponding spectrograms for two selected participants. As shown in Figures 4(a)-(b), the raw data indicates that participant (a) has a more pronounced 24-hour ambient sound cycle than participant (b).

[0045] In one specific embodiment, during the model prediction phase, the goal is to provide users with an accurate assessment of their depression probability and an analysis of possible causes of depression. To this end, this embodiment relies on the collaborative work of a Support Vector Machine (SVM) and the large language model ChatGPT-4o, integrating multimodal data to achieve prediction. The resulting probability can provide a reference for whether a user suffers from depression.

[0046] Specifically, the model primarily outputs the user's depression probability and an analysis of the causes of depression. The depression probability is obtained by weighted averaging of predictions from a support vector machine (SVM) and the large language model ChatGPT-4o, while the analysis of the causes of depression is provided by ChatGPT-4o.

[0047] In a specific embodiment, the specific steps for training the Support Vector Machine (SVM) and the large language model are as follows: First, data from multiple batches of subjects were collected, including motion data, voice features, ambient lighting, and nearby Bluetooth counts. This data was then used to train the model to improve its accuracy and performance. By analyzing data collected from users over two weeks, changes in their mood—whether their depressive tendencies worsened or improved—could be predicted. Furthermore, changes in depressive mood were assessed by comparing changes in the subjects' PHQ-9 scores before and after the experiment. The goal of this embodiment is to predict changes in depressive mood using only motion and voice data features, and to use correlation coefficients and p-values ​​to represent the correlation between the predicted results and the actual situation, as well as the statistical significance of the correlation coefficients.

[0048] Specifically, the input to the Support Vector Machine (SVM) model consists of four initial data points after feature extraction: acceleration magnitude, ambient light intensity, ambient sound intensity, and the number of Bluetooth devices around the user. Its output is the user's probability of depression.

[0049] Specifically, the ChatGPT-4o large language model takes as input data the user's 14-day activity intensity, ambient light intensity, sound intensity, number of Bluetooth connections, etc., and calculates sample entropy, spectral entropy, daily routine index, 24-hour routine index, and circadian rhythm. Outputs include the user's probability of depression and an analysis of the causes of depression.

[0050] Specifically, the depression classification probabilities obtained from Support Vector Machine (SVM) and Large Language Model are weighted and averaged to obtain the depression risk assessment results. The weighted average summation formula is as follows: ; In the formula, This is the final depression risk score after weighted averaging. The depression risk value output by the Support Vector Machine (SVM) model. The depression risk score output by the large-scale language ChatGPT-4o model. and These are the weights for the SVM and ChatGPT-4o models, respectively.

[0051] In one specific embodiment, a weighted average is used to combine the predictions from Support Vector Machine (SVM) and ChatGPT-4o according to their respective weights to obtain a more comprehensive assessment of depression risk. The weights reflect the reliability or importance of each model. This weighted averaging method effectively integrates the advantages of multiple models, improving the accuracy of the final prediction.

[0052] Specifically, the architecture of the Support Vector Machine (SVM) model is designed to classify data into different categories by finding the optimal hyperplane. In this embodiment, the SVM is used for a binary classification task: predicting a user's risk of depression (i.e., depression versus health).

[0053] x i It is a feature vector formed after a user uses a wristband to collect motion, lighting, voice, and Bluetooth data, and then performs the feature extraction process described above. y i The labels are generated after users fill out the PHQ9 depression scale and receive on-site diagnoses from mental health counselors. The label for depression is 1, and the label for health is 0. This project uses a dataset of 97 users (27 depressed and 70 healthy) collected in the early stage to train the model.

[0054] Specifically, the basic principle of Support Vector Machines (SVM) is to find a decision boundary that maximizes the margin between different classes. For a given training dataset... (x i ，y i ) ,in x i It is an eigenvector. y i The hyperplane is determined by SVM through the following optimization problem: (The hyperplane is a class label.) ; ; in, w It is a weight vector. b It's a bias. ξ i It is a slack variable. C It is a penalty parameter used to control the trade-off between training error and model complexity.

[0055] In one specific embodiment, an RBF (Radial Basis Function) kernel was chosen because it can effectively handle linearly inseparable data. The RBF kernel solves nonlinear problems by mapping to a higher-dimensional space; its mathematical expression is: ; in, γ These are kernel parameters that control the width of the RBF kernel.

[0056] Specifically, in terms of parameter selection, this embodiment uses the GridSearchCV method to find the optimal model parameters. This method iterates through all possible parameter combinations, evaluates performance using cross-validation, and finally selects the best-performing parameter combination. In this embodiment, the optimal parameter combination found is {"C": 10, "gamma": 1, "kernel": "rbf"}. GridSearchCV evaluates the performance of different parameter combinations by performing cross-validation on multiple folds. This method ensures the generalization ability of the selected parameters and improves the model's predictive performance.

[0057] The model outputs the user's depression probability, normalized to a value between 0 and 1, representing a preliminary assessment of the user's depression risk.

[0058] In one specific embodiment, the large language model ChatGPT-4o supplements the SVM's predictions by analyzing more complex behavioral and environmental features. This model particularly focuses on users' behavioral patterns and daily activity routines to provide a more in-depth analysis of the causes of depression. This embodiment tests and applies the large language model ChatGPT-4o using data from typical depressed and healthy users.

[0059] In one specific embodiment, the large language model ChatGPT-4o performs causal analysis using a lifestyle regularity index and circadian rhythms to obtain cue words. These cue words include system cue words and user data cue words. The system cue words provide ChatGPT-4o with a comprehensive data context and analytical framework, ensuring the model understands how to use input data to assess depression risk. They detail the data collection process, feature calculation methods, and the meaning of each feature, thus laying the foundation for the model's analysis. User data cue words focus on the specific information of the individual user input, including age, gender, and extracted feature sequences.

[0060] Specifically, the system prompt is: System-Prompt = "As a professional mental health specialist, please analyze the user's depression risk based on data collected from the user wearing the bracelet for 7 days, including exercise intensity, ambient light intensity, sound intensity, and the number of nearby Bluetooth devices. Data is collected once per minute, for 1440 minutes per day, with 1440 data points of each type per day. By analyzing this data, extracting features, and analyzing the user's depression risk."

[0061] Specifically, the following are the features used for analysis: Feature Data Structure Description: Features are calculated from data such as the user's 7-day activity intensity, ambient light intensity, sound intensity, and number of surrounding Bluetooth devices. Each feature is separated by a semicolon (;). Except for age and gender, features 1-4 are the means of the four data points (activity intensity, ambient light intensity, sound intensity, and number of surrounding Bluetooth devices); features 5-8 are the standard deviations of these data points; features 9-12 are the sample entropy; features 13-16 are the spectral entropy; features 17-20 are the lifestyle regularity index; features 21-24 are the 24-hour regularity index; and features 25-28 are the circadian rhythm.

[0062] Feature Description: 1) Mean and Standard Deviation: Calculates the mean and standard deviation of the four data points. 2) Sample Entropy: Calculates data complexity in the time domain; a higher value indicates a more complex and disordered lifestyle. 3) Spectral Entropy: Calculates data complexity in the frequency domain; a higher value indicates a more complex and disordered lifestyle. 4) Life Regularity Index: Calculates the ratio of a user's activity level from 7 AM to 11 PM to their activity level from 11 PM to 7 AM the next day; a higher value indicates a stronger biological clock and more regular lifestyle. 5) 24-Hour Regularity Index: A higher value indicates a biological clock cycle closer to 24 hours. 6) Circadian Rhythm: A higher index indicates a regular sleep-wake cycle and clear day-night cycle. Healthy users tend to have stronger daytime activity, light exposure, and sound intensity, while these are weaker at night; depressed users show the opposite. Based on these characteristics, determine the likelihood of the user's depression.

[0063] Required output format: Depression level: A value between 0 and 1, with lower numbers indicating better health and higher numbers indicating more severe depression; Diagnosis of cause: Please explain in detail the cause of depression.

[0064] The following features are provided. Please assess the user's risk of depression based on these features and output the data in the prescribed format.

[0065] Specifically, the user data prompt is: User-data-prompt = "The user's age is {age}, gender is {gender}, and the following are the features {fearures}".

[0066] Where {age}, {gender}, and {feature} are replaced by the user's actual age, gender, and extracted feature sequences, respectively. The feature sequences are calculated according to the above formulas (1) to (22), converted into strings, and a total of 30 features are connected with semicolons ";" to form the feature sequence feature.

[0067] On the other hand, see Figure 5 This invention discloses an auxiliary system for analyzing the risk and causes of depression by integrating a wristband and a large language model, comprising: Acquisition module: Acquires wristband signal data; Preprocessing module: preprocesses the wristband signal data to obtain multiple frequency band data; Feature extraction module: Extracts features from multiple frequency band data; Feature input module: Inputs the features into the trained Support Vector Machine (SVM) and the large language model respectively; Risk assessment module: The depression classification probabilities obtained from support vector machine (SVM) and large language model are weighted and averaged to obtain the depression risk assessment result.

[0068] In one specific embodiment, acquiring wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

[0069] In one specific embodiment, features of multiple frequency band data are extracted, including: Extract frequency band energy features, sample entropy features, Hjorth Activity features, Mobility features, Complexity features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from the frequency band data. The frequency band data is analyzed to obtain the life pattern index and circadian rhythm.

[0070] In one specific embodiment, the regularity index and circadian rhythm are obtained: The expression for the regularity index is: ; The 24-hour regularity index R24(x) is: ; The 12-hour regularity index R12(x) is: ; The expression for circadian rhythm is: ; In the formula, psd(x) represents the power spectral density at frequency x.

[0071] In one specific embodiment, the large language model also performs causal analysis through a life pattern index and circadian rhythm to obtain prompt words, wherein the prompt words include: system prompt words and user data prompt words.

[0072] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0073] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model, characterized in that, Includes the following steps: Acquire wristband signal data; The wristband signal data is preprocessed to obtain multiple frequency band data; Extract features from multiple frequency band data; The features are then input into the trained Support Vector Machine (SVM) and the large language model, respectively. The depression classification probabilities obtained from the Support Vector Machine (SVM) and the Large Language Model are weighted and averaged to obtain the depression risk assessment result.

2. The auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model according to claim 1, characterized in that, The acquisition of wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

3. The auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model, as described in claim 1, is characterized in that... Extracting features from multiple frequency band data includes: Extract frequency band energy features, sample entropy features, spectral entropy features, Hjorth features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from the frequency band data. The frequency band data is analyzed to obtain the life pattern index and circadian rhythm.

4. The auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model according to claim 3, characterized in that, Obtain the lifestyle regularity index and circadian rhythm: The expression for the regularity index is: The 24-hour regularity index R24(x) is: The 12-hour regularity index R12(x) is: The expression for circadian rhythm is: In the formula, psd(x) represents the power spectral density at frequency x.

5. The auxiliary method for analyzing the risk and causes of depression by integrating a wristband and a large language model according to claim 1, characterized in that, The large language model also performs causal analysis through lifestyle regularity index and circadian rhythm to obtain prompt words, which include system prompt words and user data prompt words.

6. An auxiliary system integrating a wristband and a large language model for analyzing the risk and causes of depression, characterized in that, include: Acquisition module: Acquires wristband signal data; Preprocessing module: preprocesses the wristband signal data to obtain multiple frequency band data; Feature extraction module: Extracts features from multiple frequency band data; Feature input module: Inputs the features into the trained Support Vector Machine (SVM) and the large language model respectively; Risk assessment module: The depression classification probabilities obtained from support vector machine (SVM) and large language model are weighted and averaged to obtain the depression risk assessment result.

7. The auxiliary system for analyzing depression risk and causes by integrating a wristband and a large language model according to claim 6, characterized in that, The acquisition of wristband signal data includes: acquiring the user's acceleration motion signal, motion signal, light intensity signal, sound intensity signal, and Bluetooth count through the wristband worn by the user according to a preset period.

8. The auxiliary system for analyzing depression risk and causes by integrating a wristband and a large language model according to claim 7, characterized in that, Extracting features from multiple frequency band data includes: Extract frequency band energy features, sample entropy features, Hjorth Activity features, Mobility features, Complexity features, Teager-Kaiser energy operator features, mean features, minimum features, and kurtosis features from the frequency band data. The frequency band data is analyzed to obtain the life pattern index and circadian rhythm.

9. The auxiliary system for analyzing depression risk and causes by integrating a wristband and a large language model according to claim 8, characterized in that, Obtain the lifestyle regularity index and circadian rhythm: The expression for the regularity index is: The 24-hour regularity index R24(x) is: The 12-hour regularity index R12(x) is: The expression for circadian rhythm is: In the formula, psd(x) represents the power spectral density at frequency x.

10. The auxiliary system for analyzing the risk and causes of depression by integrating a wristband and a large language model according to claim 6, characterized in that, The large language model also performs causal analysis through lifestyle regularity index and circadian rhythm to obtain prompt words, which include system prompt words and user data prompt words.

Images