Method for predicting depression risk of chronic disease patient

Abstract

The invention provides a chronic disease patient depression risk prediction method, and relates to the field of chronic disease patient depression risk prediction. The chronic disease patient depression risk prediction process provided by the invention comprises the steps of constructing a chronic disease patient depression risk prediction data set, constructing a deviation value sequence based on a median value of each factor data, mapping the deviation value sequence into a high-dimensional feature sequence, calculating a tension value and an overall tension mean value of the high-dimensional feature sequence, and predicting the depression risk of the chronic disease patient. Introducing an adaptive switching factor and nonlinear phase transformation to obtain a fusion feature sequence; extracting a local segment of the fusion feature sequence by adopting a sliding window; calculating the local complexity and trend bending rate of the local segment; introducing a local control factor; the final feature representation of each factor data is generated, the full-connection neural network is input to calculate the chronic disease patient depression risk prediction value of each factor data, the mean value is taken as the final prediction value, and the method can cooperatively utilize local change and overall trend to enhance prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of depression risk prediction for chronic disease patients, and specifically relates to a method for predicting depression risk in chronic disease patients. Background Art

[0002] The depression risk of chronic disease patients refers to the possibility that an individual is more likely to have depressive symptoms or develop clinical depression than the general population due to the physical pain, reduced quality of life, and increased psychological stress brought about by long-term diseases. This depression risk will exacerbate the condition of the chronic disease itself, reduce treatment compliance, increase the incidence of complications and disability rates. Timely and accurate prediction of the depression risk of chronic disease patients can help the medical team to intervene early and develop personalized physical and mental treatment plans, which can not only improve the mental health status of patients, but also significantly improve the management effect of chronic diseases and improve the quality of life of patients.

[0003] The prediction of depression risk in chronic disease patients mainly relies on means such as regular psychological questionnaires and artificial interviews, which have problems such as strong subjectivity, long cycle, and response lag, and it is difficult to achieve individualized and dynamic management. Although the prediction models of some machine learning methods have improved the automation level to a certain extent, the existing methods have not effectively characterized the complex non-linear interaction relationships between various data, resulting in insufficient generalization ability of the models. Secondly, there is a lack of a dynamic modeling mechanism for the state evolution process of chronic disease patients, and it is difficult to capture the key changes from stable accumulation to sudden outbreak.

[0004] The factor data collected for the prediction of depression risk in chronic disease patients usually covers multiple dimensions. These data not only have highly non-linear characteristics in the time series, but also show the evolution laws of multi-scale and multi-stage. Therefore, it is very important for the prediction model to simultaneously have the sensitivity to short-term fluctuations and the modeling ability of long-term trend evolution to accurately predict the depression risk of chronic disease patients. Summary of the Invention

[0005] The present invention provides a method for predicting depression risk in chronic disease patients, which uses non-linear transformation to convert the deviation sequence of various factor data relative to the median value into a high-dimensional feature sequence, calculates the tension value and the overall tension mean of the high-dimensional feature sequence, introduces an adaptive switching factor to adjust the activation intensity of the tension value, combines phase transformation to obtain an updated fused feature sequence, then uses a sliding window mechanism to divide the fused feature sequence into continuous local segments, calculates the local complexity and trend curvature rate to characterize the dynamic change characteristics, introduces a local control factor, generates multi-scale path weights according to the distance between the control factor and the segment center, realizes the weighted fusion of local segments, and finally inputs the weighted fused features into a fully connected neural network to independently predict the depression risk values corresponding to each factor data, and takes their average as the final output to improve the accuracy and stability of the prediction of depression risk in chronic disease patients.

[0006] The technical method adopted by the present invention to achieve the above object specifically includes the following steps: S1. Collect factor data affecting the depression risk of chronic disease patients, and construct a prediction data set for the depression risk of chronic disease patients; S2. The factor data constitutes time series data. Obtain the median value of each factor time series data, and calculate the deviation sequence between each factor time series data and the corresponding median value. Map the deviation sequence to a high-dimensional feature sequence through non-linear transformation; S3. Calculate the tension value and overall tension mean of the high-dimensional feature sequence, introduce an adaptive switching factor, and combine non-linear phase transformation to construct an adaptive switching evolution module. Integrate the high-dimensional feature sequence and local tension changes to obtain a fused feature sequence; S4. Use a sliding window mechanism to divide the fused feature sequence into multiple consecutive local segments, construct a local change characterization module, and calculate the local complexity and local trend curvature rate of each local segment respectively; S5. According to the local complexity and local trend curvature rate, introduce a local control factor, generate multi-scale path weights based on the distance between the local control factor and the center of each local segment, and fuse the local segments with weights to obtain the final feature representation of each factor data; S6. Use a fully connected neural network to calculate the final feature representation of each factor data, obtain the prediction value of the depression risk of chronic disease patients, add all the prediction values and take the average as the final prediction value of the depression risk of chronic disease patients.

[0007] Preferably, in S1, the factor data affecting the depression risk of chronic disease patients is collected, including blood glucose level, blood pressure, heart rate, exercise frequency, anxiety assessment value, stress assessment value, disease type, drug use record. Resampling and interpolation techniques are used to complement the missing data to ensure the continuity of the data, and a prediction data set for the depression risk of chronic disease patients is constructed.

[0008] Preferably, in S2, input the time series data of the th factor data affecting the depression risk of chronic disease patients , the time series length is , calculate the deviation sequence between the th factor data and its median value to achieve the measurement of the abnormality degree of chronic disease patients. The specific mathematical model is: ; In the formula, is the deviation sequence between the th factor data and the corresponding median value, is the sensitivity hyperparameter of the th factor data, is the Median value of factor data; Using non - linear transformation to transform the deviation sequence into a high - dimensional feature sequence, automatically suppressing extreme data input and ensuring the stability of the model. The specific mathematical model is: ; In the formula, is the high - dimensional feature sequence of the th factor data. The dimension of the high - dimensional feature is , is a learnable parameter matrix, is a learnable bias term.

[0009] Preferably, by calculating the deviation between the factor data of the depression risk of chronic disease patients and the median value and performing non - linear mapping, a deviation potential energy distribution reflecting the degree of feature abnormality is formed. Further, non - linear transformation is used for high - dimensional mapping to strengthen feature expression, automatically suppressing the influence of extreme data on the model stability, and providing a more robust and fine - grained high - dimensional feature basis for subsequent unified modeling.

[0010] Preferably, in S3, an adaptive switching and evolution module is constructed to simulate the natural evolution process of the depression risk of chronic disease patients from stable accumulation to sudden outbreak, and update the overall features of the high - dimensional feature sequence to obtain a richer overall feature representation, including the following steps: S31: Using second - order and fourth - order tension control hyperparameters, respectively perform joint weighted integration on the element - by - element square and fourth - power results of the high - dimensional feature sequence of the th factor data to obtain the tension value sequence of the th factor data. Further, normalize and average the tension values of each time step of the th factor data within the full sequence range to obtain the overall tension mean of the th factor data. The specific mathematical model is: ; ; In the formula, is the tension value sequence of the th factor data, is the second - order tension control hyperparameter, is the fourth - order tension control hyperparameter, is 's element - by - element square, is 's element - by - element fourth - power, is the overall tension mean of the th factor data, is the Tension value of the time step; S32. Calculate the absolute difference between the tension value and the overall tension mean value, construct an adaptive switching factor in combination with the normalization function, and adjust the activation degree of the tension value. The specific mathematical model is: ; In the formula, is the adaptive switching factor of the th factor data, is the absolute value operation, is the hyperparameter of the stability threshold; S33. Use the adaptive switching factor as the adjustment weight, combine the non-linear phase change, and fuse the high-dimensional feature sequence and the local tension change to obtain the fused feature sequence after fusion. The specific mathematical model is: ; In the formula, is the fused feature sequence of the th factor data after fusion.

[0011] Preferably, calculating the tension value and the overall tension mean value can help the model capture small changes and large anomalies in the factor data of the depression risk of chronic disease patients. During the change process of the factor data, not all fluctuations of the tension value are worth triggering the evolution of the model. Therefore, an adaptive switching factor is introduced, which can naturally allocate the local evolution weight according to the deviation between the local tension and the overall tension mean value. When there is a small disturbance, the switching factor suppresses the local evolution, while when the local anomaly is significant, it activates the local drastic change. The adaptive switching factor combines the non-linear phase change, enabling the model to achieve the coexistence of continuous smooth evolution and local jump phenomena, greatly enriching the fine-grained simulation ability of the model for the change process of the depression risk of chronic disease patients.

[0012] Preferably, in S4, a sliding window mechanism is used to divide the time series data into multiple continuous local segments, and the local complexity and local trend curvature are calculated based on each segment to characterize the local change characteristics, including the following steps: S41. Use the sliding window to divide the fused feature sequence of the th factor data into local segments. The size of the sliding window is , and the sliding step is , , and a total of local segments are generated. Denote the th local segment as ; S42. Construct the local complexity. Calculate the local mean in the feature direction for each local segment, calculate the square of the difference between the fused feature value within the local segment and the corresponding mean, and normalize it within the time step and feature dimension ranges to obtain the local complexity, which is used to describe the local fluctuation intensity. The specific mathematical model is as follows: ; ; In the formula, is the th local segment of the th factor data, feature mean, is the th local complexity of the th local segment of the th factor data, is the th local segment of the th time step of the th factor data, is the dimension of the high-dimensional feature; S43. Calculate the discrete second-order difference of the feature values of adjacent time steps within the local segment, and normalize it to obtain the local trend curvature rate to capture the degree of change inflection, making up for the problem that the local complexity can only perceive the intensity but not the trend change. The specific mathematical model is as follows: ; In the formula, is the th local trend curvature rate of the th local segment of the

[0013] Preferably, the fused feature sequence is locally partitioned through a sliding window mechanism, enabling the model to extract local dynamic information while maintaining the global structure. On this basis, two types of feature indicators, namely local complexity and trend curvature rate, are jointly constructed to synchronously characterize the amplitude fluctuation and trend mutation behavior within the local segments of the factor data affecting the depression risk of chronic disease patients, thereby providing a quantitative basis for subsequent feature weighting.

[0014] . Preferably, in S5, based on the local complexity and local trend curvature rate, a local control factor is introduced to perform weighted fusion of the local segment features, which specifically includes the following steps: S51. Introduce a local control factor to nonlinearly fuse two types of local feature indicators, namely local complexity and local trend curvature rate. The specific mathematical model is as follows: ; In the formula, is the th factor data, The local control factor of a local segment and are respectively the learnable complexity weight and the curvature weight is the learnable bias term; S52. Calculate the center corresponding to each local segment using a fixed equidistant strategy, and generate multi-scale path weights based on the distance between the control factor and the center of the local segment. The specific mathematical model is: ; ; In the formula, is the center of the th local segment of the th factor data, is the multi-scale weight corresponding to the th local segment of the th factor data, is the center of the th local segment of the th factor data, is the local control factor of the th local segment of the th factor data; S53. Use the multi-scale path weights to perform weighted fusion of the local segments to obtain the overall feature sequence of the final th factor data. The specific mathematical model is: ; In the formula, is the overall feature sequence of the final th factor data obtained after weighted fusion of the local segments.

[0015] Preferably, a local control factor is constructed based on the local complexity and the local trend curvature rate to achieve adaptive modeling of the importance of different local segments. Multi-scale path weights are generated according to the distance between the local control factor and the center of each local segment, enabling the model to adaptively allocate different importance according to the characteristics of each local segment, so as to perform weighted fusion at the local segment granularity, making the model more adaptable and stable when facing the non-linear and non-stationary sequence characteristics of the factor data affecting the depression risk of chronic disease patients, and enhancing the model's ability to distinguish key change regions.

[0016] Preferably, in S6, a fully connected neural network is used to calculate the predicted value of the depression risk of chronic disease patients for each factor data, and the average value of all predicted values is calculated to obtain the final prediction result. The specific mathematical model is: ; In the formula, is the final predicted value of the depression risk for chronic disease patients, is the fully connected neural network for the factor data of the depression risk of the number of types of factor data for the depression risk of chronic disease patients.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses non-linear transformation to convert the deviation amount sequence of various factor data relative to the median value into a high-dimensional feature sequence, effectively enhancing the perception ability of potential risk signals. Combining the adaptive switching factor and non-linear phase transformation, the dynamic activation and reconstruction of high-dimensional feature vectors are realized. Subsequently, the sliding window mechanism is used to extract the complexity and trend curvature rate of local segments, characterize the local change characteristics, and introduce a local control factor. According to the distance between the control factor and the segment center, multi-scale path weights are generated to achieve the soft selective fusion of key features. Finally, the fully connected neural network independently predicts the depression risk values of each factor data for chronic disease patients and averages them to obtain a more stable and accurate prediction result of the depression risk for chronic disease patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of the steps of a method for predicting the depression risk of chronic disease patients.

[0019] Figure 2 is a diagram of the adaptive switching evolution module.

[0020] Figure 3 is a diagram of the local change characterization module.

[0021] Figure 4 is a structural diagram of the weighted fusion of local segments.

[0022] Figure 5 is an effect diagram of the prediction of the depression risk of chronic disease patients. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The present invention proposes a method for predicting the depression risk of chronic disease patients. This method constructs a deviation amount based on the median value for various factor data, generates a high-dimensional feature sequence through non-linear mapping, calculates the tension value of the high-dimensional feature sequence and its overall mean, introduces an adaptive switching factor and combines non-linear phase transformation to achieve dynamic activation and regulation of the high-dimensional feature vector, and obtains a fused feature sequence. Subsequently, a sliding window mechanism is used to divide the fused feature sequence into local segments, calculates the local complexity and trend curvature rate, introduces a local control factor, and combines the distance of the center of the local segment to generate a multi-scale path weight to achieve soft selection fusion of the local segments. After obtaining the final feature representation of each factor data, the depression risk value of chronic disease patients for each factor data is predicted respectively through a fully connected neural network, and the average value is calculated to obtain the final prediction result, thereby improving the stability and accuracy of the depression risk prediction of chronic disease patients. The technical solutions in the embodiments of the present invention will be described in detail and completely below, specifically including the following steps, as Figure 1 shown.

[0024] S1. Collect factor data affecting the depression risk of chronic disease patients and construct a depression risk prediction data set for chronic disease patients.

[0025] Further, in S1, collect factor data affecting the depression risk of chronic disease patients for 80 consecutive days, including blood glucose level, blood pressure, heart rate, exercise frequency, anxiety assessment value, stress assessment value, disease type, drug use record, use resampling and interpolation techniques to complement missing data to ensure data continuity, construct a depression risk prediction data set for chronic disease patients, divide the data set into a training set and a validation set in a ratio of 8:2, and use the data of the previous 7 days at each time step as the model input.

[0026] S2. The factor data constitutes time series data. Obtain the median value of each factor time series data, and calculate the deviation amount sequence between each factor time series data and the corresponding median value, and map the deviation amount sequence to a high-dimensional feature sequence through non-linear transformation.

[0027] Further, in S2, input the time series data constituted by the factor data of the th type of depression risk of chronic disease patients where is the monitoring value of the th type of factor data of depression risk of chronic disease patients at the time step, and calculate the deviation amount sequence between the Wherein, is the deviation sequence of the th factor data from the corresponding median value, is the sensitivity hyperparameter of the th factor data, and the initial value is set to 25 during the implementation process, is the median value of the th factor data; The deviation sequence is mapped to a high-dimensional feature sequence by using a non-linear transformation, and the specific mathematical model is: ; Wherein, is the high-dimensional feature sequence of the th factor data, is the dimension of the high-dimensional feature vector, is the learnable parameter matrix, is the learnable bias term.

[0028] S3. Calculate the tension value of the high-dimensional feature sequence and the overall tension mean value, introduce an adaptive switching factor, combine with non-linear phase transformation, construct an adaptive switching evolution module, fuse the high-dimensional feature sequence and the local tension change, and obtain a fused feature sequence.

[0029] Furthermore, in the said S3, an adaptive switching evolution module is constructed. As Figure 2 shown, an adaptive switching evolution module is constructed to update the overall feature of the high-dimensional feature sequence , including the following steps: S31. Use the second-order and fourth-order tension control hyperparameters to jointly weight and integrate the element-wise square and fourth-power results of the high-dimensional feature sequence of the th factor data respectively, to obtain the tension value sequence of the th factor data. Further, normalize and average the tension values of the th factor data at each time step within the full sequence range to obtain the overall tension mean value of the th factor data. The specific mathematical model is: ; ; Wherein, is the tension value sequence of the th factor data, is the second-order tension control hyperparameter, and the initial value is set to 0.5 during the implementation process, is the fourth-order tension control hyperparameter, and the initial value is set to 0.05 during the implementation process, is 's element-wise square, is the element - by - element fourth power of the overall tension mean value of the data of the - th factor, is the tension value at the - th time step in S32. Calculate the absolute difference between the tension value and the overall tension mean value, and construct an adaptive switching factor in combination with the normalization function to adjust the activation degree of the tension value. The specific mathematical model is: ; In the formula, is the adaptive switching factor of the data of the - th factor, is the absolute - value operation, is the hyper - parameter of the stability threshold, and the initial value is set to 0.1 during the implementation process; S33. Use the adaptive switching factor as the adjustment weight, and combine it with the non - linear phase change to fuse the high - dimensional feature sequence and the local tension change to obtain the fused feature sequence after fusion. The specific mathematical model is: ; In the formula, is the fused feature sequence of the data of the - th factor after fusion, where is the - th factor data's feature representation at the - th time step.

[0030] S4. Adopt the sliding - window mechanism to divide the fused feature sequence into multiple consecutive local segments, construct a local - change characterization module, and calculate the local complexity and local trend curvature rate of each local segment respectively.

[0031] Furthermore, in S4, the time - series data is divided into multiple consecutive local segments by using the sliding - window mechanism, and the local complexity and local trend curvature rate are calculated based on each segment to characterize the local - change characteristics. As shown in Figure 3 it includes the following steps: S41. Use the sliding window to divide the fused feature sequence of the data of the - th factor into local segments. The size of the sliding window is , and the initial value is set to 3 during the implementation process. The sliding step is , and the initial value is set to 1 during the implementation process. A total of local segments are generated. Denote the - th local segment as where is and ; S42. Construct the local complexity, calculate the local mean in the feature direction for each local segment, calculate the square of the difference between the fused feature value and the corresponding mean within the local segment, and normalize it within the time step and feature dimension range to obtain the local complexity. The specific mathematical model is as follows: ; ; In the formula, is the th factor data th local segment feature mean, is the th factor data th local segment local complexity, is the th factor data th time step feature value; S43. Calculate the discrete second-order difference of the feature values of adjacent time steps within the local segment, and normalize it to obtain the local trend curvature rate. The specific mathematical model is as follows: ; In the formula, is the th factor data th local segment local trend curvature rate.

[0032] S5. According to the local complexity and the local trend curvature rate, introduce a local control factor, generate multi-scale path weights based on the distance between the local control factor and the center of each local segment, and fuse the local segments with weights to obtain the final feature representation of each factor data.

[0033] Figure 4 ; In the formula, is the th factor data th local segment local control factor,, are the learnable complexity weight and curvature rate weight respectively, is a learnable bias term; S52. Calculate the center corresponding to each local segment using a fixed equidistant strategy, and generate multi-scale path weights according to the distance between the control factor and the local segment center. The specific mathematical model is as follows: ; ; In the formula, is the center of the th local segment of the th factor data, is the multi-scale weight corresponding to the th local segment of the th factor data, is the center of the th local segment of the th factor data, is the local control factor of the th local segment of the th factor data; S53. Use the multi-scale path weights to perform weighted fusion of the local segments to obtain the overall feature sequence of the final th factor data. The specific mathematical model is as follows: ; In the formula, is the overall feature sequence of the final th factor data obtained after weighted fusion of the local segments.

[0034] S6. Use a fully connected neural network to calculate the final feature representation of each factor data, obtain the depression risk prediction value of chronic disease patients, and add all the prediction values and take the average as the final depression risk prediction value of chronic disease patients.

[0035] Furthermore, in S6, use a fully connected neural network to calculate the depression risk prediction value of chronic disease patients for each factor data, and calculate the average of all the prediction values to obtain the final prediction result. The specific mathematical model is as follows: ; In the formula, is the final depression risk prediction value of chronic disease patients, is the fully connected neural network of the th factor data of the depression risk of chronic disease patients, is the number of types of factor data of the depression risk of chronic disease patients.

[0036] Furthermore, the model is implemented based on the Python 3.8 programming language and the PyTorch framework, runs in the CUDA 11.3 environment, and uses an NVIDIA 3090 24GB GPU during the training process. To improve the training efficiency and stability of the model, the initial learning rate is , the batch size is , and the optimizer is used to iteratively update the parameters. The loss function is loss function, and a linear learning rate decay strategy is adopted.

[0037] Furthermore, the prediction effect of this method is as shown in Figure 5 . The ordinate represents the depression risk value (%) of chronic disease patients, and the abscissa represents time (days). The gray dotted line in the figure represents the actual evaluation value of the depression risk of chronic disease patients, and the black solid line is the predicted value of the depression risk of chronic disease patients generated by this method. It can be seen from Figure 5 that the two curves maintain a high degree of consistency in both the overall trend and local fluctuations, and the fitting degree between the prediction result and the actual evaluation value is relatively high, indicating that the proposed method has high accuracy and stability in the task of predicting the depression risk of chronic disease patients.

[0038] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the creative concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for predicting the depression risk of chronic disease patients, characterized in that, It includes the following steps: S1. Collect the factor data affecting the depression risk of chronic disease patients, and construct a prediction dataset for the depression risk of chronic disease patients; S2. The factor data constitutes time series data. Obtain the median value of each factor time series data, and calculate the deviation sequence between each factor time series data and the corresponding median value. Map the deviation sequence to a high-dimensional feature sequence through non-linear transformation; S3. Calculate the tension value and the overall tension mean of the high-dimensional feature sequence, introduce an adaptive switching factor, combine non-linear phase transformation, construct an adaptive switching evolution module, fuse the high-dimensional feature sequence and local tension changes, and obtain a fused feature sequence; S4. Use a sliding window mechanism to divide the fused feature sequence into multiple consecutive local segments, construct a local change characterization module, and calculate the local complexity and local trend curvature rate of each local segment respectively; S5. According to the local complexity and local trend curvature rate, introduce a local control factor, generate multi-scale path weights based on the distance between the local control factor and the center of each local segment, and fuse the local segments with weights to obtain the final feature representation of each factor data; S6. Use a fully connected neural network to calculate the final feature representation of each factor data, obtain the prediction value of the depression risk of chronic disease patients, add all the prediction values and take the average as the final prediction value of the depression risk of chronic disease patients.

2. The method for predicting the depression risk of chronic disease patients according to claim 1, characterized in that, In S1, collect the factor data affecting the depression risk of chronic disease patients, including blood glucose level, blood pressure, heart rate, exercise frequency, anxiety assessment value, stress assessment value, disease type, drug use record, and use resampling and interpolation techniques to complete the missing data, and construct a prediction dataset for the depression risk of chronic disease patients.

3. The method for predicting the depression risk of chronic disease patients according to claim 2, characterized in that, Input the time series data of the factors affecting the depression risk of chronic disease patients , the length of the time series is , obtain the median value, calculate the deviation amount sequence between the th factor data and the median value. The specific mathematical model is as follows: ； In the formula, is the deviation amount sequence of the th factor data from the corresponding median value, is the sensitivity hyperparameter of the th factor data, is the median value of the th factor data; Map the deviation sequence through non-linear transformation to a high-dimensional feature sequence. The specific mathematical model is as follows: ； In the formula, is the high-dimensional feature sequence of the th factor data, and the dimension of the high-dimensional feature is , is a learnable parameter matrix, is a learnable bias term.

4. The method for predicting the depression risk of chronic disease patients according to claim 3, wherein, In S3, an adaptive switching and evolution module is constructed to update the overall features of the high-dimensional feature sequence, including the following steps: ​ S31. Using the second-order and fourth-order tension control hyperparameters, respectively perform joint weighted integration on the element-wise squared and fourth-power results of the high-dimensional feature sequence of the th factor data to obtain the tension value sequence of the th factor data . Further, perform normalized averaging on the tension values at each time step of the th factor data within the entire sequence range to obtain the overall tension mean of the th factor data ; S32. Calculate the absolute difference between the tension value and the overall tension mean, and construct an adaptive switching factor in combination with a normalization function. The specific mathematical model is: ； In the formula, is the adaptive switching factor of the th factor data, is the absolute value operation, is the hyperparameter of the stability threshold; S33. Using the adaptive switching factor as the adjustment weight, combining the non-linear phase change, and fusing the high-dimensional feature sequence and the local tension change to obtain the fused feature sequence after fusion. The specific mathematical model is as follows: ； In the formula, is the fusion feature sequence of the th factor data after fusion.

5. The method for predicting the depression risk of chronic disease patients according to claim 4, wherein, In S4, construct a local change characterization module, including the following steps: S41. Use a sliding window to divide the fusion feature sequence of the th factor data into local segments. The size of the sliding window is , and the sliding step is . , . A total of local segments are generated. Denote the th local segment as ; S42. Calculate the local mean in the feature direction for each local segment. The th factor data, the th local segment, and the feature mean is . Calculate the square of the difference between the fused feature value and the corresponding mean within the local segment, and normalize it within the time step and feature dimension ranges to obtain the local complexity. The specific mathematical model is as follows: ； In the formula, is the local complexity of the th factor data for the th local segment, is the th eigenvalue at the th time step in the th local segment of the th factor data, and is the dimension of the high-dimensional feature; S43. Calculate the discrete second-order difference of the feature values of adjacent time steps within the local segment, and obtain the local trend curvature rate after normalization. The specific mathematical model is: ； In the formula, is the local trend curvature of the th factor data for the th local segment.

6. A method for predicting the depression risk of chronic disease patients according to claim 5, characterized in that, In S5, based on the local complexity and local trend curvature rate, introduce a local control factor to fuse the features of the local segments with weights. Specifically It includes the following steps: S51. Introduce a local control factor, and non-linearly fuse two types of local feature indicators, namely local complexity and local trend curvature rate. The specific mathematical model is: ； In the formula, is the local control factor of the th factor data for the th local segment, , are the learnable complexity weight and curvature weight respectively, is the learnable bias term; S52. Use a fixed equidistant strategy to calculate the center corresponding to each local segment, and generate multi-scale path weights according to the distance between the control factor and the center of the local segment. The specific mathematical model is: ； ； In the formula, is the center of the th local segment of the th factor data, is the multi-scale weight corresponding to the th local segment of the th factor data, is the center of the th local segment of the th factor data, is the local control factor of the th local segment of the th factor data; S53. Using multi-scale path weights to perform weighted fusion on all local segments of the -th factor data to obtain the overall feature sequence of the final -th factor data .

7. A method for predicting the depression risk of chronic disease patients according to claim 6, characterized in that, In S6, use a fully connected neural network to calculate the predicted depression risk values of chronic disease patients for each factor data, and calculate the average value of all predicted values to obtain the final predicted depression risk value of chronic disease patients .

Images