A method, apparatus, and medium for predicting risk of depression

Abstract

The application discloses a depression risk prediction method and device and a medium, and relates to the technical field of medical data processing. The method comprises the following steps: constructing a corresponding marginal value function for each feature index type to calculate the marginal value of the feature index; configuring a feature index weight for each feature index, and calculating a comprehensive value that contributes to depression risk prediction according to the feature index weight and the marginal value; combining multiple feature indexes to obtain multiple feature index pairs, and configuring an interaction coefficient for the positive and negative results of the two marginal values of each feature index pair; calculating the interaction effect coefficient of each feature index pair according to the two marginal values and the interaction coefficient of each feature index pair, and combining the interaction effect coefficient of each feature index pair to calculate the average value of the interaction effects of the multiple feature index pairs; calculating the risk prediction value of the to-be-predicted object according to the comprehensive value and the average value, and determining the depression prediction result of the to-be-predicted object based on the size relationship between the risk prediction value and a risk threshold value.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and more specifically, to a method, device, and medium for predicting the risk of depression. Background Technology

[0002] Depression is a mental disorder characterized by high incidence and high disability rates, and its early identification and accurate prediction are important issues facing the global public health field. Currently, the diagnosis of depression mainly relies on two standards: the Predictive and Statistical Manual of Mental Disorders (PSM) and the International Classification of Diseases (ICD), combined with standardized assessment tools such as the Hamilton Depression Rating Scale and the Beck Depression Rating Scale, and is made through structured or semi-structured interviews. While this predictive paradigm has been used in practice for a long time, it has the following limitations: the assessment process is time-consuming, making it difficult to meet the efficiency requirements of large-scale early screening; the predictive results rely heavily on patient self-reporting and individual clinician experience, resulting in fluctuations in the objectivity, repeatability, and consistency among different assessors, and it is highly dependent on professionals.

[0003] Currently, intelligent technologies, represented by machine learning, have been applied in the field of intelligent assisted prediction and identification of depression due to their capabilities in nonlinear pattern recognition and high-dimensional data processing. The purpose of machine learning technology is to automatically learn the complex mapping relationship between clinical characteristic indicators and predicted outcomes from large-scale historical case data, thereby constructing a computational model for early identification and risk prediction of depression, in order to improve the efficiency and objectivity of early prediction.

[0004] Although machine learning-based intelligent recognition models demonstrate good predictive performance, they still face the following challenges in practical clinical applications: 1. Currently widely used black-box models (such as deep neural networks) have internal representations and reasoning processes that are difficult to understand intuitively. They typically only output predicted labels or risk probabilities, failing to provide decision-making basis based on clinical reasoning logic and pathophysiological mechanisms. Clinicians cannot clearly understand the reasons for the model's judgments or the contribution weights of each feature indicator, which directly limits the model's clinical acceptance. Furthermore, at the interaction level, purely data-driven models cannot incorporate physicians' practical experience into the training and optimization process.

[0005] 2. Existing multimodal fusion methods often employ data-driven feature index concatenation or attention-weighted mechanisms. Their weight allocation primarily relies on the statistical correlation strength between feature indicators and outcome variables, rather than the actual importance of the feature indicators in clinical prediction. These approaches fail to incorporate prior associations between the predictive pathways of depression, symptom dimensions (such as anhedonia, sleep disorders, and cognitive symptoms), and various modalities as constraints. This makes it difficult to assign differentiated, clinically intuitive weights to feature indicators of different modalities and types. Furthermore, these methods are highly sensitive to redundant feature indicators and data noise, easily generating statistically significant but clinically weak associations, making it difficult to trace and interpret the decision-making path.

[0006] 3. Most studies on auxiliary prediction of depression rely primarily on neuroimaging data such as functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI). This type of data is costly to collect, requires stringent equipment, and relies on professional interpretation, making it unsuitable as a routine screening method for large-scale general populations or primary healthcare institutions. Meanwhile, research on depression risk modeling using readily available and widely accessible routine clinical data such as electronic health records and physical examination reports is relatively insufficient, making it difficult to integrate with routine clinical screening scenarios for depression. Summary of the Invention

[0007] The purpose of this invention is to provide a method, device, and medium for predicting the risk of depression, in order to solve the above-mentioned technical problems.

[0008] A first aspect of the present invention provides a method for predicting the risk of depression, the method comprising: Obtain multiple feature metrics of an object to be predicted; Multiple feature indicators are divided into multiple feature indicator classes, and a corresponding marginal value function is constructed for each feature indicator class to calculate the marginal value of the feature indicators included in each feature indicator class. Assign a weight to each feature indicator, and calculate the combined value of the contribution of multiple feature indicators to the prediction of depression risk based on the feature indicator weights and marginal values; Multiple feature indicators are combined in pairs to obtain multiple feature indicator pairs, and an interaction coefficient is configured for the positive and negative results of the two marginal values ​​of each feature indicator pair. Based on the two marginal values ​​and the interaction coefficient of each feature index pair, calculate the interaction effect coefficient of each feature index pair, and combine the interaction effect coefficient of each feature index pair to calculate the average value of the interaction effect of multiple feature index pairs. The risk prediction value of the subject to be predicted is calculated based on the comprehensive value and the average value. The prediction result of depression of the subject to be predicted is determined based on the relationship between the risk prediction value and the risk threshold.

[0009] A second aspect of the present invention provides a computer device comprising a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of a method for predicting the risk of depression as provided in the first aspect. A third aspect of the present invention provides a computer-readable storage medium comprising a computer program that, when executed by one or more processors, implements a method for predicting the risk of depression as provided in the first aspect.

[0010] Compared with the prior art, the present invention has the following beneficial effects: This invention differs from the ex-post interpretation of traditional black-box models (such as feature importance ranking, SHAP values, etc.) by embedding traceable clinical logic throughout the decision-making process and output results. Specifically, this invention uses multi-attribute value theory as its modeling framework, transforming the reasoning process that comprehensively considers multi-dimensional symptoms, test indicators, and vital signs in prediction into an explicit mathematical function: the independent contribution of each feature indicator to the prediction is represented by the product of its weight and marginal value, and the joint effect between feature indicators is characterized by the product of the interaction coefficient and the absolute value of the marginal value. All parameters (weights, interaction coefficients) have clear clinical meanings—feature indicator weights reflect the importance of the feature indicator, and interaction coefficients reflect the synergistic or antagonistic effects under specific value combinations. The construction of marginal values ​​is based on clinical reference ranges, mapping the feature indicator data to interpretable deviations, thus making each step of reasoning traceable to specific clinical evidence. This invention not only outputs the risk prediction probability but also fully demonstrates the positive and negative contributions of each feature indicator and pairwise combinations to the final decision, achieving full transparency of the decision-making path. Because the prediction process is highly consistent with clinical prediction logic, physicians can directly understand the judgment basis in the prediction process without relying on additional interpretation tools. This method uses readily available data such as electronic health records and physical examination data as input, and has good clinical applicability. It can assist doctors in early screening, risk assessment and predictive decision-making for depression. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating a method for predicting the risk of depression provided in an embodiment of the present invention; Figure 2 This is a block diagram illustrating the principle of depression risk prediction provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0013] It should be noted that the terms "comprising" or "may include" used in the various embodiments of this application indicate the presence of the claimed function, operation, or element, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms "comprising," "having," and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more features, numbers, steps, operations, elements, components, or combinations of the foregoing.

[0014] In various embodiments of this application, the expression "or" or "at least one of B and / or C" includes any combination or all combinations of the words listed simultaneously. For example, the expression "B or C" or "at least one of B and / or C" may include B, may include C, or may include both B and C.

[0015] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0016] Existing technologies for early identification and prediction of depression based on artificial intelligence mainly suffer from the following problems: First, multimodal clinical data lacks a unified structured representation mechanism. Symptom characteristic indicators, scale scores, biochemical indicators, and vital signs differ significantly in their value formats, dimensional ranges, and clinical semantics. Existing methods often employ direct concatenation or simple normalization, failing to transform heterogeneous characteristic indicators into standardized representations with unified predictive meanings, thus limiting the realization of subsequent interpretable modeling. To address this issue, this invention establishes a five-category marginal value function system based on prior clinical knowledge, uniformly mapping multidimensional characteristic indicators to a bounded predictive contribution space, thereby achieving structured representation and marginal value quantification of characteristic indicators.

[0017] Second, existing predictive models lack transparency and traceability in their decision-making processes. Data-driven models, such as deep neural networks, only output predictive labels or risk probability values, failing to provide decision-making basis consistent with clinical reasoning logic. Clinicians find it difficult to ascertain the specific contribution of each feature indicator to the predictive conclusion and its weight allocation. Furthermore, existing interpretable methods often employ linear weighted structures, neglecting interaction terms between feature indicators, making it difficult to reflect the synergistic or antagonistic effects of multiple co-occurring symptoms and multiple indicator linkages on predictive results in clinical practice. To address these issues, this invention constructs an interpretable value function based on multi-attribute value theory. It uses explicit feature indicator weights and interaction coefficients to characterize the predictive reasoning process and introduces an interaction effect modeling mechanism based on marginal value symbol combinations. This allows each decision step to be traced back to the contribution of specific parameters, achieving a refined expression of the joint predictive effect among feature indicators.

[0018] Third, existing interpretable methods have shortcomings in parameter learning mechanisms regarding computational efficiency and generalization ability. Traditional mathematical programming models have low solution efficiency in high-dimensional clinical data scenarios and lack generalization control mechanisms such as regularization and cross-validation, making them prone to overfitting. To address this issue, this invention employs a parameter learning strategy that combines a differentiable loss function with an optimization algorithm, and introduces regularization and hierarchical cross-validation mechanisms to improve solution efficiency and generalization performance while maintaining structural transparency.

[0019] The depression risk prediction method provided by this invention will be described in detail below with reference to specific implementation schemes, such as... Figure 1 As shown in the figure, this invention provides a method for predicting the risk of depression, the method comprising the following steps: S101, obtain multiple feature indicators of an object to be predicted.

[0020] Specifically, before obtaining multiple characteristic indicators, it is necessary to collect multimodal data of the subjects to be predicted. This multimodal data comes from electronic health records, physical examination files, and other relevant clinical data systems of patients diagnosed with depression, healthy controls, and high-risk groups, but is not limited to these sources. Specifically, diagnosed patients are those diagnosed with depression by psychiatrists according to the criteria of the *Manual of Predictive and Statistical Analysis of Mental Disorders (Fifth Edition)* or the *International Classification of Diseases (Eleventh Edition)*. Healthy controls are healthy individuals collected concurrently without depression prediction. High-risk groups are individuals identified through large-scale population screening data such as student mental health surveys and employee physical examinations, who were initially screened using psychological scales and indicated a high risk of depression but have not yet met the clinical prediction criteria.

[0021] The collected multimodal data may include: sociodemographic data (such as age, gender, marital status, education level, etc.), symptom assessment data (such as the total score and scores of each item of the Patient Health Questionnaire 9, the total score and scores of each item of the Hamilton Depression Rating Scale, etc.), vital sign data (such as systolic blood pressure, diastolic blood pressure, heart rate, body temperature, respiratory rate, etc.), blood routine data (such as white blood cell count, red blood cell count, hemoglobin concentration, platelet count, etc.), and biochemical test data (such as fasting blood glucose, total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, alanine aminotransferase, aspartate aminotransferase, serum creatinine, blood urea nitrogen, etc.).

[0022] The various characteristic indicators (such as body temperature and heart rate) of the modal data described above contribute differently to the auxiliary prediction of depression. It is necessary to screen out the characteristic indicators that show statistically significant differences between the confirmed case group and the control group through a significance test, and use these as key input variables. Specifically, for continuous numerical variables (such as blood pressure, test indicators, and scale scores), an independent samples t-test is used to compare the mean differences between the depression group and the control group; for categorical variables (such as gender), a chi-square test is used to compare the distribution differences between groups.

[0023] Screening criteria: Calculate the p-value for each feature indicator. If p < 0.05, the feature indicator is considered to have a statistically significant difference between the two groups and can be retained as a candidate important feature indicator for predicting depression. After completing the feature indicator screening, a clear classification label needs to be assigned to each data record for learning. The assignment rule is as follows: the label for patients diagnosed with depression is "1", as a positive sample. The label for control group subjects (healthy controls) is "-1", as a negative sample. The labels and feature indicator values ​​together constitute the training dataset. Each sample contains a set of feature indicator values ​​and their corresponding predicted labels. Hereinafter, both patients and samples will be referred to as "objects".

[0024] S102, divide multiple feature indicators into multiple feature indicator classes, construct a corresponding marginal value function for each feature indicator class, and calculate the marginal value of the feature indicators included in each feature indicator class.

[0025] In this embodiment, addressing the problem of inconsistent processing of different modalities and types of clinical characteristic indicators within a single model in existing solutions, the present invention first aims to solve the problem of transforming multidimensional clinical data such as symptom information, vital signs, and biochemical test indicators into a structured representation with unified predictive meaning. Existing methods often directly use raw numerical values ​​or simple normalization methods as model inputs. While preserving the data format, this approach fails to fully reflect the actual clinical significance of various characteristic indicators in predicting depression and is not conducive to the subsequent establishment of interpretable decision-making models.

[0026] To address this issue, this invention divides the collected multidimensional clinical modal data into five categories and constructs marginal value functions for each category, mapping the original feature index values ​​uniformly to the [-1,1] interval, thereby achieving a structured representation of clinical semantics.

[0027] (1) Symptom-related indicators, such as depressed mood, loss of interest, sleep disturbances, suicidal ideation, etc. These indicators can be represented by binary or ordinal scores. For core symptoms, the presence of a symptom is a positive value, with higher scores resulting in larger values, up to a maximum of 1, while the absence of a symptom is a -1 value. For secondary symptoms, the presence of a symptom is a positive value, while the absence of a symptom is a 0 value. This is used to characterize their positive or neutral / negative contribution to prediction.

[0028] (2) Nominal characteristic indicators, such as gender, marital status, and education level. Based on prior epidemiological knowledge or expert experience, each value is mapped to... Interval: Positive values ​​indicate an increased risk of depression, negative values ​​indicate a protective effect, and absolute values ​​represent the intensity of the risk. Specifically, a logarithmic mapping to base 5 is used, truncated to the boundary.

[0029] (3) Upper boundary clinical test characteristic indicators, such as C-reactive protein and some inflammatory factors. The clinical significance of these characteristic indicators is mainly reflected in the direction of elevation. Taking the midpoint of the normal range as the dividing point: when it is below the midpoint, the marginal value is 0 (no contribution), and when it is above the midpoint, the marginal value increases exponentially with the degree of deviation until it reaches the saturation threshold, where it saturates to 1.

[0030] (4) Lower boundary clinical laboratory characteristic indicators, such as high-density lipoprotein. The predictive value of these characteristic indicators is mainly reflected in the low direction. Set a reference point (the mean of the normal lower limit and the abnormal low value threshold). When the value is below the reference point, the marginal value decreases to -1 according to an exponential function, and when the value is above the reference point, the marginal value is 0.

[0031] (5) Interval clinical laboratory characteristic indicators, such as blood glucose, uric acid, white blood cell count, etc. The midpoint of the normal range is used as the reference point, and an exponential function with a uniform scale D is used: the marginal value is positive when it is above the midpoint and negative when it is below the midpoint. The further the deviation, the closer the absolute value is to 1; when the value exceeds the clinical saturation threshold, it is truncated to [-1,1].

[0032] It should be noted that merging or further subdividing the five categories of characteristic indicators while retaining the core logic of marginal value mapping based on clinical priors or epidemiological information constitutes an equivalent substitution and falls within the scope of protection of this invention.

[0033] By employing the five types of marginal value functions described above, this invention transforms clinical characteristic indicators of different dimensions and types into comparable marginal values, providing standardized characteristic indicators for subsequent decision-making. It is important to understand that the marginal value function used for the characteristic indicators of each modality of data is different. The following will provide a detailed explanation, in conjunction with specific embodiments, of how the marginal value function is used to calculate the marginal value of each characteristic indicator as described in this embodiment.

[0034] To unify the transformation of feature indicators with different dimensions and clinical meanings into comparable predictive contribution measures, this invention introduces the concept of a marginal value function. For the original value of the i-th object on the j-th feature indicator... Its marginal value function is defined as... Mapped to The mathematical function over an interval, the output of which is called the marginal value of that feature index. The absolute value of this marginal value represents the degree of abnormal deviation of the current value relative to the clinical reference point. The larger the absolute value, the further it deviates from the normal state, and the greater its potential contribution to prediction; the sign indicates the direction of deviation. Specifically: This indicates that the current value has reached an extreme level in the positive deviation direction (such as extremely high indicators or severe symptoms). This indicates that the current value has reached an extreme level in the negative deviation direction (such as extremely low indicators or lack of core symptoms). This indicates that the current value is close to the clinical reference point and there is no significant abnormal deviation.

[0035] The rationale for defining the marginal value function is as follows: Clinical prediction requires a comprehensive consideration of the presence or absence of symptoms and the direction and degree of deviation of test indicators from the normal range. In the original modal data, symptoms are discrete values, while blood pressure, blood sugar, etc., are continuous values, and their normal reference ranges vary, making direct comparison and fusion impossible. The marginal value function, by simulating the logic of clinicians judging predictive contribution based on the degree of deviation from normal, transforms all feature indicators into anomaly deviation measures on the same scale, providing standardized input for subsequent decision-making models. The absolute value of the marginal value represents the degree of abnormal deviation of the current value relative to the clinical reference point, and the sign indicates the direction of deviation. For example, values ​​above the reference point are positive, and values ​​below the reference point are negative; the presence of symptoms is positive, and the absence of symptoms is negative.

[0036] Explanation of the rationality of the marginal value function: This mapping method has clear traceability, and each marginal value can be directly mapped to a clinical interpretation: For continuous test indicators, the magnitude and direction of deviation from the normal range are mapped to the corresponding degree and direction of abnormal deviation (positive value is positive deviation, negative value is negative deviation); for symptom feature indicators, the presence of core symptoms is mapped to +1 (extreme positive deviation), and the absence is mapped to -1 (extreme negative deviation); the presence of secondary symptoms is mapped to +1, and the absence is mapped to 0 (no deviation); at the same time, the boundary values ​​+1 and -1 respectively represent reaching extreme abnormality in the positive and negative directions, and 0 represents no abnormal deviation. This method avoids black-box normalization. It should be noted that the marginal value itself does not presuppose the direction of support for the risk of depression; the specific direction of support is determined by the positive or negative of the feature indicator weights. This design achieves the separation of prior knowledge-guided anomaly quantification and data-driven relational learning, enhancing the interpretability and flexibility of prediction.

[0037] The marginal value function is defined for each type of characteristic indicator, and the design rationale, rational basis, and typical examples are explained.

[0038] 1) Symptom-based indicators: Symptom-based indicators for depression can use binary variables to represent the presence or absence of symptoms, or multi-level rating scales to reflect the severity of symptoms (such as 0-3 points for each item in the PHQ-9). Without loss of generality, the presence or increased severity of symptoms tends to support the prediction of depression, while the absence of symptoms helps to exclude prediction or makes a neutral contribution.

[0039] Let the original value of symptom characteristic index j be... Its value range is ,in The maximum possible score for this symptom (when it is a binary variable) When using multi-level rating (The highest score). This symptom was marked as a core or secondary symptom according to clinical predictive criteria. This indicates that the symptoms are completely absent; This indicates the presence of symptoms (including varying degrees). Marginal value The definition is as follows: .

[0040] The presence of typical depressive symptoms is a necessary condition for prediction, so it is assigned a maximum positive value of 1 when it is present. For core symptoms (depressed mood, loss of interest), their absence strongly supports prediction, so they are assigned -1; for secondary symptoms (such as decreased appetite), their absence only indicates a normal range, so they are assigned 0. This design directly corresponds to the principle in clinical consultation that prediction is considered only when core symptoms are present, and excluded when core symptoms are absent. The classification of core and secondary symptoms is based on the prediction criteria of the *Manual of Predictive and Statistical Analysis of Mental Disorders (Fifth Edition)*. For multi-level rating data, the above formula linearly maps severity to... (Secondary symptoms) or (Core symptoms, with -1 as an exception for 0 points) This approach preserves gradient information while ensuring the clinical rationality that the presence of core symptoms at any level does not have a negative contribution.

[0041] For example, core symptoms (anhedonia, PHQ-9 item 1, 0-3): score 0, v = -1; score 1, v = 1 / 3 ≈ 0.333; score 2, v = 2 / 3 ≈ 0.667; score 3, v = 1. Secondary symptoms (sleep disturbance, binary variable): missing (0), v = 0; present (1), v = 1. Secondary symptoms (altered appetite, PHQ-9 item 5, 0-3): score 0, v = 0; score 1, v = 0.333; score 2, v = 0.667; score 3, v = 1.

[0042] 2) Nominal characteristic indicators: Nominal characteristic indicators include binary variables (such as gender) and multi-category variables (such as marital status, education level, occupation, etc.). These characteristic indicators do not have clinically meaningful "normal / abnormal" or "present / missing" values. Their contribution to the risk of depression is mainly based on prior epidemiological knowledge (such as relative hazard ratio, prevalence differences) or clinical expert experience.

[0043] Let the set of values ​​for the nominal characteristic index X be . .make For taking values The relative risk of depression relative to a reference baseline (such as the general population or the lowest risk level) is the ratio of the incidence of depression in the target population to the incidence in the reference population (in case-control studies, the odds ratio can be used as an approximation). Marginal values ​​of Category II characteristic indicators. Defined as: .in, As a monotonic mapping function, it compresses the range of relative risk to The interval. Positive and negative values ​​indicate the direction of the influence of that value on the risk of depression: positive values ​​indicate an increased risk, negative values ​​indicate a decreased risk, and the absolute value reflects the strength of the influence. In a specific embodiment of this invention, f is taken as a logarithmic function to the base 5, with upper and lower bounds set for truncation: ,in, Indicated by The base is the logarithm. In practical applications, the base can be flexibly selected based on the empirical distribution range of relative risk. The value of is taken as (e.g., 10, 5, e). This mapping has the following properties: hour (No impact); hour (Strong risk factors); hour (Strong protective factors); Follow Monotonically increasing and satisfying symmetry: .

[0044] The individual values ​​of nominal characteristic indicators are not predictive "symptoms" themselves, but rather participate in decision-making as background risk factors. Through logarithmic mapping based on relative risk, prior epidemiological knowledge can be quantitatively embedded into the model, ensuring that the model conforms to population-level risk distribution patterns from the initial decision-making stage. A logarithmic function with base 0 makes exist arrive The marginal value changes smoothly and symmetrically between these points, and is taken at the boundary. To avoid extreme values ​​excessively dominating subsequent weighted calculations, this mapping method maintains a scale consistent with the marginal values ​​of continuous indicators and symptom characteristic indicators. This ensures that different types of characteristic indicators can be integrated and compared within a unified framework. During the mapping process, the positive and negative signs and the intensity are based on clear epidemiological or clinical evidence, avoiding the interpretive loss caused by black-box normalization.

[0045] For example, regarding gender: the prevalence of depression in women is approximately 1.5 to 2 times that in men. The value is 5, with males as the reference ( , ),female ,but This indicates that women contribute a moderate level of positive risk. Marital status: Divorced / widowed individuals have a relatively higher risk. ), (Low risk); the risk is relatively low for married individuals. ), (Weak protection); Unmarried individuals face moderate risk. ), (No impact). Education level: Low education level (junior high school or below) , (Low risk); Higher education (undergraduate and above) , (Weak protection); secondary education as a reference, , .

[0046] 3) Upper boundary clinical laboratory characteristic indicators: According to clinical guidelines, the normal range for these indicators in healthy individuals is... ,in These are the upper limits of the normal range, such as some inflammatory factors and enzyme indicators. The clinical significance of these indicators is usually manifested as follows: when the characteristic indicator value is below or close to the reference point, its contribution to disease prediction is weak; when the characteristic indicator value exceeds the reference point, its contribution to prediction gradually increases with the increase of the value.

[0047] Midpoint of normal range As a reference point, and set For clinically meaningful high-value saturation thresholds, for example, taking... (Normal upper limit) times, Greater than 1) or the 95th percentile of the sample (whichever is smaller, to balance clinical significance and data suitability). Marginal values ​​of Class III characteristic indicators. Defined as: ,in, Indicated by The base is 10. In practical applications, other bases (such as 10 or e) can also be used, and the choice can be made flexibly based on the empirical distribution range of relative risk. is the original test value of the i-th object on the j-th feature index; U is the upper limit of the normal range of the feature index; The midpoint of the normal range serves as the dividing point (when the characteristic index value is below this value, the contribution is considered to be 0). For high-value saturation threshold, we can take... ,in The 95th percentile of this feature in the sample (or the 99th percentile).

[0048] The marginal value changes according to the following pattern: when When: The characteristic indicator value is at a low level within the normal range, its contribution to disease prediction and differentiation is weak, and the marginal value is set to 0; when Time: The characteristic index value gradually transitions from a normal high value to an abnormally high value, and the marginal value increases exponentially with the degree of deviation from the reference point, and through... Limit the results to the interval [0,1]; when At that time, the marginal value saturates to 1.

[0049] The predictive significance of such indicators mainly stems from their elevated or abnormally high values. Once an indicator's value exceeds the midpoint of the normal range, the higher the value, the more pronounced the deviation from the normal state, and the greater its predictive value. Using an exponential function can reflect clinically relevant indicators where slight deviations contribute little, while significant deviations contribute rapidly. The selection of a high-value saturation threshold avoids oversensitivity to extreme values ​​while ensuring stable mapping across datasets: a fixed clinical multiple (e.g., twice the upper limit of normal) ensures universality, while high percentile truncation prevents abnormal expansion of the mapping range due to individual extreme values ​​in the dataset.

[0050] For example, taking C-reactive protein (CRP) as an example, the normal range is... The upper limit of the normal range It is 5, midpoint It is 2.5. Assume the high-value saturation threshold is... .

[0051] Object A: .because The marginal value is 0. This value is in the low range of the normal range and does not contribute to disease prediction.

[0052] Object B: .because Substitute into the calculation: This value is slightly higher than the midpoint, indicating a mild predictive contribution.

[0053] Object C: Substitute the values ​​into the calculation: This exceeds the normal upper limit, and the predicted contribution has increased significantly.

[0054] Object D: (Maximum value of the sample). Substitute into the calculation: It reaches the highest level of this feature index in the sample, with a marginal value of 1, thus forming the strongest predictive support.

[0055] 4) Lower boundary clinical laboratory characteristic indicators: According to clinical guidelines, the normal range for these indicators in healthy individuals is... ,in This refers to the lower limit of the normal range, such as high-density lipoprotein (HDL) and some immune indicators. Since these indicators do not have a clearly defined upper limit, the midpoint of the normal range cannot be directly defined. Therefore, reference points are constructed based on the lower limit of normal and the abnormally low value range. Values ​​below the reference point indicate a transition from normal to abnormally low, with a marginal value less than 0; values ​​above the reference point are considered within the normal range, with a marginal value set to 0.

[0056] set up This is a clinically meaningful low-value saturation threshold. For example, it could be taken as... (lower limit of normal) (times) or 5th percentile of the sample (whichever is greater, to balance clinical significance and data robustness).

[0057] Reference point Defined as: .

[0058] The marginal value function is: ,in, Indicated by The base is 10. In practical applications, other bases (such as 10 or e) can also be used, and the choice can be made flexibly based on the empirical distribution range of relative risk. Let L be the original value of the i-th object on the j-th feature index; L is the lower limit of the normal range of the feature index. As a low-value saturation threshold, it can be taken as follows: ,in The 5th percentile of this characteristic indicator in the sample (the 1st percentile can also be used or adjusted according to clinical knowledge); As a reference point, between Between L and L.

[0059] when When the indicator value is at a normal or relatively normal level, its contribution to disease prediction is weak, and the marginal value is set to 0; when When the indicator value is lower than the reference point, it indicates a transition from a normal state to an abnormally low state, and the lower the value, the closer the marginal value is to -1; when At that time, the marginal value saturates to -1.

[0060] For indicators like high-density lipoprotein (HDL), the general clinical understanding is that "low levels are harmful, while normal or high levels are beneficial." For indicators with a normal range from positive to infinity, their predictive value primarily lies in abnormally low values. If the indicator value is higher than the reference point, it is considered to be within the normal range or slightly above normal, offering no additional predictive contribution; if it is lower than the reference point, the greater the degree of low value, the stronger the abnormality, and the greater the marginal contribution. Exponential normalization can more reasonably describe the non-linear enhancement of the predictive contribution of abnormally low values. Using a fixed clinical proportion (e.g., 0.5 times the lower limit of normal) and the maximum value of the lowest percentile of the sample avoids the excessive influence of extreme outliers on the saturation threshold, ensuring stable mapping across datasets.

[0061] For example, taking high-density lipoprotein (HDL) as an example, the lower limit of the normal range Assuming a low-value saturation threshold The reference point is calculated as follows: .

[0062] Object A: .because The marginal value is 0. This value is considered to be at a normal to high level and does not contribute to prediction.

[0063] Object B: .because Substitute into the calculation: It has just touched the lower limit of normal and shows a slight predictive contribution.

[0064] Object C: Substitute into the calculation: It is significantly lower than the lower limit of normal, and the predicted contribution has increased significantly.

[0065] Object D: (Minimum value of the sample). Substitute into the calculation: When the sample reaches its lowest level, the marginal value saturates to -1, thus forming the strongest predictive support.

[0066] 5) Interval-based clinical laboratory characteristic indicators: According to clinical guidelines, the normal range for these indicators in healthy individuals is as follows: ,in and These represent the lower and upper limits of the normal range, respectively. These are the most common test indicators, such as blood pressure, blood glucose, uric acid, and white blood cell count. The midpoint of the normal range is used as a reference point, with deviations in different directions represented on either side. If the indicator value is higher than the reference point, the marginal value is positive; if the indicator value is lower than the reference point, the marginal value is negative. To avoid inconsistencies in scale caused by different value ranges on both sides, a uniformly scaled exponential marginal value function is used for normalization.

[0067] Let the midpoint of the normal range be... Define a low-value saturation threshold. and high saturation threshold . ,in Coefficients less than 1 (e.g.) ), This is the 5th percentile of this feature in the sample. ,in Coefficients greater than 1 (e.g.) ), It is the 95th percentile of this feature index in the sample. Adjustments can be made based on specific indicators and clinical knowledge; if clinical knowledge is insufficient, quantiles can be used directly without multiplying by a coefficient.

[0068] Scale parameters Take the maximum distance between the reference point and the saturation thresholds on both sides: .

[0069] The marginal value function is defined as: ,in, Indicated by The base is 10. In practical applications, other bases (such as 10 or e) can be used, and the choice can be flexible based on the empirical distribution range of relative risk.

[0070] when When the value is positive, the marginal value is higher, and the higher the value, the closer the marginal value is to +1; when When the value is negative, the marginal value is negative, and the lower the value, the closer the marginal value is to -1; when When the marginal value saturates to -1; when At that time, the marginal value saturates to +1; through and Truncate the marginal value to ensure the result always lies within the range of the marginal value. Within the range.

[0071] Indicators such as blood pressure, blood glucose, and white blood cell count have clearly defined normal ranges. Clinically, the further a value deviates from any range, the stronger its pathological nature; the midpoint of the range is ideal. For continuous indicators with bilateral normal ranges, the directions of deviation on either side of the reference point may correspond to different disease states or different predictive tendencies. Using the midpoint of the normal range as the dividing point maps values ​​below and above the midpoint to negative and positive marginal contributions, respectively. Simultaneously, normalization using a uniform scale D avoids the impact of asymmetry in the value ranges on both sides. The saturation threshold combines a clinically fixed multiple with robust sample quantiles to avoid excessive influence of extreme outliers on the scale parameter, ensuring stable mapping across datasets. The exponential function further reflects the non-linear variation where the further the feature indicator value deviates from the reference point, the more significant the predictive contribution.

[0072] For example, taking systolic blood pressure as an example, the normal range is... Then the lower limit upper limit Reference point Assume the low-value and high-value saturation thresholds are respectively... and Calculate the normalization scale: .

[0073] Object A: (Midpoint R). Substitute into the calculation (take...) Branches): It is exactly at the normal midpoint, with a marginal value of 0, and has no predictive contribution.

[0074] Object B: (Lower limit of normal). Because Substitute into the calculation: It is within the normal range and its predictive contribution is weak.

[0075] Object C: (Below the lower limit). Substitute into the calculation: If it is below the lower limit of normal, the predictive contribution increases.

[0076] Object D: (Above the upper limit). Because Substitute into the calculation: This exceeds the normal upper limit, indicating a significant contribution to prediction.

[0077] Object E: (High-value saturation threshold). Substitute into the calculation: The sample reached its highest level, and the predicted contribution saturation was 1.

[0078] Object F: (Low-value saturation threshold). Substitute into the calculation: Reaching the lowest sample level results in the largest predictive contribution (here, because...). ,by To unify the denominator, the low-value saturation threshold side has not yet reached -1. If Then, at the low saturation threshold, it can saturate to -1).

[0079] In this embodiment, regarding the marginal value function, replacing the exponential marginal value function with a piecewise linear function, a sigmoid function (such as the logistic function), a power function, or other monotonic mapping functions constitutes an equivalent replacement, as long as it maps the original feature index to the normalized interval and reflects the degree of deviation.

[0080] S103: Assign a feature indicator weight to each feature indicator, and calculate the comprehensive value of the contribution of multiple feature indicators to the prediction of depression risk based on the feature indicator weight and marginal value.

[0081] In this embodiment, depression prediction requires the integration of information from multiple dimensions. Each feature index, after being transformed by a marginal value function, yields a value located at... Marginal value of the interval This reflects the degree and direction of deviation of the current value of the feature index from the normal reference point. However, different feature indices have different importance for prediction, and there may be "interactions" between feature indices that reinforce or cancel each other out. To address this, this invention designs an interaction value function based on multi-attribute value theory, which integrates the independent contribution of each feature index and the interaction effect between feature indices into a single decision model.

[0082] One of the core aspects of the decision-making model is assigning a feature weight to each feature indicator. This indicates the importance of the feature indicator in prediction, and the direction of its value in relation to the risk of depression (positive or negative correlation). However, the final contribution of this feature indicator to the prediction is... Therefore, the direction of contribution is determined by both symbols.

[0083] Different characteristic indicators vary in their importance as auxiliary predictors of depression. For example, core symptoms such as "depressed mood" and "loss of interest" are more important than secondary symptoms such as "loss of appetite." Furthermore, some characteristic indicators (such as higher education level) may act as protective factors, meaning that a higher value of this characteristic indicator may actually reduce the risk of depression. To characterize this difference, a weight is assigned to each characteristic indicator j. And satisfy Its meaning is as follows: The larger the value, the more important the characteristic indicator is in the prediction (whether it is a risk factor or a protective factor). The sign of the symbol reflects the direction of the statistical association between the feature indicator and the risk of depression. The final contribution of a single feature indicator to the prediction is... Its positive and negative values ​​are determined by and The sign is determined by both. If Then the marginal value A positive marginal value (i.e., the presence of symptoms or abnormally high indicators) tends to increase the risk of depression; a negative marginal value tends to decrease the risk. When the marginal value is negative (i.e., the indicator is abnormally low), it tends to increase the risk of depression; when the marginal value is positive, it tends to decrease the risk.

[0084] The sum of the marginal values ​​and weights of each feature indicator yields the combined contribution of all feature indicators to the prediction of depression risk. In other words, the combined value of the independent contributions of all feature indicators is: When the composite value is greater than 0, it indicates that the overall independent contribution tends to support depression, and the larger the value, the stronger the support; when the composite value is less than 0, it indicates that the overall independent contribution tends to exclude depression, and the smaller the value (i.e., the larger the absolute value of the negative number), the stronger the exclusion.

[0085] S104 uses a pairwise combination method to combine multiple feature indicators to obtain multiple feature indicator pairs, and configures the interaction coefficients for the positive and negative results of the two marginal values ​​of each feature indicator pair.

[0086] In this embodiment, the present invention considers the interaction between pairwise feature indicators: when two feature indicators are simultaneously in a certain state (e.g., both are high, one is high and the other is low), they will generate additional positive or negative contributions (i.e., synergistic enhancement or antagonistic weakening), and this additional contribution is determined by the interaction coefficient. Decide.

[0087] When two specific values ​​of two characteristic indicators coexist, their combined predictive effect may differ from the simple sum of their individual effects; this phenomenon is called the interaction effect. Interaction effects can be categorized into positive interaction (synergistic enhancement) and negative interaction (antagonistic weakening). A positive interaction means that when two characteristic indicators coexist, their contribution to prediction is greater than the sum of their individual contributions. For example, "anhedonia" and "loss of energy" are two different items in the predictive criteria for depression, but clinical experience shows that when both are present, the patient's likelihood of depression is much higher than when only one is present, demonstrating the synergistic enhancement effect between core symptoms. A negative interaction means that when two characteristic indicators coexist, their contribution to prediction is less than the sum of their individual contributions. For example, when "psychomotor retardation" and "psychomotor agitation" coexist, the presence of either symptom alone supports a prediction of depression, but when both are present, it often suggests the possibility of mixed characteristic indicators or bipolar disorder. The predictive strength for depression is significantly lower than the sum of their individual contributions, and may even become evidence of exclusion.

[0088] The interaction effect of two feature indices may differ under different combinations of state values, and the sign of the marginal value is precisely a quantitative representation of these state directions. To distinguish the interaction effect under different pairings of marginal value signs, for each feature index pair... The first sub-interaction coefficient when both marginal values ​​are positive, the second and third sub-interaction coefficients when the two marginal values ​​of the feature index pair are one positive and one negative, and the fourth sub-interaction coefficient when both marginal values ​​of the feature index pair are negative are as follows: First sub-interaction coefficient (Both present a risk (or protection) direction).

[0089] Second sub-interaction coefficient (One is in a risky direction, and the other is in a protective direction).

[0090] Third sub-interaction coefficient (One is in a risky direction, and the other is in a protective direction).

[0091] Fourth sub-interaction coefficient (Both are in a protective (or risky) direction).

[0092] A positive interaction coefficient indicates that the combination produces an additional positive contribution (synergistic enhancement), while a negative coefficient indicates an additional negative contribution (antagonistic weakening). The larger the absolute value of the coefficient, the stronger the interaction between the two feature indicators.

[0093] S105, calculate the interaction effect coefficient of each feature index pair based on the two marginal values ​​and the interaction coefficient of each feature index pair, and calculate the average value of the interaction effect of multiple feature index pairs by combining the interaction effect coefficient of each feature index pair.

[0094] In this embodiment, the interaction coefficient of each feature index pair is obtained by multiplying the absolute values ​​of the two marginal values ​​of each feature index pair by the interaction coefficient. Therefore, the interaction effect coefficient of each feature index pair is... The interaction effect coefficient under different combinations of state values ​​is defined as follows: ;in, The learnable interaction coefficients correspond to four combinations of the two marginal value signs and satisfy the condition that they represent the absolute value of the product of the two marginal values.

[0095] When two feature indicators are simultaneously abnormal, the strength of their impact on prediction is positively correlated with the degree of their abnormality: mild abnormalities produce only weak interactions, while severe abnormalities may lead to synergistic enhancement or significant cancellation. This embodiment introduces... The product of these factors ensures that the magnitude of the interaction effect is proportional to the degree to which the characteristic indices deviate from the normal reference point, thus accurately reflecting the clinical reality that the more significant the abnormality, the more pronounced the combined effect. Simultaneously, setting independent interaction coefficients for the four combinations of marginal value signs allows the model to flexibly capture potential directional differences under different value combinations; for example, two positively biased characteristic indices may synergistically enhance (…). A positive and a negative can weaken each other ( ), ), while the two negatively biased eigenvalues ​​may not provide additional support ( This design separates the continuous scaling of interaction intensity from the directional control of value state combinations, achieving refined modeling of the interaction effects of feature indices while maintaining the learnability of parameters. Furthermore, it requires the sum of the four interaction coefficients to be zero, aiming to avoid systematic bias in the model for any value state combination, ensuring the overall balance of interaction effects, thereby improving the identifiability and generalization ability of the parameters.

[0096] The expression for the average value of the interaction effect based on multiple characteristic indicators is as follows: ;in, This represents the interaction coefficient between the marginal values ​​of the j-th feature index and the k-th feature index of the i-th object under the combination of positive and negative results, and its value range is... . For a sign function, when The value is +1 when... The value is -1. (Superscript) This indicates the four interaction coefficients corresponding to this feature index. Select one that matches the current marginal value symbol combination.

[0097] S106, calculate the risk prediction value of the subject to be predicted based on the comprehensive value and the average value, and determine the prediction result of depression of the subject to be predicted based on the relationship between the risk prediction value and the risk threshold.

[0098] In this embodiment, to comprehensively assess the overall depression risk of the i-th subject, it is necessary to integrate the independent contributions of each feature indicator with the interaction effects between all feature indicator pairs. The independent contribution is the weighted sum of the marginal values ​​of each feature indicator, and the interaction component is the average of the interaction effects between all feature indicator pairs. Both parts are constrained to... To ensure that the summation result also falls within this interval, an equal-weighted average is used. In summary, the expression for calculating the risk prediction value is as follows: ;in, This represents the feature index weight of the j-th feature index. Let represent the marginal value of the j-th feature index of the i-th object, and n represent the total number of feature indices. This represents the marginal value of the k-th feature index of the i-th object.

[0099] The first item of the above risk prediction values This represents the composite value of the contributions of all characteristic indicators, and its value range is... ; Second item This represents the average of all characteristic indicators for the interaction effect, and also falls within... Interval. Multiplied by Then, the range of the comprehensive predicted value is: The prediction decision rule is: if If the predicted risk value is higher, the individual is at risk of depression; if the predicted risk value is higher, the confidence level of the prediction is higher. If the predicted risk value is 0, then the subject has no risk of depression, and the smaller the predicted risk value, the higher the confidence level of the prediction. This function transforms the process of integrating multi-dimensional feature indicators in clinical prediction into a computable and interpretable mathematical form. It retains the clinical significance of the importance of each feature indicator to the prediction and its interaction, and ensures the consistency and interpretability of the predicted value range through averaging and weighting operations.

[0100] Regarding the calculation of risk prediction values, without changing the core structure of summing the average value of independent contributions and the average value of interaction effects, adjusting coefficients (such as 1 / 2, 1 / p, etc.), changing the order of summation of terms, or adding linear biases are all considered equivalent implementation methods and are within the protection scope of the embodiments of this invention.

[0101] The expression for the risk prediction value described in this embodiment includes two types of parameters to be learned: feature index weights. and interaction coefficient The task of parameter learning is to find the optimal parameters from historical prediction data so that the model predictions are as close as possible to the true labels. (Where 1 represents the object as having depression, and -1 represents not having depression). This problem can be formalized as a constrained mathematical programming model and can be solved using optimization algorithms.

[0102] Specifically, the feature index weights and interaction coefficients are determined as follows: For each feature index, a feature index weight is configured, and for each feature index pair, an interaction coefficient is configured based on the positive and negative results of the two marginal values. The values ​​of the feature index weights and interaction coefficients are randomly generated. Multiple historical feature indices and true labels of the depressed individuals are obtained. The historical marginal value of each historical feature index is calculated using the marginal value function. Based on the values ​​of the feature index weights and interaction coefficients, and combined with the historical marginal values ​​of each historical feature index, a comprehensive value and an average value are calculated. The predicted value of the depressed individual is calculated based on the comprehensive value and the average value. The predicted value is transformed using the sigmoid activation function to obtain the predicted label of the depressed individual. The predicted label and the true label are substituted into the objective function to obtain the target loss value. When the target loss value meets a preset loss threshold, the randomly generated value under the current iteration conditions is taken as the optimal value of the feature index weights and interaction coefficients. The objective function is an empirical loss function.

[0103] The objective function of this mathematical programming model is expressed as follows: Where m represents the total number of objects, This represents the predicted label after the risk prediction value of the i-th object is transformed by the sigmoid function. This represents the actual label of the i-th object. and These are the regularization coefficients for the first and second regularization terms, respectively. This represents the first sub-interaction coefficient of the feature index pair (j, k) consisting of the j-th feature index and the k-th feature index. This represents the second sub-interaction coefficient of the feature index pair (j, k) consisting of the j-th feature index and the k-th feature index. Let represent the third sub-interaction coefficient of the feature index pair (j, k) formed by the j-th feature index and the k-th feature index. Let represent the fourth sub-interaction coefficient of the feature index pair (j, k) consisting of the j-th feature index and the k-th feature index, and n represent the total number of feature indices.

[0104] It should be noted that the sigmoid transform is: ,in, This is an adjustable parameter used to control the steepness of the sigmoid function: The larger the value, exist The more dramatic the jump in the vicinity, the more sensitive the prediction probability is to small changes in the overall prediction value; conversely, the response is more gradual. It can be used as a decision variable or fixed to a preset value. The first and second regularization terms can be L2 regularization terms. Of course, L2 regularization can also be replaced by L1 regularization, ElasticNet, Lasso, Ridge regression, etc., or other regularization strategies such as early stopping and Dropout can be used, all without departing from the essence of the present invention.

[0105] The decision variables of the objective function must satisfy the following constraint: the sum of the absolute values ​​of the weights of all feature indicators is 1, i.e. And the weight of each feature indicator ranges from 1 to 1. The range of values ​​for each interaction coefficient is: And for each feature index satisfy .

[0106] The objective function is a continuously differentiable optimization problem, which can be solved using gradient-based optimization algorithms (such as AdamW). AdamW is an improved version of the adaptive moment estimation optimization algorithm, which decouples weight decay from gradient update and has better generalization performance than the traditional Adam. It should be noted that replacing AdamW with adaptive or non-adaptive gradient descent optimization algorithms such as Adam, RMSprop, SGD (with or without momentum), and AdaGrad still falls within the parameter learning framework of this invention. This invention uses AdamW to solve the mathematical function defined above, and the solution process is explained in detail: First, all parameters are initialized: the feature index weights are randomly selected near zero and scaled until the sum of their absolute values ​​is 1; the interaction coefficients are randomly selected so that the sum of the four coefficients in each group is zero; and the first and second moment momentum terms are cleared. In each iteration, the algorithm randomly samples a small batch of samples, calculates the loss function under the current parameters (including mean squared error empirical loss and L2 regularization term), and obtains the gradient of the loss with respect to each parameter through automatic differentiation. AdamW's core update mechanism comprises two key components: an adaptive learning rate and decoupled weight decay. The adaptive learning rate maintains the first moment (momentum) and second moment (variance) of the gradient using an exponential moving average, and eliminates the bias in the initial estimate using a bias correction term. Subsequently, during parameter updates, the descent direction is determined by the corrected momentum and variance, while a weight decay term proportional to the parameters themselves is subtracted independently. This design maintains Adam's fast convergence while enhancing the regularization effect.

[0107] Because the constraints of the decision variables in a multi-objective function require the parameters to always satisfy specific constraints (the sum of the absolute values ​​of the feature index weights is 1, the sum of the interaction coefficients is 0, and each coefficient is within a certain range), the parameters must always satisfy specific constraints. Within each round of parameter updates, a projection operation is performed: the weights and interaction coefficients are truncated to the required interval, the mean of each group of interaction coefficients is subtracted to make their sum zero, and finally the weights are renormalized so that the sum of their absolute values ​​is 1. Projection ensures the feasibility of the solution during the optimization process without compromising the convergence of AdamW. Through multiple iterations, the algorithm gradually reduces the loss function value and finally outputs the optimal feature index weights and interaction coefficients that satisfy all constraints. Compared with traditional mathematical programming solvers, AdamW can efficiently handle high-dimensional continuous optimization problems and is insensitive to hyperparameter selection, making it particularly suitable for scenarios with large clinical data scales and high feature index dimensions.

[0108] To evaluate the performance and generalization ability of the risk predictions, k-fold cross-validation was used. Positive and negative samples were randomly divided into k equal-sized subsets. In each round, one subset from each sample was used as the test set, and the remaining subsets were used as the test set. A subset is used as the training set to ensure that the proportion of positive and negative samples is the same in each fold. Algorithm 1 is executed independently in each fold to obtain a set of parameters and performance metrics on the test set: AUC (Area Under the Receiver Operating Characteristic), sensitivity (recall), specificity, and accuracy. Calculations are based on the confusion matrix (where TP is true positive, FN is false negative, TN is true negative, and FP is false positive) and follow standard classification evaluation criteria: Sensitivity = TP / (TP+FN), Specificity = TN / (TN+FP), Accuracy = (TP+TN) / (TP+FN+TN+FP).

[0109] To reduce the randomness of a single data partitioning, the parameters obtained from k-fold training are combined into a comprehensive decision model using their average values: the feature index weights are taken as the arithmetic mean. .

[0110] The interaction coefficient is also taken as the average value. Since the predictions depend only on the sign and relative magnitude of the predicted risk values, scaling the absolute values ​​of the parameters does not affect the decision, thus direct averaging is effective. The four metrics of the integrated decision model are then re-evaluated on the full sample set as the final performance result. This method effectively suppresses overfitting and provides robust performance estimates.

[0111] For any new subject, firstly, their various clinical characteristic indicators (symptoms, laboratory indicators, vital signs, etc.) are collected, and the marginal value of each characteristic indicator is calculated according to the marginal value function. Then, these values ​​are substituted into the comprehensive decision-making model to calculate the risk prediction value.

[0112] Calculate the probability of depression risk based on the sigmoid transform: ,in, The sigmoid steepness parameter is the same as during training. The value range of P is... When the overall prediction value is 0, When the overall predicted value is greater than 0, When the overall predicted value is less than 0, The final prediction rule is: if This predicts that the subject has a risk of depression, and the higher the probability value, the greater the risk; if This predicts that the subject has no risk of depression. This probability output can be directly used for clinical decision support or population screening.

[0113] Furthermore, this invention is primarily applied to clinical and health management scenarios related to depression. In psychiatric outpatient clinics or inpatient wards, it can assist doctors in rapidly generating interpretable depression risk probabilities and decision-making basis by integrating multidimensional data such as symptoms, scales, laboratory tests, and vital signs, improving prediction consistency and efficiency. In primary healthcare institutions and community mental health screenings, early identification and stratification of high-risk groups can be achieved using routine physical examinations or simple questionnaires, guiding referrals. In non-psychiatric wards of general hospitals, it can help identify depressive states accompanying physical illnesses, indicating the need for consultation. During the treatment of depression, it can longitudinally assess the dynamic changes in the contribution of characteristic indicators, assisting in individualized treatment decisions. By providing a transparent and traceable reasoning path, this invention enables medical personnel at different levels to understand and trust the model's recommendations, promoting the implementation of interpretable artificial intelligence in the clinical practice of depression. In addition, the core framework of this invention is not limited to auxiliary prediction of depression risk and can be applied to the following areas: Auxiliary prediction and risk prediction for other mental disorders: such as anxiety disorders, bipolar disorder, schizophrenia, obsessive-compulsive disorder, etc., especially mental illnesses that require comprehensive multi-dimensional clinical characteristic indicators for interpretable assessment.

[0114] Risk assessment and auxiliary prediction of chronic diseases: such as hypertension, diabetes, coronary heart disease, and chronic obstructive pulmonary disease, using routine physical examination and test indicators in electronic health records for early risk warning and prediction support.

[0115] Large-scale population mental health screening based on physical examinations or electronic health records: For large populations such as schools, enterprises, and communities, use routinely available data to conduct rapid initial screening for mental health risks and identify high-risk individuals.

[0116] Treatment response prediction and personalized treatment decision support: By combining patient baseline characteristics and dynamic indicators during treatment, the individualized response probability of different treatment options (such as antidepressants, cognitive behavioral therapy, etc.) can be predicted to assist clinicians in developing optimized treatment strategies.

[0117] Interpretable clinical decision-making scenarios based on multi-source medical data fusion: Applicable to any clinical auxiliary prediction system that needs to integrate structured (laboratory indicators, scale scores) and unstructured (medical history texts, imaging reports) data and output interpretable decision-making results.

[0118] This invention also provides a computer device. The computer device includes a processor, a memory, a communication interface, and at least one communication bus for connecting the processor, the memory, and the communication interface. The memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (PROM), or portable read-only memory (CD-ROM), and is used for related instructions and data.

[0119] The communication interface is used to receive and send data. The processor can be one or more CPUs; if the processor is a single CPU, it can be a single-core CPU or a multi-core CPU. The processor in the computer device reads one or more programs stored in memory and performs the following operations: acquires multiple feature indicators of a target object; divides the multiple feature indicators into various feature indicator classes, constructs a corresponding marginal value function for each feature indicator class to calculate the marginal values ​​of the feature indicators included in each feature indicator class; assigns feature indicator weights to each feature indicator, and calculates the comprehensive value of the contribution of multiple feature indicators to the prediction of depression risk based on the feature indicator weights and marginal values; combines multiple feature indicators in a pairwise manner to obtain multiple feature indicator pairs, and assigns interaction coefficients to the positive and negative results of the two marginal values ​​of each feature indicator pair; calculates the interaction effect coefficient of each feature indicator pair based on the two marginal values ​​and interaction coefficients of each feature indicator pair, and calculates the average value of the interaction effect of multiple feature indicator pairs based on the interaction effect coefficients of each feature indicator pair; calculates the risk prediction value of the target object based on the comprehensive value and the average value, and determines the depression prediction result of the target object based on the relationship between the risk prediction value and the risk threshold.

[0120] It should be noted that the specific implementation of each operation can be described above. Figure 1 The corresponding description of the method embodiments shown herein indicates that a computer device can be used to execute a depression risk prediction method according to the above method embodiments of this application, which will not be described in detail here.

[0121] This invention also provides a computer-readable storage medium, which is a memory device in a computer device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of a terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or nonvolatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the depression risk prediction method in the above embodiments. Those skilled in the art should understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0122] This invention also provides a computer program product containing program instructions. The computer program product may be software or program products containing program instructions, capable of running on a computing device or stored on any usable medium. When the computer program product is run on at least one computer device, it causes the at least one computer device to perform a method for predicting the risk of depression.

[0123] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the risk of depression, characterized by the following indicators: The methods include: Obtain multiple feature metrics of an object to be predicted; Multiple feature indicators are divided into multiple feature indicator classes, and a corresponding marginal value function is constructed for each feature indicator class to calculate the marginal value of the feature indicators included in each feature indicator class. Assign a weight to each feature indicator, and calculate the combined value of the contribution of multiple feature indicators to the prediction of depression risk based on the feature indicator weights and marginal values; Multiple feature indicators are combined in pairs to obtain multiple feature indicator pairs, and an interaction coefficient is configured for the positive and negative results of the two marginal values ​​of each feature indicator pair. Based on the two marginal values ​​and the interaction coefficient of each feature index pair, calculate the interaction effect coefficient of each feature index pair, and combine the interaction effect coefficient of each feature index pair to calculate the average value of the interaction effect of multiple feature index pairs. The risk prediction value of the subject to be predicted is calculated based on the comprehensive value and the average value. The prediction result of depression of the subject to be predicted is determined based on the relationship between the risk prediction value and the risk threshold.

2. The method according to claim 1, characterized in that, The various characteristic indicator categories include five types: symptom characteristic indicator category, nominal characteristic indicator category, upper boundary clinical laboratory characteristic indicator category, lower boundary clinical laboratory characteristic indicator category, and interval clinical laboratory characteristic indicator category.

3. The method according to claim 2, characterized in that, The marginal value function is used to map the parameter values ​​of each feature indicator to... A function on an interval.

4. The method according to claim 1, characterized in that, The comprehensive value of the contribution of multiple feature indicators to the prediction of depression risk is calculated based on the feature indicator weights and marginal values. Specifically, the marginal value and feature indicator weight of each feature indicator are multiplied together and then summed to obtain the comprehensive value of the contribution of all feature indicators to the prediction of depression risk.

5. The method according to claim 1, characterized in that, The interaction coefficients include the first sub-interaction coefficient when both marginal values ​​of the feature index pair are positive, the second and third sub-interaction coefficients when the two marginal values ​​of the feature index pair are one positive and one negative, and the fourth sub-interaction coefficient when both marginal values ​​of the feature index pair are negative.

6. The method according to claim 5, wherein the characteristic index is to calculate the interaction effect coefficient of each characteristic index pair based on the two marginal values ​​and the interaction coefficient of each characteristic index pair, specifically by multiplying the absolute values ​​of the two marginal values ​​and the interaction coefficient of each characteristic index pair to obtain the interaction effect coefficient of each characteristic index pair.

7. The method according to claim 1, characterized in that, The weights and interaction coefficients of the feature indicators are determined as follows: For each feature indicator, the feature indicator weights are configured, and the interaction coefficients are configured for the positive and negative results of the two marginal values ​​of each feature indicator pair. The values ​​of the feature indicator weights and interaction coefficients are randomly generated. Obtain multiple historical characteristic indicators and real labels of individuals with depression; Calculate the historical marginal value of each historical characteristic indicator based on the marginal value function; Based on the values ​​of the feature index weights and interaction coefficients, and combined with the historical marginal values ​​of each historical feature index, the comprehensive value and average value are calculated. The predicted value for individuals with depression is calculated based on the composite value and the average value. The predicted values ​​were transformed using the sigmoid activation function to obtain the predicted labels for individuals with depression. The predicted label and the true label are substituted into the objective function to obtain the target loss value. When the target loss value meets the preset loss threshold, the value randomly generated under the current iteration condition is taken as the optimal value of the feature index weight and the interaction coefficient. The objective function is an empirical loss function.

8. The method according to claim 7, characterized in that, The loss term of the objective function also includes regularization terms for the feature index weights and interaction coefficients, added to the empirical loss function.

9. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the steps of a method for predicting the risk of depression as claimed in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program that, when executed by one or more processors, implements a method for predicting the risk of depression as described in any one of claims 1 to 8.

Images