Chronic actinic dermatitis onset risk prediction method based on risk prediction model

Abstract

The invention relates to a chronic actinic dermatitis onset risk prediction method based on a risk prediction model. The method comprises the following steps: firstly, obtaining multi-dimensional indexes such as photobiology, immune function, molecular markers, clinical basis and life behaviors of a to-be-predicted object, and establishing a data subset; a standard data set is obtained through hierarchical preprocessing, correction and credibility scoring; risk indexes are extracted through variance analysis and recursive feature elimination, the comprehensive association degree of the indexes is calculated in combination with Pearson correlation coefficients and mutual information values, and high-association-degree indexes are screened to construct risk feature vectors; inputting the vector into a multi-modal fusion risk prediction model constructed by a feature attention mechanism and an improved gradient boosting tree algorithm, and outputting a morbidity probability, a risk level and a core risk factor ranking; and in combination with skin typing and photosensitive contact history calibration risk levels, a preset prevention and control measure table is called to execute personalized prevention and control. According to the method, accurate prediction and personalized prevention and control guidance of CAD onset risks are realized, and the early recognition rate and intervention accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a chronic actinic dermatitis pathogenic risk prediction method based on a risk prediction model. BACKGROUND

[0002] Chronic actinic dermatitis (CAD) is a type of idiopathic photosensitive skin disease characterized by chronic inflammation at light-exposed sites. Its pathogenesis is complex and its clinical diagnosis and treatment are difficult, and it has become an important threat to the skin health of middle-aged and elderly male groups. According to clinical statistics, the global incidence of CAD is increasing year by year, and it is more common in tropical and subtropical regions or summer. Some severe cases can progress to cutaneous T-cell lymphoma, significantly reducing the quality of life of patients and increasing the medical burden.

[0003] The pathogenesis of CAD is related to the synergistic action of multiple factors. Currently known ultraviolet (especially UVA, UVB) irradiation is the core inducer, but its pathogenic process also involves multiple mechanisms such as immune dysfunction, skin barrier damage, and abnormal molecular regulation. Clinical studies have shown that CAD patients have peripheral blood T cell subpopulation imbalance (CD3 + CD4 + ratio decreases, CD3 + CD8 + ratio increases), Th1 cell overactivation (IFN-γ / IL-4 ratio increases), and abnormal expression of skin tissue-specific lncRNA (such as RP11-356I2.4). The discreteness and correlation of these indicators increase the difficulty of early diagnosis. In addition, there are individual differences in the sensitivity of CAD patients to ultraviolet light. Even in an indoor environment, 50% UVA that can penetrate ordinary glass may still induce the disease, further increasing the complexity of risk prediction.

[0004] Current clinical assessment of CAD risk mainly relies on single indicator detection and physician experience, and lacks systematicness and precision. For example, the minimal erythema dose (MED) determined by a solar simulator can reflect skin photosensitivity, but it does not combine key dimensions such as immune function and molecular markers, and is prone to miss potential patients with normal photosensitivity but immune disorder; traditional photopatch tests can only identify known photosensitizing substance contact history and cannot quantify the impact of environmental photosensitive load (such as residential ultraviolet radiation intensity and sunscreen adherence) on the onset. In addition, existing methods do not consider the differentiated risk of skin types (Fitzpatrick classification), and the mean UVA-MED of type III skin patients is significantly lower than that of type IV skin, but a unified risk determination standard is used, resulting in a high misdiagnosis rate.

[0005] The incidence of CAD has the characteristics of chronicity and recurrence. The existing technology focuses on single diagnosis, and does not construct a dynamic risk model based on long-term index monitoring. For example, the CD3 + CD4 + / CD3 + CD8 + When the ratio returns to 1.2 or more, the probability of disease remission, but the traditional method does not include such dynamic indicators in the risk update mechanism.

[0006] In the existing CAD related research, the feature screening depends on single variance analysis (ANOVA) or single variable correlation analysis. For example, there is a segmented correlation between the environmental light sensitivity load index and the incidence of risk, and the Pearson correlation coefficient cannot capture such non-linear relationship, resulting in that the contribution of the index in the traditional model is underestimated. At the same time, without removing low contribution indicators, there is redundancy in the model input, such as height, weight and other indicators that have no significant correlation with CAD, which reduces the prediction accuracy.

[0007] The existing CAD risk prediction model mostly uses a single algorithm, without considering the modal difference of multi-dimensional indicators. The weight contribution of photobiology indicators and molecular markers is different, with the former being about 0.35 and the latter being about 0.3. A single model cannot dynamically allocate dimension weight.

[0008] The incidence probability output by the existing model is not combined with individual skin type and light sensitivity contact history for secondary adjustment, resulting in that the risk of patients with type III skin + light sensitivity contact history is underestimated, and the risk of patients with type IV skin + no light sensitivity contact history is overestimated. This lack of calibration directly affects the accuracy of risk level determination, and further leads to mismatched intervention measures.

[0009] With the development of precision medicine, the demand for CAD risk prediction in clinical practice has shifted from qualitative diagnosis to quantitative stratification + dynamic intervention. The present application aims to provide a prediction method that integrates multi-dimensional indicators, optimizes feature screening, and has risk calibration and intervention guidance capabilities, to solve the problems of early identification difficulty, low prediction accuracy and insufficient personalized intervention in the prior art. Through the innovative design of hierarchical preprocessing, multi-modal fusion model and dynamic risk calibration, the accuracy and clinical adaptability of CAD incidence risk prediction can be effectively improved. SUMMARY

[0010] The purpose of the present application is to provide a risk prediction model based method for predicting the incidence of chronic photodermatitis, to solve the above technical problems in the prior art.

[0011] To solve the above technical problems, the technical solutions adopted by the present application are as follows: The risk prediction model based method for predicting the incidence of chronic photodermatitis comprises the following steps: S1: Obtain the photobiological index, immune function dynamic index, molecular marker combination index, clinical basic index and life behavior risk index of the to-be-predicted object within a specified time period, and establish corresponding data subsets respectively; S2: The data subsets are preprocessed in a hierarchical manner, and then corrected and scored for credibility. Data with a score greater than or equal to a specified threshold is retained and merged into a data set; S3: Risk index extraction is performed on the data set: first, filter out low-variation irrelevant indicators, then gradually eliminate indicators with low prediction contribution from the remaining indicators, and finally retain a specified number of candidate indicators; S4: Calculate the correlation of each candidate indicator with CAD incidence risk: Pearson correlation quantifies linear correlation, mutual information value quantifies nonlinear correlation, and the comprehensive correlation of each indicator is calculated by weighting; S5: Sort the candidate indicators according to the comprehensive correlation from high to low, select indicators with a comprehensive correlation greater than or equal to a threshold to construct a risk feature vector, and assign weights to each dimension according to the comprehensive correlation; S6: Input the risk feature vector into the pre-constructed multi-modal fusion risk prediction model, and the multi-modal fusion risk prediction model outputs the incidence probability of the to-be-predicted object. According to the incidence probability, the incidence risk level and the core risk factor ranking are obtained.

[0012] Preferably, it also includes a risk level calibration process, which is as follows: S7: According to the Fitzpatrick skin type standard, divide it into type III and type IV, and determine it by the history of contact with photosensitizing substances or no photosensitizing substances. Set groups based on skin type and photosensitizer contact history, and set probability adjustment coefficients respectively: S8: Determine the grouping of the to-be-predicted object, obtain the corresponding adjustment coefficient, calibrate the incidence probability predicted by the model based on the adjustment coefficient, and set the incidence probability boundary constraint. According to the adjusted probability value, combined with the risk level division standard, the final incidence risk level is determined; S9: Based on the final incidence risk level and the core risk factor ranking, call the preset risk prevention and control measure table to perform corresponding risk prevention and control measures on the to-be-predicted object.

[0013] Preferably, in step S1, the photobiological index is the minimum erythema dose of long-wave ultraviolet light UVA-MED, the minimum erythema dose of medium-wave ultraviolet light UVB-MED and the erythema recovery time ERT measured by a sunlight simulator, the immune function dynamic index is the ratio of peripheral blood T cell subsets detected by flow cytometry, CD3 + CD4 + / CD3 + CD8 +The ratio and Th1 / Th2 cytokine balance state, the combination of molecular marker indicators including serum IFN-γ, IL-12, sFas / sFasL, skin tissue-specific lncRNA and TNFAIP3 gene promoter region methylation rate, clinical basic indicators including age, gender, Fitzpatrick skin type and history of previous photosensitive diseases, and life behavior risk indicators including daily sun exposure time, sunscreen compliance, photosensitive substance intake frequency and environmental photosensitive load index.

[0014] CD3 + CD4 + / CD3 + CD8 + The ratio of CD3 + CD4 + , that is, the ratio of CD3 + CD4 + , that is, the ratio of CD3 + CD8 + , that is, the ratio of CD3 + CD8 + , that is, the ratio of CD3 + CD4 + / CD3 + CD8 + The ratio is obtained by dividing CD3 + CD4 + by CD3 + CD8 + , which reflects the balance state of T cell subsets.

[0015] Preferably, the 5 data subsets in step S2 are preprocessed by stratification preprocessing, and then corrected and scored for reliability as follows: S21: Box-Cox transformation is performed on the photobiological indicators to eliminate dimensions; The immune function dynamic indicators are subjected to Z-score method to remove outliers; The molecular marker combination indicators are filled with missing values by multiple interpolation method; The clinical basic indicators and life behavior risk indicators are converted into classification variables by one-hot encoding; Each indicator is standardized; S22: The data distribution of each data subset is verified by 3σ criterion, and the abnormal subset is corrected by local weighted regression; S23: Based on the detection instrument precision, sample collection stability, generate 0-10 data credibility score, reserve the data with score ≥8, merge to get the standardized data set.

[0016] Preferably, the specific process of step S3 is as follows: S31: Obtain the index data of the healthy subjects with the same clinical basic indicators as the to-be-predicted subjects as the healthy control group, and take the data set in step S2 as the CAD case group; S32: Calculate the overall variance, intra-group variance, inter-group variance of each indicator, and calculate the F value through the ANOVA formula, F value = inter-group variance / intra-group variance; The following screening criteria are set for index screening: The overall variance of the index is greater than 0.1; The F value is greater than 3.85; The index meets the above two points at the same time, otherwise it is rejected; S33: Further index screening is performed through the incremental index elimination method RFE.

[0017] Preferably, the specific process of calculating the correlation degree of each candidate indicator with the CAD incidence risk in step S4 is as follows: S41: Quantify the linear correlation between each candidate indicator and the CAD incidence risk by calculating the Pearson correlation coefficient; S42: Quantify the nonlinear correlation between each candidate indicator and the CAD incidence risk by calculating the mutual information value; S43: Calculate the comprehensive correlation degree based on the Pearson correlation coefficient and the mutual information value: adopt the weighted formula of linear correlation weight 0.6+ nonlinear correlation weight 0.4.

[0018] S44: For all candidate indicators screened by ANOVA, repeat steps S41-S43 to calculate the R 综合 .

[0019] Preferably, the specific process of step S5 is as follows: S51: Calculate the R 综合 According to the order from high to low, select the indicators with R 综合 ≥0.6 to construct the risk feature; S52: Based on the 8-12 indicators screened, construct the risk feature vector, and give the basic weight to each dimension according to the comprehensive correlation degree.

[0020] Preferably, the multi-modal fusion risk prediction model of step S6 is constructed by a feature attention mechanism + an improved gradient boosting tree algorithm, the risk feature vector of the historical case is taken as a training sample, the feature attention mechanism dynamically allocates the basic weight of each dimension in the risk feature vector, and the improved gradient boosting tree algorithm optimizes the training efficiency of the model by introducing an adaptive learning rate; After the risk feature vector of the to-be-predicted object is input into the multi-modal fusion risk prediction model, the multi-modal fusion risk prediction model outputs the incidence probability of the to-be-predicted object, and the incidence risk level and the core risk factor ranking are obtained according to the incidence probability, and the risk level includes low risk, medium risk and high risk.

[0021] The beneficial effects of the present application include: The chronic actinic dermatitis incidence risk prediction method based on the risk prediction model provided by the present application solves the core problems of single index, insufficient precision and blind intervention in the existing CAD risk prediction through multi-dimensional index integration, refined feature screening, multi-modal model construction and personalized risk calibration, and achieves significant breakthroughs in clinical practicability, prediction accuracy and personalized adaptability.

[0022] 1. The photobiological index, immune function index, molecular marker, clinical basic index and life behavior index are integrated into a unified system, and the cause- immune- molecular- behavior pathogenic chain of CAD is comprehensively covered. For example, through the cooperative detection of UVA-MED and TNFAIP3 methylation rate, the recognition rate of early asymptomatic patients is improved, and the traditional single index is avoided.

[0023] 2. Through the two-step strategy of ANOVA filtering low-variation index + RFE removing low-contribution index, combined with the comprehensive correlation calculation of the Pearson correlation coefficient of linear correlation + the mutual information value of nonlinear correlation, 8-12 core features are accurately reserved, the effective correlation rate of model input features is improved, and the reliability of model input is ensured through the 3σ criterion and the credibility score control data quality.

[0024] 3. The feature attention mechanism + improved gradient boosting tree is used to construct the model, the feature attention mechanism can dynamically allocate the weight of each dimension, the improved gradient boosting tree optimizes the training efficiency through the adaptive learning rate, can capture the linear correlation between indexes, and can identify the nonlinear correlation, so that the prediction accuracy of the model is improved.

[0025] 4. According to the risk difference of Fitzpatrick classification, the grouping adjustment coefficient of skin type-light sensitive contact history is set, and the incidence probability is optimized through a quadratic calibration formula. The calibration mechanism reduces the misjudgment rate of type III skin patients and type IV skin patients, and solves the adaptability problem of the traditional unified standard.

[0026] 5、The model output not only contains the risk level, but also analyzes and generates the core risk factor ranking, which clearly identifies the key driving factors of the patient's disease, facilitates targeted intervention based on the core risk factor ranking, avoids one-size-fits-all clinical intervention, and reduces waste of medical resources, such as unnecessary genetic testing for medium-risk patients.

[0027] 6、Promote the transformation of CAD prevention and control from passive diagnosis to active early warning: early risk identification and delay of disease progression, through early warning of potential patients with normal photosensitivity but immune disorders or abnormal molecular markers, the risk can be identified 6-12 months before clinical lesions appear. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The flowchart of the chronic photodermatitis onset risk prediction method based on the risk prediction model of the application.

[0029] Figure 2 The logic principle diagram of the chronic photodermatitis onset risk prediction based on the risk prediction model of the application.

[0030] Figure 3 The architecture diagram of the multi-modal fusion risk prediction model of the application. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings Figures 1-3 Further detailed description of the application: Example 1: Referring to the drawings Figure 1 and Figure 2 As shown, the chronic photodermatitis onset risk prediction method based on the risk prediction model comprises the following steps: S1: Obtain the photobiological index, immune function dynamic index, molecular marker combination index, clinical basic index and life behavior risk index of the to-be-predicted object within a specified time period, and establish corresponding five data subsets respectively.

[0032] S2: The five data subsets are preprocessed in a hierarchical preprocessing manner, and then corrected and scored for credibility, and the data greater than or equal to the specified score threshold is retained to form a data set.

[0033] S3: Risk index extraction is performed on the data set: first, filter out low-variation irrelevant indexes, then gradually eliminate indexes with low prediction contribution from the remaining indexes, and finally retain 15-20 candidate indexes.

[0034] S4: Calculate the correlation degree of each candidate index with the risk of CAD: Pearson correlation coefficient quantifies linear correlation, the value range is [-1, 1], the closer to 1, the higher the correlation degree, mutual information value quantifies nonlinear correlation, the value range is [0, log2(n)], n is the sample size, the larger the value, the higher the correlation degree, the comprehensive correlation degree of each index is calculated by a weighted formula, the value range is [0, 1].

[0035] S5: Sort the candidate indexes according to the comprehensive correlation degree from high to low, select the indexes with a comprehensive correlation degree greater than or equal to 0.6 to construct a risk feature vector, the vector dimension is 8-12, each dimension is assigned a weight according to the correlation degree, the weight of the index with the highest correlation degree is set to 0.2, and the weights of the remaining indexes decrease in proportion to the correlation degree.

[0036] S6: Input the risk feature vector into the pre-constructed multi-modal fusion risk prediction model, and the multi-modal fusion risk prediction model outputs the incidence probability of the object to be predicted, and the incidence risk level (low risk, medium risk, high risk) and the ranking of the core risk factors are obtained according to the incidence probability.

[0037] In this embodiment, the photobiological indexes in step S1 are the minimum erythema dose of long-wave ultraviolet UVA-MED, the minimum erythema dose of medium-wave ultraviolet UVB-MED and the erythema regression time ERT measured by a sunlight simulator, the dynamic immune function indexes are the peripheral blood T cell subsets (CD3 + CD4 + proportion, CD3 + CD8 + proportion), CD3 + CD4 + / CD3 + CD8 + ratio and Th1 / Th2 cytokine balance state (IFN-γ / IL-4 ratio), the IFN-γ / IL-4 ratio refers to the value obtained by dividing the concentration of IFN-γ (interferon-γ) in serum by the concentration of IL-4 (interleukin-4) in serum; the combination of molecular markers includes serum IFN-γ, IL-12, sFas / sFasL (ABC-ELISA method), skin tissue-specific lncRNA (RP11-356I2.4, LNC_000310, qRT-PCR method) and TNFAIP3 gene promoter region methylation rate (bisulfite sequencing method, detection region chr6: 1378200-1378500); the clinical basic indexes include age, gender, Fitzpatrick skin type and history of previous photosensitive diseases, and the life behavior risk indexes include daily average sun exposure time, sunscreen measure compliance, photosensitive substance intake frequency and environmental photosensitive load index (calculation formula: index = UV radiation intensity of residence × sun exposure time × indoor glass UVA penetration rate).

[0038] In another embodiment of the present embodiment, a risk level calibration process is also included, specifically as follows: S7: According to the Fitzpatrick skin type standard, it is divided into type III and type IV, type III is light brown skin, easy to sunburn and moderate sunburn after sun exposure, type IV is brown skin, not easy to sunburn, easy to sunburn after sun exposure. The ultraviolet tolerance threshold of the two types of skin is significantly different, and the UVA-MED of type III skin is 20%-30% higher than that of type III.

[0039] Through the past medical history, medication history (such as taking quinolone antibiotics, tetracycline drugs), occupational exposure history (such as long-term contact with asphalt, coal tar) and light patch test results of the object to be predicted, it is determined to have clear contact history (contact history ≥1 times and light patch test positive) or no photosensitive contact history (no contact history or light patch test negative).

[0040] Based on skin type and photosensitive substance contact history, set groups, and set probability adjustment coefficients respectively K : Type III has clear contact history, and the adjustment coefficient K is 1.10-1.15; Type III has no clear contact history, and the adjustment coefficient K is 1.00-1.08; Type IV has clear contact history, and the adjustment coefficient K is 1.05-1.08; Type IV has no clear contact history, and the adjustment coefficient K is 0.92-0.95.

[0041] S8: Determine the group of the object to be predicted, obtain the corresponding adjustment coefficient, and adjust the incidence probability P init calibration, the calibration formula is as follows: P adj = P init ×K ; Among them, P adj is the incidence probability after calibration; and set the incidence probability boundary constraint: If P adj >1.0, then set P adj =1.0 (judged as extremely high risk); If P adj <0.05, then set Padj = 0.05 (judged as very low risk), to ensure the rationality of subsequent risk level determination; According to the adjusted probability value P adj , combined with the risk level classification standard, the final incidence risk level is determined: low risk: P adj < 0.25; Medium risk: 0.25≤ P adj ≤ 0.75; High risk: P adj > 0.75.

[0042] S9: Based on the final incidence risk level and the ranking of core risk factors, the preset risk prevention and control measure table is called to perform corresponding risk prevention and control measures on the to-be-predicted object. The risk prevention and control measure table is shown in Table 1 below.

[0043] Table 1 Risk prevention and control measure table

[0044] The final incidence risk level of the to-be-predicted object is obtained from the multi-modal fusion model output result, and from the ranking of core risk factors output by the model, the top 3 factors with the highest comprehensive correlation degree R 综合 are selected, and the abnormal state of each factor is determined, such as TNFAIP3 methylation rate < 20%, RP11-356I2.4 expression lower than 50% of the healthy mean value.

[0045] For example: the top 3 core factors of a high-risk object are: 1. TNFAIP3 methylation rate < 20%, R 综合 = 0.75; 2. RP11-356I2.4 expression is low, R 综合 = 0.72; 3. (IFN-γ) / (IL-4) > 2.0, R 综合 = 0.68; Form the "high risk-TNFAIP3+RP11-356I2.4+IFN-γ / IL-4" combination label of the object.

[0046] In the risk prevention and control measure table, first locate vertically according to the final risk level, such as the high risk column, and then locate horizontally according to the top 3 core risk factors, such as the "TNFAIP3 methylation rate+RP11-356I2.4+IFN-γ / IL-4" row, and the intersection cell is the exclusive control scheme of the object.

[0047] Example 2: On the basis of embodiment 1, the specific process of targeted pretreatment of 5 data subsets in step S2 through hierarchical pretreatment is as follows, and then correction and reliability scoring are performed respectively: S21: Box-Cox transformation is performed on the photobiological index to eliminate dimension; The dynamic index of immune function is removed by Z-score method (Z-score absolute value > 3); The combination index of molecular markers is filled with missing values by multiple interpolation method (based on the same disease stage and age group distribution); The clinical basic index and life behavior risk index are converted into classification variables by one-hot encoding; Each index is standardized.

[0048] S22: The data distribution of each data subset is verified by 3σ criterion, 99.7% data falls within mean ± 3 times standard deviation, and abnormal subset is corrected by local weighted regression LOESS.

[0049] S23: Based on the accuracy of detection instrument and the stability of sample collection, 0-10 score data reliability score is generated, data with score ≥8 is reserved, and standardized data set is obtained by merging.

[0050] The specific process of step S3 is as follows: S31: Obtain the index data of healthy subjects with the same clinical basic index as the to-be-predicted object as the healthy control group, and the data set in step S2 as the CAD case group.

[0051] S32: Calculate the overall variance, within-group variance, within-group variance, and between-group variance of each index, and calculate F value by ANOVA formula: F value = between-group variance / within-group variance, between-group variance reflects the difference between the two groups, and within-group variance reflects the data fluctuation within the group.

[0052] The following screening criteria are set for index screening: The overall variance of the index is greater than 0.1 to ensure that the index has sufficient variation; F value > 3.85 to ensure that the difference between the two groups has statistical significance; The index that meets the above two points is retained, otherwise it is removed, and the number of retained indexes is M; Take CD3 + CD4 + Proportion index as an example, assuming that the sample size of the case group n1 = 200, the sample size of the control group n2 = 200, and the basic parameters are as follows: n1: sample size of the case group; n2: sample size of the control group; total sample size N = n1 + n2 = 400; Mean values ​​of indicators in the case group, such as CD3 in the case group. + CD4 + The average proportion is 30.2%; Mean values ​​of control group indicators, such as CD3 in the healthy group. + CD4 + The average proportion is 42.5%; Total sample CD3 + CD4 + The average percentage of the proportional indicator is 36.35%: (200 × 30.2% + 200 × 42.5%) / 400 = 36.35; Case group number i CD3 of each sample + CD4 + Proportional index values, such as CD3 in case 1 of the case group + CD4 + =28.5%; : Control group j CD3 of each sample + CD4 + Proportional index values, such as CD3 in case 1 of the control group. + CD4 + =41.8%; k: number of groups, k=2.

[0053] The process of calculating the variance between groups is as follows: First calculate the sum of squared deviations between groups from the mean. SSB Between-group degrees of freedom dfB Then, the variance between groups can be obtained by comparing the two values: Between-group sum of squared differences SSB This measures the sum of squared deviations of each group's mean from the overall mean, using sample size weighting to ensure fairness. The specific formula is as follows: ; Substitute the above data: SSB = 200 × (30.2 - 36.35) 2 +200×(42.5-36.35) 2 =200×(-6.15) 2 +200×(6.15) 2 =200×37.82+200×37.82 =7564+7564=15128; Between-group degrees of freedom (dfB): the amount of information from independent groups, expressed by the formula:dfB = k - 1 = 1; Between-group variance MSB = SSB / dfB = 15128 / 1 = 15128.

[0054] The within-group variance reflects the fluctuation of the index within the same group (such as within the case group). The smaller the fluctuation, the stronger the stability of the index and the higher the reliability.

[0055] The calculation process of the within-group variance MSW is as follows: First, calculate the sum of squares of within-group deviations SSW and the within-group degrees of freedom dfW , and then get the within-group variance through their ratio: Sum of squares of within-group deviations SSW : measures the sum of squares of deviations of each sample within the group from the mean of the group, excluding the interference of group differences.

[0056] ; Substitute the above data: SSW = 850 + 920 = 1770; Within-group degrees of freedom dfW is the number of independent samples within the group, and the formula is: dfW=N-k = 400 - 2 = 398; Within-group variance MSW = SSW / dfW= 1770 / 398 ≈ 4.45.

[0057] Total variance: reflects the overall variation of all samples (case group + control group), which is a comprehensive reflection of the between-group variance and the within-group variance. First, calculate the sum of squares of total deviations SST and the total degrees of freedom dfT , and then get the total variance through their ratio.

[0058] The sum of squares of total deviations SST measures the sum of squares of deviations of all samples from the total mean, which is equal to the sum of the between-group sum of squares and the within-group sum of squares: That is, SST = SSB + SSW = 15128 + 1770 = 16898; Total degrees of freedom dfT is the number of total independent samples, dfT = N - 1 = 400 - 1 = 399; Total variance σ 2 = 16898 / 399 = 42.35; Set the following screening criteria for index screening: Total variance of index > 0.1 (to ensure that the index has sufficient variation); F value = 15128 / 4.45 = 3400 > 3.85, indicating that the group difference is significant, and the index can be retained; S33: Further index screening by incremental index elimination method RFE: Eliminate the 5% of the M indexes with the lowest importance scores in each iteration process of the logistic regression model, the specific process is as follows: First iteration: Train the logistic regression model with the M indexes, calculate the importance score of each index by the absolute value of the model coefficient, the higher the score, the greater the contribution to prediction, assuming N = 22; Eliminate the 1 index with the lowest score (5% x 22 = 1), leaving 21 indexes; Re-train the logistic regression model with 21 indexes, calculate the 5-fold cross-validation AUC value, assuming it increases from 0.82 to 0.83.

[0059] 2-5 iterations: Repeat the process of training the logistic regression model, calculating the importance, eliminating the lowest score index, and verifying the AUC. Eliminate 1 low contribution index each time, and the number of indexes decreases from 21 to 18; If the AUC value does not improve after an iteration, such as 0.85, then continue to eliminate; if the AUC value does not improve for 3 consecutive times, such as 0.86 for the 3rd-5th time, then stop iteration; Assuming that after 5 iterations, the number of indexes decreases from 22 to 18, and the AUC value stabilizes at 0.86 (0.04 higher than the initial 22 indexes), then the 18 indexes are the candidate indexes after RFE screening.

[0060] Example 3: On the basis of Example 1 or Example 2, the specific process of calculating the correlation degree of each candidate index with the risk of CAD in step S4 is as follows: Data preparation: Obtain 500 cases of CAD confirmed cases, meet the criteria of disease course > 1 year, rash distribution in light exposed parts, and clear photosensitivity symptoms; 500 healthy control samples, no history of photosensitivity diseases, no recent photosensitive drugs, total sample size N ≥ 1000, pre-process the samples. Construct a "candidate index-disease state" data set, where the disease state is a binary variable, the case group is marked as 1, and the control group is marked as 0.

[0061] Define core parameters: X: Numerical vector of candidate indexes, such as UVA-MED values of 500 cases + 500 controls, a total of 1000 data points; Y: Disease state, 1000 data points, 1 = CAD case, 0 = healthy control; r: Pearson correlation coefficient, value [-1, 1], the larger the absolute value |r|, the stronger the linear correlation; MI(X, Y): mutual information value of index X and disease state Y, taking value [0, log2n], when n = 1000, log21000≈9.97, the greater the value, the stronger the nonlinear correlation; maxMI: maximum value of mutual information value of all candidate features, used for normalization, so that the mutual information value is mapped to the range [0, 1]; R 综合 : comprehensive correlation value [0, 1], the closer the value to 1, the stronger the correlation between the feature and the risk of CAD.

[0062] S41: Quantify the linear correlation between each candidate index and the risk of CAD by calculating the Pearson correlation coefficient: ; Wherein, n is the total sample size, is the product sum of index X and disease state Y, ∑X is the total sum of index X, ∑Y is the total sum of disease state, is the square sum of index X, ∑Y 2 is the square sum of disease state.

[0063] Take UVA-MED feature as an example: assuming that the mean of UVA-MED of 500 cases (Y = 1) is 8.2 J / cm², the mean of UVA-MED of 500 controls (Y = 0) is 18.5 J / cm², the total sample size n = 1000, and the calculation is as follows: ∑X = 500 × 8.2 + 500 × 18.5 = 13350; ∑Y = 500 × 1 + 500 × 0 = 500; ∑XY = 500 × 8.2 × 1 + 500 × 18.5 × 0 = 4100; ∑X 2 = 500 × (8.2 2 + variance) + 500 × (18.5 2 + variance) = 218925; variance = 4.96; ∑Y 2 = 500 × 12 + 500 × 02 = 500.

[0064] Substituting the formula gives: r = (1000 × 4100 - 13350 × 500) / ([1000 × 218925 - 13350 2 ] [1000 × 500 - 5002]) 1 / 2 ≈ -0.78; Interpretation of results: r≈-0.78, the negative sign indicates that UVA-MED is negatively correlated with the risk of CAD, that is, the lower the UVA-MED, the higher the risk of CAD. |r|=0.78, the linear association is relatively strong.

[0065] S42: Quantifying the nonlinear association between each candidate indicator and the risk of CAD incidence by calculating mutual information values: Discretization: First, the continuous feature X is discretized into K intervals. For example, UVA-MED uses 5J / cm² intervals to divide it into 5 intervals: [0-5), [5-10), [10-15), [15-20), and ≥20J / cm². This ensures that each interval contains enough samples, with ≥50 samples.

[0066] Calculate the marginal probabilities and joint probabilities: Marginal probability P( X=x i Feature X takes the value of x i The sample proportion is given by the formula: P( X=x i The characteristic X takes the value of ) = x i Number of samples / n; Marginal probability P( Y=y j The disease state Y takes the value of y j Sample proportion, formula is P( X=y j ) = Number of samples in disease state Y / n; Joint probability P( X=x i P( Y=y j Feature X takes the value of x i And the disease state Y takes the value y j The sample proportion, P( X =x i P( Y=y j The feature X takes the value of x i And the disease state Y takes the value y j Number of samples / n.

[0067] The formula for calculating mutual information value is as follows: ; like X and YIrrelevant, P( X=x i P( Y=y j ))=P( X=x i )·P( Y=y j At this point, log2(1) = 0, MI = 0; The stronger the correlation between X and Y, the larger the ratio and the higher the MI value.

[0068] Taking the UVA-MED index as an example: Assuming UVA-MED is discretized into 5 intervals, the joint probability and marginal probability of each interval are calculated, as shown in Table 2 below: Table 2 Joint Probability and Marginal Probability of Each Interval

[0069] Substitute into the formula: MI(X, Y)=0.1425×log2(0.1425 / (0.15×0.5))+0.0075×log2(0.0075 / (0.15×0.5))+ ...+0.0950×log2(0.0950 / (0.10×0.5)) ≈4.2; When MI≈4.2 and n=1000, maxMI≈9.97, it indicates that there is a strong nonlinear association between UVA-MED and the risk of CAD.

[0070] S43: Calculate the overall correlation degree based on Pearson correlation coefficient and mutual information value: Since the MI value range varies for different features, depending on the feature discretization interval and sample distribution, it needs to be normalized to the [0,1] range first. The formula is: MI 归一化 =MI(X,Y) / maxMI; If the maxMI of all candidate features is approximately 9.97, then the MI of UVA-MED is... 归一化 =MI(X, Y) / maxMI=4.2 / 9.97≈0.42.

[0071] Using a weighted formula of 0.6 for linear association and 0.4 for nonlinear association, and based on clinical data validation, linear association contributes more to CAD risk prediction. Therefore, the formula gives a greater weight to linear association: R 综合 =0.6×|r|+0.4×MI 归一化 ; Input data from the UVA-MED instance: R 综合= 0.6 * 0.78 + 0.4 * 0.42 = 0.636; Then the comprehensive correlation degree of UVA-MED is approximately 0.64 (≥ 0.6), which meets the high correlation degree characteristic standard and can be included in the risk characteristic vector.

[0072] S44: For all candidate indicators (such as UVA-MED, CD3 + CD4 + ratio, TNFAIP3 methylation rate, etc.) screened by ANOVA, repeat steps S41-S43 to calculate the R 综合 .

[0073] The specific process of step S5 is as follows: S51: Calculate the R 综合 According to the order from high to low, filter the indicators with R 综合 ≥ 0.6 to construct the risk characteristics, usually retaining 8-12 indicators; S52: Based on the 8-12 indicators screened, construct the risk characteristic vector, i.e. the vector dimension is 8-12, and assign the basic weight to each dimension according to the comprehensive correlation degree, with the highest correlation degree indicator weight set to 0.2, and the rest of the indicators decreasing in proportion to the correlation degree.

[0074] Referring to Figure 3 , the multi-modal fusion risk prediction model is constructed based on feature attention mechanism + improved gradient boosting tree algorithm, with the risk characteristic vector of historical cases as the training sample. The feature attention mechanism dynamically allocates the basic weight of each dimension in the risk characteristic vector. The improved gradient boosting tree algorithm optimizes the model training efficiency by introducing an adaptive learning rate. The initial learning rate is 0.05, and it is dynamically decayed according to the formula learning rate = 0.05 * [1-iteration number / 1000*0.6] with the number of iterations). After the risk characteristic vector of the object to be predicted is input into the multi-modal fusion risk prediction model, the multi-modal fusion risk prediction model outputs the incidence probability of the object to be predicted, and according to the incidence probability, the incidence risk level (low risk, medium risk, high risk) and the ranking of core risk factors are obtained.

[0075] The specific process of the trained multi-modal fusion risk prediction model outputting the incidence probability and the ranking of core risk factors of the object to be predicted is as follows: Input layer: The risk characteristic vector of the object to be predicted, i.e. the 8-12 high correlation degree indicators screened out, such as UVA-MED, TNFAIP3 methylation rate, etc., each dimension with a basic weight, is input into the input layer of the model, completing the preliminary reception and format adaptation of data, ensuring that the subsequent modules can directly call the indicator data in the vector.

[0076] The feature attention mechanism dynamically adjusts the feature weight: Based on the model-based mean square error loss function, the gradient of each risk feature dimension is calculated, and the weight of each feature is dynamically adjusted according to the gradient size: If a certain indicator (such as TNFAIP3 methylation rate) has a large impact on the loss function (large absolute value of gradient), it means that this feature is more critical to the prediction result and will increase its weight; If a certain feature (such as some clinical basic indicators) has a small impact on the loss function (small absolute value of gradient), its weight will be reduced.

[0077] Every 50 model iterations, the weight of each feature is recalculated and updated, making the weight distribution more suitable for the current data prediction needs.

[0078] Improved gradient boosting tree calculation of incidence probability: First, perform hyperparameter optimization, and the model completes the optimization configuration of hyperparameters in advance: The number of trees is set to 1000 to ensure sufficient learning ability of the model; The maximum depth of each tree is set to 8 layers to avoid overfitting; The minimum number of split samples is set to 20 to ensure that the node splitting has enough data support.

[0079] Then perform adaptive learning rate update: The initial learning rate is set to 0.05, and then it is dynamically adjusted according to the formula learning rate iteration times. As the number of iterations increases, the learning rate gradually decreases, allowing the model to learn quickly in the early stage of training and fine-tune in the later stage, improving training efficiency and stability.

[0080] Then perform L1 regularization constraint: Through the L1 regularization unit (penalty coefficient set to 0.01), the model parameters are constrained. This can effectively suppress the interference of abnormal values on the model, making the model pay more attention to the relationship between general and stable risk features and incidence probability, and enhancing the generalization ability of the model.

[0081] Then calculate the incidence probability: Based on the adjusted feature weights, improve the gradient boosting tree to calculate the node splitting of each tree based on the squared error loss function. Combine the prediction results of all trees to get the preliminary incidence probability of the object to be predicted.

[0082] Core risk factor ranking generation: Use the SHAP (SHapley Additive exPlanations) value analysis method to calculate the contribution of each risk feature to the prediction result of the incidence probability. The higher the contribution, the more critical the feature in prediction. According to the contribution from high to low, the risk features are sorted to generate the core risk factor ranking, to clarify which indicators (such as TNFAIP3 methylation rate, UVA-MED, etc.) are the main factors leading to high incidence risk of the object to be predicted.

Claims

1. A method for predicting the risk of onset of chronic actinic dermatitis based on a risk prediction model, characterized by, Comprise the following steps: S1: Obtain the photobiological index, immune function dynamic index, molecular marker combination index, clinical basic index and life behavior risk index of the object to be predicted within a specified period of time, and establish the corresponding data subsets respectively; S2: The data subsets are preprocessed by hierarchical preprocessing, then corrected and scored for reliability, and data with a score greater than or equal to a specified threshold are retained to form a data set; S3: Risk index extraction is performed on the data set: first filter low-variation irrelevant indicators, then gradually eliminate indicators with low prediction contribution from the remaining indicators, and finally retain a specified number of candidate indicators; S4: Calculate the correlation of each candidate indicator with CAD incidence risk: Pearson correlation quantifies linear correlation, mutual information value quantifies nonlinear correlation, and the comprehensive correlation of each indicator is calculated by weighting; S5: Sort the candidate indicators according to the comprehensive correlation from high to low, select indicators with a comprehensive correlation greater than or equal to a threshold to construct a risk feature vector, and assign weights to each dimension according to the comprehensive correlation; S6: Input the risk feature vector into the pre-constructed multi-modal fusion risk prediction model, and the multi-modal fusion risk prediction model outputs the incidence probability of the object to be predicted, and the incidence risk level and core risk factor ranking are obtained according to the incidence probability.

2. The method of claim 1, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model. It also includes a risk level calibration process, which is as follows: S7: According to the Fitzpatrick skin type standard, divide it into type III and type IV, determine it by the light-sensitive contact history or no light-sensitive contact history of the object to be predicted, set groups based on skin type and light-sensitive substance contact history, and set probability adjustment coefficients respectively: S8: Determine the group of the object to be predicted, obtain the corresponding adjustment coefficient, calibrate the incidence probability predicted by the model based on the adjustment coefficient, set the incidence probability boundary constraint, determine the final incidence risk level according to the adjusted probability value combined with the risk level division standard; S9: Based on the final incidence risk level and the core risk factor ranking, call the pre-set risk prevention and control measure table to perform corresponding risk prevention and control measures on the object to be predicted.

3. The method of claim 1, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The photobiological indicators in step S1 are the minimum erythema dose of long-wave ultraviolet light (UVA-MED), the minimum erythema dose of medium-wave ultraviolet light (UVB-MED) and the erythema regression time (ERT) measured by a sunlight simulator, the dynamic immune function indicators are the proportion of peripheral blood T cell subsets detected by flow cytometry, the CD3 + CD4 + / CD3 + CD8 + ratio and the Th1 / Th2 cytokine balance state, the combination of molecular markers includes the serum IFN-γ, IL-12, sFas / sFasL, skin tissue-specific lncRNA and TNFAIP3 gene promoter region methylation rate, the clinical basic indicators include age, gender, Fitzpatrick skin type and history of previous photosensitivity diseases, and the life behavior risk indicators include the average daily sun exposure time, the compliance of sun protection measures, the frequency of photosensitizer intake and the environmental photosensitivity load index.

4. The method of claim 1, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The specific process of step S2 is as follows: S21: Box-Cox transformation is performed on the photobiological index to eliminate dimension; Z-score method is used to remove outliers from the immune function dynamic index; Multiple imputation method is used to fill in missing values in the molecular marker combination index; Clinical basic index and life behavior risk index are converted into categorical variables by one-hot encoding; Standardize each index; S22: Verify the data distribution of each data subset by 3σ criterion, and correct the abnormal subset by local weighted regression; S23: Generate a 0-10 data reliability score based on the accuracy of the detection instrument and the stability of sample collection, retain data with a score greater than or equal to 8, and merge to obtain a standardized data set.

5. The method of claim 4, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The specific process of step S3 is as follows: S31: Obtain the index data of healthy subjects with the same clinical basic index as the object to be predicted as the healthy control group, and use the data set in step S2 as the CAD case group; S32: Calculate the overall variance, within-case group variance, within-control group variance, and between-group variance for each indicator, and calculate the F value through the ANOVA formula, F value = between-group variance / within-group variance; The following screening criteria are set for indicator screening: The overall variance of the indicator is greater than 0.1; The F value is greater than 3.85; The indicators that meet the above two points are retained, otherwise they are rejected; S33: Further indicator screening is performed by incremental indicator elimination method RFE.

6. The method of claim 1, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The specific process of calculating the correlation degree of each candidate indicator with the risk of CAD in step S4 is as follows: S41: Quantify the linear correlation between each candidate indicator and the risk of CAD by calculating the Pearson correlation coefficient; S42: Quantify the nonlinear correlation between each candidate indicator and the risk of CAD by calculating the mutual information value; S43: Calculate the comprehensive correlation degree based on the Pearson correlation coefficient and the mutual information value: use the weighted formula of linear correlation weight 0.6 + nonlinear correlation weight 0.4; S44: For all the candidate indicators after ANOVA screening, repeat steps S41-S43 to calculate the R2of each 综合 .

7. The method of claim 6, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The specific process of step S5 is as follows: S51: Calculate R for all candidate indicators calculated in step S4 综合 Select indicators with R ranked from high to low 综合 Construct risk features from indicators with R ≥ 0.6 S52: Based on the screened 8-12 indicators, construct a risk feature vector, and assign a basic weight to each dimension according to the comprehensive correlation degree.

8. The method of claim 7, wherein the method is a method of predicting the risk of developing chronic actinic dermatosis based on a risk prediction model, characterized by, The multi-modal fusion risk prediction model of step S6 is constructed by feature attention mechanism + improved gradient boosting tree algorithm, using the risk feature vector of historical cases as training samples, feature attention mechanism dynamically allocates the basic weight of each dimension in the risk feature vector, and the improved gradient boosting tree algorithm optimizes the model training efficiency by introducing adaptive learning rate; After inputting the risk feature vector of the object to be predicted into the multi-modal fusion risk prediction model, the multi-modal fusion risk prediction model outputs the incidence probability of the object to be predicted, and according to the incidence probability, the incidence risk level and the core risk factor ranking are obtained, the risk level includes low risk, medium risk and high risk.

Images