This was a retrospective cohort study utilizing real-world clinical data from a multicenter database. The patients with psoriasis who received secukinumab treatment at the Chinese Psoriasis Standardized Diagnosis and Treatment Center database from June 2020 to September 2024, and all patients sign informed consent forms upon entry into the database. Inclusion criteria were: (1) diagnosis of plaque psoriasis; (2) treatment with secukinumab; and (3) availability of baseline and follow-up PASI scores. The primary outcome was achieving PASI 100 (100% improvement from baseline PASI) after a defined treatment period. Baseline characteristics included demographic data (age, gender, BMI, education, job, marriage, region), disease characteristics (duration, disease status, family history, comorbidities, history of malignancy, allergy history), and clinical assessments (bPASI, bBSA, bDLQI, bIGA). Treatment history (topical therapy, biologic usage status, traditional systemic medications) and medical insurance status were also recorded.

Patients were eligible for inclusion if they met the following criteria: (1) age ≥18 years; (2) clinical and/or histopathological diagnosis of plaque psoriasis; (3) received standard-dose secukinumab treatment (300 mg subcutaneously at weeks 0, 1, 2, 3, and 4, followed by more than 3 months maintenance); and (4) had complete baseline and follow-up data available. Exclusion criteria included: (1) concurrent diagnosis of other psoriasis subtypes (pustular, erythrodermic, or guttate psoriasis); (2) pregnancy or lactation; (3) severe systemic comorbidities that could interfere with efficacy assessment; and (4) loss to follow-up or substantial missing data.

The primary outcome was PASI100 response, defined as achievement of 100% improvement from baseline PASI score (complete skin clearance) at the designated assessment time point.

Candidate predictor variables were selected based on clinical relevance and data availability, encompassing five domains: (1) Demographic characteristics: age, sex, body mass index (BMI), education level, occupation, marital status, and insurance status. (2) Disease-related characteristics: disease duration, baseline body surface area involvement (bBSA), baseline PASI (bPASI), baseline Investigator Global Assessment (bIGA), baseline Dermatology Life Quality Index (bDLQI), disease status (new-onset vs. recurrent), family history of psoriasis, presence of comorbidities, and history of malignancy. (3) Allergy history: history of drug allergy and history of allergic diseases. (4) Treatment history: prior topical therapy, prior conventional systemic therapy (including methotrexate, cyclosporine, and acitretin) and biologic usage status (biologic-experienced vs. biologic-naïve). (5) Lifestyle factors: smoking status. (6) Geographic factors: China region (including North area and South area).

Continuous variables were compared between PASI100 response and non-response using independent samples t-test. Categorical variables were compared using Pearson’s chi-square test. Variables with p < 0.01 in univariate analysis were considered for inclusion in the feature selection process. To reduce overfitting risk and identify the most informative predictors, LASSO regression with 10-fold cross-validation was performed. The optimal regularization parameter (λ) was determined based on the λ.1se criterion (the largest λ within one standard error of the minimum cross-validation error). The relationship between the number of features and model AUC was plotted to guide final feature selection, balancing predictive performance, clinical interpretability, and practical applicability.

The dataset was randomly partitioned into training (70%) and testing (30%) sets, stratified by outcome. Continuous variables were standardized using z-score normalization. Categorical variables were encoded using one-hot encoding. Eight ML algorithms were implemented: Decision Tree (DT), Random Forest (RF), Multi-layer Perceptron (MLP), Radial Basis Function Support Vector Machine (RBF-SVM), eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Light Gradient Boosting Machine (LightGBM). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), with higher values indicating better discriminative ability. AUC values were compared between training and testing sets to assess model generalizability. An AUC of 0.5 indicates no discriminative ability (equivalent to random guessing), while 1.0 indicates perfect discrimination.

SHAP (SHapley Additive exPlanations) analysis was applied to the best-performing model to interpret feature contributions. SHAP values quantify each feature’s marginal contribution to individual predictions based on Shapley values from cooperative game theory (18). SHAP summary plots were generated to visualize feature importance rankings and the direction of feature effects. SHAP dependence plots were constructed to examine the relationship between individual feature values and their SHAP contributions.

All analyses were performed using R 4.5 and Python 3.9 version software. Continuous variables were compared using the two-sample t-test, and categorical variables using the Chi-squared test. To reduce overfitting and remove redundant variables, the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to variables with p < 0.01 in the univariate analysis. The optimal penalty parameter (λ) was determined via 10-fold cross-validation based on the 1-standard-error (1-se) rule. Machine learning models were implemented using scikit-learn 1.0, XGBoost 1.5, and LightGBM 3.3. SHAP analysis was conducted using the shap 0.40 package. A two-sided p < 0.05 was considered statistically significant.

A total of 11,134 patients with plaque psoriasis who received secukinumab treatment were included in the analysis. Among them, 4,593 patients (41.25%) achieved PASI100 response, while 6,541 patients (58.75%) did not. The baseline characteristics of the two groups are presented in Table 1. Among continuous variables, disease duration was significantly longer in non-response compared to response (p = 0.024). Additionally, BMI was significantly higher in non-response than response (p = 0.007). Notably, indices of baseline disease severity, including bBSA, bPASI, and bDLQI were consistently higher in non-response, with all comparisons yielding p-values < 0.001. Age did not differ significantly between the two groups (p = 0.059). For categorical variables, gender distribution revealed a small but statistically significant difference (p < 0.001), with 32% of response being female compared to 35% of non-response. The baseline disease severity as measured by bIGA differed markedly between groups (p < 0.001). Significant group differences were also found in socioeconomic variables; education background (p < 0.001) and job status (p < 0.001) differed significantly. Regarding clinical history, comorbidity status showed a modest difference (p < 0.001), while history of malignancy did not demonstrate significant difference (p = 0.2). Family history of psoriasis showed a strong association with PASI100 response (p < 0.001). Smoking habits were not significantly associated with PASI100 response (p = 0.089). In contrast, drug allergy history exhibited significant differences (p < 0.001), whereas allergic disease history was similar across groups (p > 0.9). The baseline disease situation (stable, worsening, relieving) was significantly different between groups (p < 0.001). Topical therapy acceptance rates were almost identical in both groups (p = 0.4). However, traditional systemic therapy was significantly less common among response (p = 0.003). Medical insurance status also revealed significant differences; those with insurance constituted 98% of non-response compared to 97% of response (p = 0.027). There were significant differences in disease status (p < 0.001), with 91% of non-response classified as recurrent compared to 87% of response. The biologic usage status was also notably different, with a higher proportion of biologic-experienced patients among non-response (p = 0.0335). Lastly, when considering geographic factors, the China region did not demonstrate significant differences in PASI100 response between non-response and response (p = 0.3).

Overall, variables reflecting baseline disease severity (bBSA, bPASI, bDLQI, bIGA), as well as gender, education background, job status, family history, drug allergy history, disease situation, traditional systemic therapy, medical insurance status, disease status and biologic usage status were demonstrated statistically significant differences between PASI100 response and non-response and were considered candidates for subsequent feature selection and model development (Table 1).

Variables with p < 0.05 in univariate analysis were entered into LASSO regression for feature selection. Figure 1A shows the relationship of the binomial deviance (measured on the y-axis) versus the logarithmic transformation of the regularization parameter [log(Lambda), x-axis] and the number of features included in the model, and the optimal binomial deviance point indicated by 14 features. The maximum AUC was achieved with 30 features, while the λ.1se criterion selected 14 features (Figure 1B). Considering model performance, clinical interpretability, and practical applicability, 5 core predictive variables were ultimately retained: Gender, bIGA, bBSA, bPASI and bDLQI.

The performance estimates of several machine learning models across different evaluation metrics were used, including accuracy, balanced accuracy, detection prevalence, F-measure, L-index, kappa (kap), Matthew’s correlation coefficient (mcc), negative predictive value (npv), positive predictive value (ppv), precision, recall, ROC AUC (roc_auc), sensitivity (sens), and specificity (spec). The results indicate that the performance of different models is relatively consistent, with most models exhibiting estimate values around the 0.7 mark across key metrics. Notably, the Random Forest and LightGBM models demonstrated particularly robust performance, achieving among the highest estimates across various metrics (Figure 2A). A summary of the average performance across models were showed in Figure 2B, showcasing average estimate values for each model. The bar chart demonstrates that both Random Forest and XGBoost scored were 0.76, respectively, indicating their superior predictive capabilities. Other models, including MLP, DT, and RBF-SVM, achieved average scores close to 0.72, while Elastic Net and K-Nearest Neighbors rounded out the performance with estimates of approximately 0.71 (Table 2; Figure 3).

SHAP analysis was applied to the Random Forest model to interpret feature contributions. Figure 4A illustrated the feature impact direction derived from a Random Forest model, presented as a bar graph that depicted the mean SHAP values. This graph highlighted the average contribution of each feature to the model’s predictions. Notably, Gender and bBSA emerged as the most influential features, with positive contributions indicating that increases in these values correlated with a higher probability of achieving a PASI100 response. In contrast, features such as bPASI, bIGS, and bDLQI demonstrated negative contributions, suggesting that higher values in these metrics decreased the likelihood of achieving PASI100. Figure 4B presented a beeswarm plot that showcased the SHAP feature importance for the Random Forest model. Each point in the plot represented the SHAP value for an individual feature across various samples and was color-coded to reflect the feature value. Importantly, Gender and bBSA exhibited substantial positive SHAP values of 0.188 and 0.121, respectively, signifying their significant impact on the model’s predictions and their positive influence on the likelihood of achieving PASI100. Conversely, features like bPASI, bIGA, and bDLQI displayed lower SHAP values, indicating a relatively minor but still negative effect on the prediction. Figure 4C displayed a force plot for Sample 76, outlining the model’s prediction process. The actual outcome for this sample was classified as “No” for achieving PASI100, which aligned with the model’s prediction of the same. The SHAP values illustrated how each feature contributed to this prediction: bBSA (45) provided a positive contribution (+0.128), while Gender (1, representing female) also contributed positively (+0.127). Other features, such as bPASI (20.3) and bDLQI (20), made minor positive contributions, ultimately leading to a final prediction score of f(x) = 0.978. In Figure 4D, the force plot for Sample 21 revealed a different scenario, as this sample was predicted to achieve PASI100 (“Yes”), coinciding with the actual outcome. Here, bBSA (4) exhibited a significant negative contribution (−0.22), while Gender (2, representing male) also had a negative impact (−0.179). Other features, including bDLQI (12), bPASI (12), and bIGA (3), similarly provided negative contributions, which together resulted in a final prediction score of f(x) = 0.

SHAP analysis was applied to the Light Gradient Boosting Machine model to interpret feature contributions. Notably, Gender had the highest mean SHAP value, indicating it was the most influential factor in increasing the prediction likelihood. Other influential features included bPASI and bIGA, both of which also contributed positively. In contrast, features such as bBSA and bDLQI were less influential, with bDLQI showing a negative mean SHAP value, suggesting it negatively impacted the prediction (Figure 5A). As illustrated in Figure 5B, gender demonstrated the highest SHAP value among the features, with a mean SHAP score of 0.666, signaling a strong positive influence on predictions. bPASI and bBSA followed, with SHAP values of 0.592 and 0.315, respectively. bIGA had a moderate positive impact (0.249), while bDLQI exhibited a negative value of −0.102, indicating that an increase in bDLQI was associated with a lower likelihood of the predicted positive outcome. The sample 139 was displayed as Figure 5C, demonstrating the model’s prediction mechanics. The actual outcome for this sample was categorized as “Yes,” and the model also predicted “Yes.” The contributions of various features to the prediction were highlighted. Positive contributions were shown in yellow, while the overall prediction score was indicated on the right side of the plot. Specifically, bPASI contributed +0.535, Gender contributed +0.968, bBSA contributed +0.436, and bDLQI contributed +0.436. The expected base value, denoted as E[f(x)], was −0.469, summing the contributions to yield a final prediction score of f(x) = 1.43, confirming a positive prediction. The other Sample 1,633 was presented as Figure 5D, illustrating a different prediction trajectory. The actual outcome for this sample was “No,” and the model similarly predicted “No.” In this case, the contributions were visualized, with negative contributions shown in purple. Here, bBSA had a significant negative impact (−0.728), while bPASI (−0.407), bIGA (−0.344), and Gender (−0.907) all contributed negatively. The expected base value E[f(x)] was also −0.469, and the final prediction score was f(x) = −0.876, indicating a strong prediction that aligned with the actual outcome.