Machine learning-based prediction of PASI100 response to secukinumab in patients with psoriasis: a real-world study with SHAP interpretability analysis

TL;DR

This study uses machine learning to predict which psoriasis patients will achieve complete skin clearance with secukinumab, highlighting baseline disease severity as a key factor.

Contribution

Development of Random Forest and LightGBM models with SHAP interpretability for predicting PASI100 response to secukinumab in real-world psoriasis patients.

Findings

41.25% of patients achieved PASI100 response after 3 months of secukinumab treatment.

Baseline disease severity indicators (bIGA, bBSA, bPASI, bDLQI) were the most important predictors of PASI100 response.

Random Forest and LightGBM models showed moderate discriminative ability with testing AUCs of 0.757 and 0.761, respectively.

Abstract

Secukinumab, an interleukin-17A (IL-17A) inhibitor, has demonstrated significant efficacy in treating moderate-to-severe plaque psoriasis. Achieving complete skin clearance (PASI 100) is the ideal therapeutic goal. However, individual responses vary, and tools to accurately predict PASI 100 response in real-world settings are lacking.

In this retrospective study, we analyzed data from 11,134 psoriasis patients who were treated with secukinumab for 3 months. The dataset was randomly split into training (70%) and testing (30%) sets. Univariate analysis and LASSO regression were used for feature selection. Eight machine learning algorithms, including Random Forest, LightGBM, and Logistic Regression, were developed to predict treatment response. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC). SHapley Additive exPlanations (SHAP) analysis was employed to interpret the optimal model.

A total of 4,593 (41.25%) patients achieved PASI 100 response. The factors of Disease duration, BMI, bBSA, bPASI, bDLQI, Gender, bIGA, Education background, Job status, Comorbidity, Family history, Drug allergy history, Disease situation, Traditional systemic therapy, Medical insurance, Disease status and Biologic usage status were significantly associated with PASI 100 response (all p < 0.05), while others not. LASSO regression identified 5 key predictors, including Gender, bIGA, bBSA, bPASI and bDLQI. Among the algorithms, Random Forest (training AUC = 0.879, testing AUC = 0.757) and LightGBM (training AUC = 0.834, testing AUC = 0.761) demonstrated the best performance in those machine learning algorithms. SHAP analysis revealed that gender and baseline disease severity indicators (bIGA, bBSA, bPASI and bDLQI) were important predictors.

We successfully developed Random Forest and LightGBM-based prediction model for PASI100 response to secukinumab with moderate discriminative ability. Baseline disease severity emerged as the dominant predictor of complete skin clearance. These findings provide evidence-based support for personalized treatment goal setting and patient selection in clinical practice.

Figures