Development and multiple visualization methods for the therapeutic effects prediction model of five-flavor Sophora Flavescens enteric-coated capsules in the treatment of active ulcerative colitis: A study on model development and result visualization

Introduction

The aim of this study was to develop a Therapeutic Effects Prediction Model (TEPM) for the treatment of active ulcerative colitis (UC) using Five-flavor Sophora Flavescens Enteric-coated Capsules (FSEC). This study also aimed to systematically review various visualization methods for the TEPM results and present the model results of FSEC as an example.

Methods

274 patients were randomly assigned to the training and testing datasets in a 7:3 ratio. We employed Least Absolute Shrinkage and Selection Operator (LASSO) regression to select predictive factors and constructed TEPM using logistic regression to assess the probability of disease remission. We assessed model performance by the area under the curve (AUC) and calibration curve. We utilized interactive nomograms, online calculators, scoring systems, graphical scoring tables, as well as the SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) methods to present the model results.

Results

LASSO regression selected several predictors, including erythrocyte sedimentation rate, age, disease type, microscopic bleeding, pus, bridge, disease location, and pain. The AUC of the testing datasets was 0.699, and the calibration curve showed poor performance. The interactive nomogram, online calculator, and the SHAP method were suitable for datasets with predominantly continuous predictors, while scoring systems, graphical scoring tables, and the LIME method might be more appropriate for datasets with fewer continuous predictors. Physicians, researchers, and policymakers could benefit from detailed visualizations using interactive nomogram, the SHAP method, and the LIME method. Scoring systems, graphical scoring tables, and online calculator were available to the general public and non-experts. Scoring systems, graphical scoring tables, and online calculator could provide an overview of the model prediction results, while interactive nomogram, the SHAP method, and the LIME method were recommended for illustrating the complexity and rationality of the model prediction results.

Conclusion

Our study demonstrated that TEPM could predict the potential of FSEC to induce disease remission in patients with active UC. However, the poor calibration curve might be due to the limited sample size. Larger-scale multicenter studies will be needed in the future. Selecting an appropriate visualization method for TEPM should be based on the datasets, audience, and research objectives.