An interpretable AutoML-based prediction model for enteral nutrition intolerance in severe pulmonary tuberculosis...

This study aims to construct an AutoML-based predictive model for enteral nutrition intolerance (ENI) in severe pulmonary tuberculosis (PTB) patients and develop a visualized clinical decision support system to inform personalized nutrition management.

Using a multicenter retrospective cohort design, clinical data from 645 severe PTB patients were analyzed. An Improved Dimension-wise Gaussian-mutated Chaotic Divine Religions Algorithm (IDRA) was proposed and integrated into the AutoML framework to simultaneously optimize feature selection and hyperparameter tuning. SHapley Additive exPlanations (SHAP) were employed for both global and individual-level model interpretability, including waterfall plots, force plots, and dependence plots. A visualized clinical decision system was subsequently developed based on the optimal model.

(1) IDRA demonstrated superior convergence speed and lower local optima entrapment risk on CEC2022 benchmark functions; (2) Eight key predictors were identified and ranked by SHAP importance: hypoalbuminemia, anti-TB drug regimen (rifampicin-containing or not), formula type (short peptide vs. intact protein), age, BMI, EN initiation time, tube type (nasoenteric vs. nasogastric), and Chinese herbal application; (3) The optimal AutoML model achieved an AUC of 0.910 in internal validation and 0.880 in external validation; (4) Individual-level SHAP analysis of representative high-, medium-, and low-risk cases using waterfall plots and force plots demonstrated clinically coherent differential model outputs under varying feature combinations, corroborating global interpretation results.

This study established an interpretable AutoML-based prediction model and clinical decision system for enteral nutrition intolerance in severe PTB patients. Its core innovation lies in creating a transparent, user-friendly, and high-efficacy precision risk-assessment paradigm that provides both population-level feature attribution and case-level decision logic visualization, offering a novel approach to individualized EN tolerance prediction.