Development and validation of an interpretable machine learning model for predicting chemotherapy-induced neutropenia in small cell lung cancer: a web-based clinical decision support tool

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objective To develop and validate an interpretable machine learning model for predicting grade ≥3 chemotherapy-induced neutropenia (CIN) in patients with small cell lung cancer (SCLC), and to establish a web-based clinical decision support tool for individualized risk assessment. Methods We conducted a retrospective observational analysis of prospectively collected clinical data and enrolled 209 patients with SCLC who underwent 840 chemotherapy cycles between 2019 and 2022 from Tianjin Medical University General Hospital. Patients rather than chemotherapy cycles were randomly assigned into the training (80%) and testing (20%) cohorts. Feature selection was performed using least absolute shrinkage and selection operator regression and the Boruta algorithm. Machine learning algorithms were compared using patient-level 5-fold GroupKFold cross-validation. Hyperparameters were optimized using Optuna, and model interpretability was evaluated using Shapley additive explanations, Local interpretable model-agnostic explanations, Partial dependence plots. Three clinically oriented probability thresholds were further assessed to support different clinical decision-making scenarios. A web-based clinical decision support tool was subsequently developed. Results CatBoost achieved the highest cross-validation performance while demonstrating the most stable results across folds, and was selected as the final model. Five predictors were selected for model development, including chemotherapy cycle, chemotherapy regimen, prophylactic medication, absolute neutrophil count (ANC) and rescue treatment in the prior cycle. In the independent testing cohort, the optimized CatBoost model achieved an AUC of 0.785 (95% CI, 0.706–0.865), with good calibration and favorable clinical net benefit. Threshold-specific analyses illustrated the expected sensitivity–specificity trade-offs for exploratory screening- and confirmatory-oriented operating points. Interpretability analyses identified early chemotherapy cycles, rescue treatment in the prior cycle, absence of prophylactic medication, VP16 plus platinum regimen, and lower baseline ANC as predictors associated with higher CIN risk in this dataset. The model was further deployed as an interactive web-based application, available at http://39.96.172.15:8080/ . Conclusions An interpretable CatBoost model was developed for predicting grade ≥3 CIN in patients with SCLC. The integration of the model with a web-based clinical decision support tool may facilitate individualized risk stratification.