Development of an explainable machine-learning-based predictive model for cisplatin-induced nephrotoxicity using real-world data

Cisplatin (CDDP)-induced nephrotoxicity (CIN) is a major dose-limiting adverse effect. The early detection and precise prediction of CIN are essential for optimizing CDDP-based chemotherapy. In this single-center study, we developed novel machine learning-based predictive models for CIN using a large-scale clinical dataset incorporating demographic characteristics and time-series laboratory data. To address the limitations of previous studies, class imbalance was handled while preserving the full dataset, and the model performance was evaluated using an independent test dataset. Among the 1,705 enrolled patients, 551 (32%) experienced nephrotoxicity. Four models were constructed using random forest, support vector machine, logistic regression, and artificial neural network algorithms. The logistic regression model achieved the highest performance among the four constructed models with scores of 0.626 (F1.5), 0.791 (AUC-ROC), 0.702 (AUC-PR), 0.739 (accuracy), 0.645 (recall), and 0.188 (Brier score). Shapley additive explanation analysis demonstrated that baseline serum creatinine and serum creatinine levels on days 4–6 after CDDP administration were the key contributors to model predictions. In conclusion, we developed a machine learning-based model for predicting CIN using real-world clinical data. Following external validation, this model may have potential to support the early identification of patients susceptible to CIN.