Presentation (CTI)

OBJECTIVES: Predicting chemotherapy-induced nephrotoxicity is essential to prevent severe adverse drug reactions from treatment interruption. The study aimed to develop machine learning prediction models to support clinicians in preventing chemotherapy-induced acute kidney injury (AKI) and acute kidney disease (AKD).
METHODS: Patients receiving cisplatin, carboplatin, ifosfamide, methotrexate, pemetrexed, and gemcitabine were identified from the Taipei Medical University Clinical Research Database and the Big Data Center, Taipei Veterans General Hospital. Data from Shuang Ho Hospital, Taipei Veterans General Hospital, and Taipei Medical University Hospital were split into training and validation datasets, and Wanfang Hospital served as the testing dataset. Logistic regression, histogram-based gradient boosting classification tree, and light gradient boosting machine (LGBM) were employed. The main performance was evaluated by the area under the receiver operating characteristic curve (AUROC) and other performance metrics. Fairness was evaluated based on Equalized Odds (EO) across different sensitivity groups, and Shapley additive explanation (SHAP) was performed to explain the feature importance.
RESULTS: A total of 6,126 patients were included, with AKI incidence rates of 3.51%, 3.41%, and 8.28%, and AKD rates of 26.95%, 26.54%, and 32.00% in the training, internal, and external validation datasets, respectively. The LGBM demonstrated an AUROC of 0.762 and 0.718 in predicting AKI and AKD in external validation, respectively. The model achieved fairness based on EO across gender, age, and drug type. SHAP analysis revealed the top 5 features for AKD included the latest albumin, age, cancer stage, the difference between the latest and baseline estimated glomerular filtration rate, and the latest hemoglobin level. Features for AKI included the latest albumin level, diuretics, cisplatin dose, treatment interval, and alcohol consumption.
CONCLUSIONS: This study established machine learning models to predict chemotherapy-induced AKI and AKD, which could be incorporated in a clinician decision support system to aid prevention of chemotherapy-induced nephrotoxicity in clinical practice.