Mortality Prediction in Imbalanced Pediatric Oncology Data: A Cost-Sensitive Hybrid Resampling Framework Using Recurrent Deep Learning

Background: Severe class imbalance hampers mortality prediction in pediatric oncology, where clinical-improvement discharges greatly outnumber deaths and bias standard models toward the majority class. Objectives: We aimed to develop and evaluate a Cost-Sensitive Hybrid Resampling Framework (CSHRF) across recurrent deep-learning architectures to mitigate this bias and improve discharge-outcome classification. Methods: A retrospective cross-sectional study was conducted using electronic health record data from a National Institute of Health in Mexico City during 2023. We analyzed 1607 hospital-discharge records, comprising 1554 clinical improvements and 53 deaths. The CSHRF combines cost-sensitive learning, which amplified the mortality class by a factor of 2.5, with a fixed-seed hybrid resampling protocol embedded in a regularized funnel-shaped architecture (N → N/2 → N/4) trained with gradient clipping (clipnorm = 1.0). Simple Recurrent Neural Network (SimpleRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models were compared with a naïve recurrent baseline and standard tabular classifiers (Weighted Logistic Regression, Balanced Random Forest, and XGBoost) using stratified cross-validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), G-mean, sensitivity, specificity, precision, F1-score, mean absolute error (MAE), and root mean squared error (RMSE). Results: The naïve baseline showed limited mortality detection (sensitivity 0.150 ± 0.147). The proposed framework substantially improved minority-class detection, with CSHRF-GRU achieving sensitivity of 0.740 ± 0.145, specificity of 0.700 ± 0.076, G-mean of 0.714 ± 0.045, and AUC of 0.761 ± 0.075. Ablation analysis demonstrated that cost-sensitive weighting contributed most substantially to mortality detection, whereas hybrid resampling alone produced more modest gains. Among tabular baselines, Weighted Logistic Regression achieved competitive performance, with sensitivity of 0.684 ± 0.130 and G-mean of 0.767 ± 0.073. Precision remained low (0.079 ± 0.015) and probabilistic calibration errors were substantial (MAE 0.448; RMSE 0.488), reflecting the trade-off required to prioritize mortality detection under severe imbalance. Conclusions: The CSHRF improved discharge-outcome classification in a highly imbalanced pediatric oncology dataset and provides a methodological framework for data-constrained hospital environments. Given its high false-positive rate and probability overestimation, the approach should be interpreted as a discharge-related outcome classification model rather than a standalone clinical decision-support system, and future work should incorporate richer clinical and disease-specific variables to improve predictive performance and calibration.