Abstract
<title>Abstract</title> <p>The lack of data and significant class imbalance between the groups in low-resource intensive care unit (ICU) settings are two major challenges to the development of machine learning models capable of predicting mortality and length of stay accurately. This paper assesses three synthetic data generation techniques-SMOTE-NC, CTGAN, and Variational Autoencoder-alongside three machine learning algorithms-Logistic Regression, Random Forest, and XGBoost-under real-only, synthetic-only, and hybrid training configurations. We analyzed a dataset of 10,669 patient records from five Ethiopian public hospitals. The original cohort (N=10,669) has shown severe class imbalance with 86.1% survival (13.9% mortality). To handle balanced model training, we undersampled the majority class, creating a balanced training set of 2,942 patients (50% mortality). We acknowledge that this required discarding 7,631 patients (72% of original cohort) and that performance metrics on this artificially balanced set do not reflect real-world class distributions. Hybrid models combining real and synthetic data improved performance, with CTGAN + XGB achieving the highest accuracy (0.998, 95% CI: 0.995–1.000) for mortality prediction. Random Forest with CTGAN outperfomed with MAE 5.46 days, 95% CI: 4.89–6.12 for length of 1 stay prediction). CTGAN generated higher-quality synthetic samples (Logistic Regression accuracy: 91.7% vs. 86.4%, p ¡ 0.001) but required 90× longer training time (187.4s vs. 2.1s). The data obtained from VAE had no predictive power (51.6% accuracy, 95% CI: 46.9–55.6%, no better than random guessing). Oxygen requirement (r = 0.327, p ¡ 1e 266) and SpO2 (r = 0.299, p ¡ 1e 222) emerged as the strongest mortality predictors, hemoglobin showed negligible association with mortality (r = 0.023, p = 0.016) while age was not statistically significant (p = 0.39). In conclusion, our results demonstrate that hybrid modeling with SMOTE-NC can be a potential solution for the creation of effective prediction models in settings with limited resources.</p>