OAS-XGB: An OptiFlect Adaptive Search Optimization Framework using XGBoost to predict Length of Stay for CAD Patients

Accurately predicting the length of hospital stay for patients with coronary artery disease is important for effective healthcare resource management and enhancing patient outcomes. However, to achieve a high predictive performance in machine learning, one of the key challenges is to balance exploration and exploitation during the hyperparameter tuning process. To address this challenge, this work proposes an OptiFlect Adaptive Search (OAS) optimization with XGBoost, an advanced framework designed to enhance the predictive performance of XGBoost models. Initially, data preparation and preprocessing are performed, followed by representation learning using Variational Autoencoders with t-SNE to capture complex data patterns through low-dimensional embeddings. With the transformed latent features, the proposed framework initiates a random search and then transitions to a Gaussian Process-based Bayesian optimization, employing an adaptive switching strategy to achieve an optimal balance between exploration and exploitation. The experiment is performed on three comprehensive datasets, i.e., MIMIC-III, MIMIC-IV, and eICU. Across these datasets, the OAS optimization algorithm achieved the best log-loss. Utilizing the optimal hyperparameters identified from OAS, the XGBoost model was retrained, resulting in superior performance with an accuracy of 97.92% on MIMIC-III, 98.99% on MIMIC-IV, and 96.58% on eICU. This model also outperformed other baseline models, achieving the highest overall and class-specific metrics, including precision, recall, F1-score, AUC-ROC, and AUPRC, while maintaining the lowest overall inference time as evaluated across nested five-fold cross-validation. The SHapley Additive exPlanations are also performed for the OAS-XGB model's interpretability by identifying crucial predictive features, helping healthcare professionals optimize resource allocation and enhance care quality.