EXPLORING INTERPRETABILITY IN MACHINE LEARNING MODELS FOR HEALTHCARE DECISION SUPPORT SYSTEMS

This project aims to study Machine Learning applications in health-care decision support systems. The main objective is to use an explainable method (SHAP/LIME) to understand how black-box models make their decisions, promoting transparency between systems decisions in such a critical context. The research problem was set as a binary classification problem using the MIMIC-IV-ED dataset, we intend to predict patients with chest pain will be discharged based on features collected during triage and first medical contacts. The pipeline integrates a comprehensive preprocessing step, where numerical features are standardized, and categorical features are one-hot encoded. The models tested include logistic regression, random forest, support Vector Machines, XGBoost, and LightGBM. To address the class imbalance, the pipeline compares a random under-sample and over-sample method. Hyperparameter optimization is performed using random search to choose the best parameters for each model, and model parameters are fine-tuned to enhance predictive performance. The primary goal is to create an efficient and accurate binary classification model, providing valuable insights for clinical decision-making in emergency department settings.