Developing Predictive Models for Healthcare Outcomes Using Machine Learning and Electronic Health Records

Thesis Abstract

The increasing adoption of electronic health records (EHR) alongside advancements in machine learning (ML) techniques presents a significant opportunity to enhance predictive analytics in healthcare, ultimately improving patient outcomes and optimizing resource allocation. However, there remains a substantial knowledge gap in systematically developing, validating, and deploying robust predictive models that can accurately forecast healthcare outcomes using complex EHR data across diverse patient populations. This study aims to develop and evaluate machine learning-based predictive models that can reliably predict key healthcare outcomes such as hospital readmission rates, mortality risk, and disease progression, utilizing comprehensive EHR datasets. The specific objectives include (1) identifying critical clinical and demographic predictors within EHR data relevant to selected healthcare outcomes; (2) constructing multiple predictive models employing supervised learning algorithms, including random forests, support vector machines, and gradient boosting machines; (3) assessing model performance using metrics such as accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC); and (4) implementing explainability techniques like SHAP (SHapley Additive exPlanations) to elucidate feature contributions and enhance model interpretability for clinical decision-making. The research adopts a quantitative, cross-sectional design integrating retrospective EHR data from a tertiary hospital over a five-year period, encompassing approximately 150,000 patient records. The study's population comprises adult patients (>18 years) diagnosed with chronic conditions such as diabetes mellitus, cardiovascular diseases, or chronic respiratory illnesses. A stratified random sampling approach is employed to select a representative subset of 20,000 records, ensuring inclusion of diverse demographic groups and disease profiles. Data collection involves extracting structured data, including demographics, clinical history, laboratory results, medication lists, hospitalization details, and procedural codes, via automated data mining tools validated for accuracy. The study also incorporates unstructured clinical notes, processed through natural language processing (NLP) techniques to augment predictive features. Data pre-processing includes cleaning, normalization, feature engineering, and handling missing values through multiple imputation. Model training and validation utilize k-fold cross-validation, with hyperparameter tuning performed via grid search. Statistical analysis comprises descriptive, inferential, and comparative assessments using Python-based machine learning libraries such as scikit-learn and XGBoost. Theoretical grounding references the Health Belief Model and the Diffusion of Innovations Theory to examine factors influencing model adoption and integration into clinical workflows. Expected findings include the development of predictive models exhibiting AUC-ROC scores exceeding 0.80 across outcomes, with the gradient boosting algorithm demonstrating superior performance due to its capacity to handle heterogeneous data. Feature importance analyses are anticipated to highlight critical predictors such as age, comorbidity indices, laboratory values, and specific medication variables. The study also aims to demonstrate that explainability techniques can effectively identify key clinical features influencing individual risk assessments, thereby fostering clinician trust and facilitating model deployment. By comparing model performances, this research seeks to establish best-practice frameworks for deploying ML models within healthcare systems, emphasizing model transparency and interpretability. This study contributes new knowledge by systematically integrating advanced machine learning algorithms with comprehensive EHR data to improve healthcare outcome predictions, addressing existing limitations related to model interpretability and data heterogeneity. It offers a practical pathway for healthcare providers to leverage data-driven insights for proactive patient management, disease prevention, and resource planning. The main conclusion emphasizes the potential of ML-driven predictive analytics to transform clinical decision support systems, contingent upon rigorous validation, transparent algorithms, and clinician engagement. Recommendations include adopting standardized protocols for EHR data integration, investing in clinician training on AI literacy, and further longitudinal validation studies across different healthcare settings to enhance model generalizability and sustainability.