Exploring Various Diseases Based on Regularized Machine Learning Approaches

Abstract

The early diagnosis of physical and mental health issues is critically important in avoiding complications and enhancing patient outcomes. This thesis investigates the predictive po tential of regularized machine learning methods that predict and quantify the risk of disease; especially in eating disorders and stroke, on three different datasets. The thesis focuses on trade-offs between the performance and interpretability of regularized models (Lasso, Ridge, ElasticNet) as well as comparing them to non-regularized methods (ensemble trees, Support Vector Regression, etc). In the prediction of eating disorders, four preprocessing techniques: standard scaling, PCA, power transformation, and combination of these meth ods were compared through different models. Ensemble methods and SVR in every case produced a higher predictive accuracy, but regularized models were more interpretable and performed consistently across preprocessing conditions. The use of the same set of models on synthetic mental health data revealed lower accuracy across all models but especially the tree-based and SVR models. The regularized models showed, however, more consis tent performance, revealing the limitations of synthetic data in storing complex correlations between variables. Partial Least Squares (PLS) regression achieved the best accuracy when predicting stroke. Regularized models were used on PLS component and showed similar performance supporting their usefulness in terms of capturing important signal patterns and model simplicity. In sum, this thesis emphasizes the comparative advantage of regular ized machine learning in disease prediction extending to balanced accuracy, stability, and interpretability.