Machine Learning for Clinical Decision Support: A Review of Feature Selection in Disease Prediction

Machine learning-based clinical decision support systems offer transformative potential for improving diagnostic accuracy and risk stratification, but adoption remains constrained by the challenge of identifying minimal clinically meaningful feature sets achieving adequate predictive performance while supporting interpretability and reducing data collection burden. This paper presents a comprehensive review of feature selection approaches applied to disease prediction machine learning models, examining filter methods, wrapper methods, embedded methods, and evolutionary optimisation including genetic algorithms, with systematic evaluation across liver disease, cardiovascular disease, diabetes, and oncological risk stratification domains. A consolidated best-practice framework integrating complementary filter and wrapper approaches within a cross-validated evaluation protocol is proposed. Genetic algorithm feature selection on the UCI Indian Liver Patient Dataset achieved six-feature XGBoost AUC-ROC of 0.841 with 40 percent feature count reduction. Comparative tables of feature selection methods and NLP approaches are included.