Abstract
Machine learning (ML) has the potential to drastically improve clinical decision-making by predicting diseases early, accurately, and based on data. This study evaluated and compared the performance of several machine learning models, including a feedforward neural network (FNN), XGBoost, Random Forest (RF), and Support Vector Machine (SVM), across four datasets: (i) clinical laboratory biomarkers for diabetes and cardiovascular disease, (ii) patient history and lifestyle factors for diabetes risk, (iii) symptom and ECG data for heart disease, and (iv) syndromic surveillance data for communicable disease. The AUC (area under the receiver operating characteristic curve) was used to assess model performance. The findings show that prediction accuracy is highly context-dependent: laboratory data produced modest accuracy (AUC = 0.62), patient history-based models performed well for diabetes (AUC = 0.84), heart disease prediction from symptoms and ECG performed best (AUC = 0.87), and communicable disease prediction from broad symptom data was poor (AUC = 0.49). This work is unique in that it conducts a systematic comparison analysis of numerous ML models across a variety of clinical data scenarios, offering vital insights into how different data modalities and modelling approaches affect diagnostic accuracy. These findings highlight the significance of dataset selection, context-aware feature design, and rigorous model tuning for deploying effective clinical ML models, especially in resource-constrained healthcare settings. Overall, the study found that prediction success is highly influenced by data type, modelling approach, and clinical context, underlining the importance of tailored, data-specific solutions in a variety of healthcare situations.