Toward reliable Lassa fever detection: A comparative analysis of ensemble and classical models under class imbalance constraints

Lassa fever remains a major public health concern in West Africa, causing an estimated 100,000 to 300,000 infections and over 5,000 deaths annually. Early detection is hindered by inadequate diagnostic infrastructure, symptom overlap with other febrile illnesses, and delays in laboratory confirmation. While machine learning offers potential for disease detection, existing models face challenges from class imbalance, redundant features, and suboptimal configurations. This paper compares four machine learning algorithms: Logistic Regression, Support Vector Machines (SVM), LightGBM, and Gradient Boosting, for Lassa fever detection using 20,062 clinical records from Nigeria's disease surveillance system (2017–2022). All four algorithms achieved similar accuracy, ranging from 75% to 76%. However, they showed different trade-offs between precision and recall. Logistic Regression, SVM, and Gradient Boosting used conservative, precision-focused strategies, achieving near-perfect or perfect precision (0.99–1.00) but missing about 32% of actual positive cases (recall = 0.68). LightGBM took the opposite approach, achieving perfect recall (1.00) by identifying every positive case, but at the cost of lower precision (0.75) and many false positives. LightGBM achieved the highest F1 Score (0.86), indicating the best balance between precision and recall among the models. These findings show that algorithm selection for Lassa fever detection involves trade-offs that cannot be resolved by focusing only on overall accuracy. LightGBM's perfect recall is more suitable for active outbreak surveillance, where missing any case is unacceptable. This study provides clear, evidence-based guidance for selecting algorithms for Lassa fever diagnostic systems and establishes a performance baseline for future research in this domain.