Lassa fever prediction using data from Nigeria's national surveillance registry

Lassa fever is a dangerous viral illness that is endemic in Nigeria and clinically difficult to separate from malaria and other febrile conditions in its early phase, which delays diagnosis and treatment. Machine learning has been proposed as a triage aid, yet several reported successes rest on small samples. Using the openly available national surveillance dataset of 20,027 suspected cases, of which 15.8% were laboratory-confirmed, we predicted laboratory confirmation from pre-laboratory bedside features alone. Synthetic oversampling was confined to training folds. Nine classifiers, including soft-voting and stacking ensembles, were compared by stratified five-fold cross-validation and a held-out test set using metrics suited to imbalanced data, alongside calibration, SHAP, missing-data and decision-threshold analyses. Discrimination from bedside features was modest, with the area under the receiver-operating-characteristic curve reaching 0.719 in cross-validation and 0.709 on the test set. Logistic regression, selected by cross-validation, achieved the best balanced accuracy (0.652) and F1 score (0.377), and the higher test value of the soft-voting ensemble was not statistically distinguishable from it. Gradient-boosted ensembles reached the highest accuracy (about 0.85) but recall below 0.22. The strongest predictors of the logistic regression were ordinary symptoms and documented source-case contact, and its performance was unchanged in fully documented records (0.702). Symptom-based prediction offers a bounded triage aid rather than a substitute for laboratory confirmation, and the study provides a reproducible benchmark for machine learning in neglected tropical disease diagnosis. The code is available at https://github.com/EgwuSamuel/lassa-fever-benchmark.