Forecasting Elevated Lassa Fever Transmission Weeks in Nigeria from National Surveillance Data: A Baseline-Benchmarked Proof of Concept

Background: Lassa fever is endemic in Nigeria, which reports the highest documented burden globally. The Nigeria Centre for Disease Control and Prevention (NCDC) publishes weekly situation reports, but the public-health response remains largely reactive. A tool that flags weeks of elevated transmission a month ahead could support resource pre-positioning, but only if it is honestly benchmarked against trivial baselines and validated on real data. Methods: We assembled a national weekly Lassa fever incidence series for Nigeria (2020-2025) from NCDC Weekly Epidemiological Reports, cross-validated against an independent extraction and against NCDC's published annual totals (agreement within ~1%). We engineered nine features from recent case history (lagged and rolling confirmed counts, trend) and calendar seasonality, and trained an XGBoost classifier to predict whether the next four weeks would carry an above-median confirmed-case burden ("elevated-transmission week"). The model was evaluated on a strict out-of-time hold-out (training years up to and including 2023; test year 2024) and benchmarked against naive persistence baselines. Individual-level SORMAS surveillance data (2018-2021) provided a per-state and clinical cross-check. Results: On the 2024 hold-out, the model achieved AUROC = 0.880 (precision 0.93, recall 0.78, F1 0.85) for a non-degenerate target (56% of weeks positive). A naive baseline using only the recent 4-week case average achieved AUROC = 0.849, so the model improved on trivial persistence by a modest ~0.03 AUROC. The strongest predictor was dry-season timing (gain importance 53%), consistent with the well-established November-April Lassa season and with the SORMAS data (peak incidence in February). Conclusions: Real national surveillance data support a modest but genuine ability to anticipate elevated Lassa transmission weeks in Nigeria, driven chiefly by seasonality and recent incidence. We make no claim of exceptional forecasting skill: the model beats a naive baseline only slightly, and the true test is prospective. The contribution is an open, reproducible, honestly-benchmarked pipeline, with append-only forecast-logging infrastructure in place to enable prospective validation. All data are from public sources with full provenance.