A research-use national early-warning study for Nigeria built on NCDC weekly surveillance data. Retrospective results are shown as development context; the binding 2026–27 test is prospective and must beat the strongest pre-specified persistence-or-seasonal comparator.
We expect the 2026–27 prospective study to end as inconclusive–underpowered, and we are saying so before the season starts.
Whether a week can be forecast depends on when NCDC publishes it. Under the frozen v1.1 protocol, the four-week forecast window for a week starts the day after that week ends, and a forecast counts only if it is logged before its window starts. So a week is eligible only if NCDC publishes its report by the Sunday that week ends. No evidence we hold shows NCDC publishing that fast: our own collector timed the fastest release at 6–8 days after the week ended, and an upload-time measure across 36 weeks of 2026 ranged from 9 to 44 days (median 17).
Expected eligible forecast windows in the primary window (2026-W45 to 2027-W17, 26 ISO weeks): 0 of 26 at the lag we measured, 0 of 26 in the best case and 0 of 26 in the worst case. The protocol needs at least 10 scoreable weeks.
An underpowered readout is a pre-specified outcome of the protocol (§11). It is not evidence that the model failed: it means the v1.1 design cannot be scored at NCDC's publication speed. We will publish it as that, and we will not change v1.1 to rescue it. A different forecast lead would be a new, separately versioned study.
Computed from the frozen v1.1 window rule, our first-observation ledger and the upload-time measure:
data/lassa/v1-1-eligibility-yield.json. The window has 25 weeks if NCDC numbers 2026 with 52 weeks; the answer is the same.
Latest directly verified 2026 national checkpoint, cumulative through Epi Week 40 (28 September – 4 October 2026): 8,414 suspected · 1,138 confirmed · 6 probable · 268 confirmed-case deaths (case fatality rate 23.6%; 24 states, 119 LGAs).
Source: Nigeria Centre for Disease Control and Prevention (NCDC), official Week 40 situation report ↗, Table 1, read directly from the PDF. These figures are a display checkpoint. The committed weekly series used by the prospective study currently runs through Epi Week 38; later weeks enter it only through the gated prospective cycle. None of this is part of the frozen 2020–2025 retrospective model-development dataset shown below. Newer weeks are published here only after direct primary-document verification.
Pre-season freeze complete.
Frozen at commit
under protocol .
All eight frozen files and the freeze manifest are published byte-identical in github.com/oluwafemidiakhoa/lassaai; VERIFY.md gives the commands to recompute every hash, and to regenerate the feature importance below.
These retrospective numbers were used to develop the model. They are not a prospective result, and the prospective study is judged only by the frozen rules above.
Research use only. Not a validated diagnostic, clinical decision tool, or replacement for NCDC field surveillance. No prospective result exists until the frozen readouts.
National model: will the next 4 weeks carry an above-median confirmed-case burden (an "elevated-transmission week")? Trained on real NCDC weekly data with a strict out-of-time hold-out (train ≤2023, test 2024). The prospective v1.1 study adds explicit seasonal and seasonal+persistence comparators.
Honest read. The 2024 retrospective hold-out AUROC was — versus — for the recent-4-week persistence baseline, and that difference was not statistically significant in the original analysis. A later pre-freeze challenger audit showed why persistence alone was too weak a scientific test: explicit seasonality is a strong comparator, and no challenger earned promotion. Protocol v1.1 therefore requires the prospective model to beat the strongest pre-specified persistence-or-seasonal comparator, with a paired 4-week moving-block confidence interval above zero and an 8-week sensitivity analysis. This is still not a breakthrough or a prospective result. The 2026–27 study is now frozen under protocol v1.1; the one-time manifest is committed, and no frozen scientific rule may be silently changed. Scored forecasts and later outcomes will be published in the public prospective receipt ledger.
Open-access preprint. The full method, results, and honest limitations are published open access: Forecasting Elevated Lassa Fever Transmission Weeks in Nigeria from National Surveillance Data: A Baseline-Benchmarked Proof of Concept (Zenodo, CC-BY, DOI 10.5281/zenodo.21122486). Code & data: github.com/oluwafemidiakhoa/lassaai.
Data: NCDC Weekly Epidemiological Reports (2020–2025), cross-validated against an independent extraction and NCDC published annual totals (agreement ±~1%), plus SORMAS individual-level data (2018–2021) for a per-state/clinical cross-check.
Limitations: Validated result is national; the per-state map below is illustrative pending real per-state weekly data (NCDC reports it in image tables). Confirmed counts under-report true incidence. The model beats a naive baseline only modestly in the original 2024 comparison, and the later audit did not establish stable superiority over explicit seasonality. Research use only — not a validated diagnostic, clinical decision tool, or replacement for NCDC field surveillance.
Enter LASV protein targets into GaiaLab's AI analysis pipeline. The viral pathogen mode routes GPC, NP, L, and Z through a Lassa-specific drug repurposing engine with known antiviral candidates pre-loaded.