LassaAI — One-Page Summary

Model Performance (National model, 2024 out-of-time hold-out)

0.880

Model AUROC

0.849

Naive baseline AUROC

0.93

Precision

0.85

F1 Score

±1%

vs NCDC published

Target: whether the next 4 weeks carry above-median confirmed burden. Trained on real, cross-validated NCDC weekly data (train ≤2023, test 2024). The model beats a naive persistence baseline by only ~0.03 AUROC; the dominant predictor is dry-season timing. All metrics are retrospective — prospective validation is the pre-registered next step.

What LassaAI Does

National Early-Warning Model

Predicts elevated (above-median) national transmission over the next 4 weeks
Trained on real NCDC weekly data, 2020–2025 (cross-validated ±~1% vs published totals)
AUROC 0.88 vs 0.85 naive baseline — modest, honest, seasonality-driven
Per-state view is illustrative pending real per-state weekly data (future work)
Live dashboard: gailabai.com/lassa

Clinical Copilot

Literature-informed probability score over 20 clinical and epidemiological inputs, for healthcare workers at point of care
Symptom weights informed by published Lassa fever clinical literature — not independently validated
Freely accessible: gailabai.com/lassa-copilot
Always recommends RT-PCR confirmation

Dataset

Item	Value
Time period	2020–2025 (national weekly)
Observations	313 epidemiological weeks
Confirmed cases (2020–25)	6,456
Case source	NCDC Weekly Epidemiological Reports (cross-validated)
Per-state / clinical	SORMAS 2018–2021 (CC-BY, DOI 10.5281/zenodo.7309567)
Elevated-week rate	56% (balanced, non-degenerate target)
Preprint (open access)	Zenodo, CC-BY · DOI 10.5281/zenodo.21122486

Top Feature Importance

Dry-season flag

52.6%

Rolling trend (4−8 wk)

8.8%

Epidemiological week

7.5%

Cases lag 1 week

7.2%

8-wk rolling mean

6.8%

4-wk rolling mean

4.9%

Cases lag 8 weeks

4.2%

Cases lag 2 weeks

4.1%

Cases lag 4 weeks

3.9%

Dry-season timing dominates (52.6% gain importance) — the model is learning the real November–April Lassa season, not a data artefact. Recent case history and trend supply the rest. The national model uses no meteorological covariates; national-average weather is not epidemiologically meaningful.

Proposed Research Partnership

We are preparing a NIH Fogarty International Center R21 application (up to $275,000 / 2 years) with two aims:

Aim 1 (Year 1): Prospective validation with NCDC — 12-month blinded prediction logging vs. confirmed case outcomes
Aim 2 (Year 2): Clinical validation of the Copilot at a Nigerian Lassa fever treatment centre (target: ISTH Irrua or FMC Owo)

Application deadline: October 16, 2026 (LOI due September 16, 2026)

Most-Affected States (endemic ranking)

From the individual-level SORMAS 2018–2021 data, the endemic states rank Edo > Ondo > Ebonyi > Bauchi, consistent with published NCDC reporting. Edo (ISTH Irrua) and Ondo (FMC Owo / ILFRC) host the principal treatment and reference centres.

Validated per-state weekly case counts are not yet available — NCDC publishes them as image tables not amenable to extraction. Obtaining structured per-state data (proposed in partnership with NCDC) to build and prospectively validate a per-state model is the immediate next step. No per-state cumulative totals are shown here because none have been validated.

Key Limitations (Disclosed)

(1) Retrospective only — prospective validation not yet complete. (2) Confirmed cases only — true Lassa incidence substantially higher. (3) State-capital weather — rural microclimates may differ. (4) Reporting variability — testing capacity varies by state and year. (5) Clinical Copilot — not validated in a prospective clinical study; not a diagnostic device.

Not a medical device. LassaAI is a research and clinical decision support tool. It has not been approved by any regulatory authority. All clinical decisions must be made by qualified healthcare professionals with RT-PCR laboratory confirmation. For research and public health purposes only.