Manuscript · Research Presentation
May 2026
Research Presentation
Predictive Modeling of Typhoid Incidence in Nepal
Under extreme climate-change scenarios using machine learning
Kritika Baral
Department of Health Informatics · Kathmandu University · Dhulikhel, Nepal
ORCID 0009-0007-1517-1063 · kritka.baral@gmail.com
Can climate-induced flood events, precipitation, temperature, and humidity predict typhoid incidence across Nepal’s seventy-seven districts well enough to make public-health action anticipatory rather than reactive?
The setting
Nepal records between eighty thousand and one hundred thousand clinically diagnosed typhoid cases annually, concentrated in the flood-prone Terai plains. Existing surveillance is reactive: cases are counted after presentation — leaving no lead time for water purification, chlorination, or hygiene campaigns.
77 Administrative districts of Nepal
108 Months, January 2015 – December 2023
7,327 Valid district-month observations
The first national-scale, district-level, multi-model machine-learning study of typhoid fever in Nepal, integrating four validated open-access national datasets within a single analytical framework.
| Source | Variable | Provider |
|---|---|---|
| HMIS | Monthly outpatient typhoid cases | Ministry of Health and Population |
| CHIRPS | Precipitation | Climate Hazards Group |
| ERA5-Land | Temperature, humidity | ECMWF · Copernicus |
| DRR Portal | Flood event counts | NDRRMA |
Bikram Sambat to Gregorian calendar conversion; CBS three-digit district-code reconciliation; area-weighted zonal aggregation of climate grids to district polygons.
Eleven predictors, grouped into five functional categories. All lag and rolling features are computed within each district to prevent cross-district data leakage.
| Category | Features |
|---|---|
| Autoregressive | Prior month’s case count |
| Lagged climate | Precipitation · Temperature · Humidity · Flood events |
| Cumulative climate | Three-month rolling precipitation · temperature |
| Seasonal indicator | Monsoon flag (June–September) |
| Cyclical encoding | Sine and cosine of calendar month |
Target variable: \(\log(1 + \text{cases})\) to stabilize variance over the heavy right tail.
| Model | Architecture |
|---|---|
| Random Forest | Bootstrap-aggregated decision trees · 300 trees · depth 10 |
| Extreme Gradient Boosting | Regularized boosting · 300 estimators · depth 4 · L1 + L2 |
| Multilayer Perceptron | Feedforward neural network · hidden layers (128, 64, 32) · ReLU |
Strictly chronological train–test partition. The most recent twelve months are held out for testing; the preceding eighty-four months are used for training and hyperparameter selection.
| Feature | Pearson correlation with \(\log(\text{cases})\) |
|---|---|
| Prior month’s case count | 0.731 |
| Lagged mean temperature | 0.599 |
| Three-month rolling temperature | 0.559 |
| Lagged precipitation | 0.313 |
| Monsoon indicator | 0.245 |
| Lagged humidity | 0.180 |
| Lagged flood events | 0.122 |
Autoregression dominates. Lagged temperature is the strongest climate driver — consistent with the multi-pathway influence of temperature on pathogen biology, vector dynamics, and food safety.
| Model | RMSE | MAE | R² | MAPE |
|---|---|---|---|---|
| Random Forest | 131.43 | 72.10 | 0.8563 | 41.48 % |
| XGBoost | 129.07 | 69.50 | 0.8614 | 44.64 % |
| Multilayer Perceptron | 128.07 | 71.45 | 0.8635 | 43.50 % |
All three architectures converge within 0.01 R² of one another.
The climate × autoregressive signal — not the model family — sets the predictive ceiling.
Selected on four criteria, each with practical consequence for field deployment.
The four criteria
Interpretability. Gain-based feature importance is accessible to non-technical decision-makers.
Robustness under missing data. Real-world surveillance ingestion is incomplete; XGBoost degrades gracefully.
Reproducibility. Tree ensembles are deterministic under a fixed seed; neural-network trajectories vary with batch order.
Lowest mean absolute error. Sixty-nine and a half cases per district-month — the most operationally relevant absolute-error metric.
| Scenario | Precipitation | Temperature | Predicted cases | Δ |
|---|---|---|---|---|
| Baseline · 2015 – 2023 | 127 mm | 14.3 °C | 375,000 | — |
| SSP2-4.5 · moderate · 2050 | 146 mm | 15.3 °C | 469,000 | +25 % |
| SSP5-8.5 · high · 2050 | 165 mm | 15.8 °C | 525,000 | +40 % |
Monte Carlo propagation with one thousand draws; climate perturbations from CMIP6 regional projections.
These are indicative scenario-conditional projections, not precise forecasts.
Disease control cannot rely solely on vaccination or clinical management without parallel investment in climate-resilient water, sanitation, and hygiene infrastructure.
The one-month predictive lead time achievable from this framework is sufficient to pre-position oral rehydration therapy, antibiotics, water-purification tablets, and community health workers ahead of forecast contamination spikes.
This work aligns directly with the health-sector priorities of Nepal’s National Adaptation Plan 2021 – 2050.
The predictive ceiling is not set by the choice of model, but by the climate–autoregressive signal already present in the data Nepal already collects.
The infrastructure for climate-informed early warning of typhoid fever therefore exists today.
What remains is the institutional decision to deploy it.
Kritika Baral · Manuscript · 2026