Predictive Modeling of Typhoid Incidence in Nepal

Manuscript — three-model machine-learning framework anchored on XGBoost

Manuscript · Research · About

About	Abstract · Highlights · Figures · Read the paper · How to cite
Paper (APA)	Landing · Introduction · Literature Review · Methodology · Results · Discussion · Data · Sources
Presentation	Slides with APA in-text citations & References
Notebook	Python notebook — runs manuscript/main.py with supporting modules

Abstract

Typhoid fever, caused by Salmonella enterica serovar Typhi, imposes a persistent and disproportionate public-health burden on Nepal. This manuscript presents a tightened, three-model machine-learning framework for predicting district-level typhoid incidence across Nepal’s 77 districts and 108 months (January 2015 – December 2023; N = 7,327 district-month observations). Four open-access national datasets are integrated: HMIS outpatient surveillance, DRR-portal flood records, CHIRPS precipitation, and ERA5-Land temperature and humidity. Three machine-learning models — Random Forest, XGBoost, and a Multilayer Perceptron — are trained on a strictly chronological partition (last 12 months held out for testing). XGBoost is selected as the operational model (R² = 0.8614, RMSE = 129.07, MAE = 69.50, MAPE = 44.64 %) on the basis of lowest MAE, interpretable gain-based feature importance, graceful missing-value handling, and reproducibility under a fixed random seed. All three architectures converge within 0.01 R², indicating that the climate × autoregressive signal — not the model family — sets the predictive ceiling.

Highlights

0.86 XGBoost R² · held-out 12 months

129 Test-set RMSE · cases / district-month

7,327 District-month observations

+25 % National burden by 2050 under SSP2-4.5

• National-scale, district-level multi-model ML study of typhoid fever in Nepal across all 77 districts and nine years of HMIS surveillance.
• XGBoost selected as the operational pick (R² = 0.8614, lowest MAE = 69.50); Random Forest and MLP each within 0.01 R².
• One-month lagged temperature is the strongest climate predictor after the autoregressive cases_lag1 term — consistent with Nepal’s extraordinary altitudinal range.
• Pre-monsoon (April–May) window is the optimal intervention opportunity; August records the peak monsoon incidence (≈ 3 × dry-season median).
• End-to-end reproducible — every number is regenerable by running manuscript/main.py.

Figures

Five publication-quality figures produced by manuscript/main.py, all rendered on the Results page and committed to manuscript/output/figures/:

#	Figure
4.1	National monthly typhoid time series (2015–2023)
4.2	Pearson correlation matrix — climate × cases
4.3	XGBoost gain-based feature importance
4.4	Actual vs. predicted scatter plots — three-model comparison
4.5	Ecological-zone seasonal distribution of district-month cases

Read the paper

• Manuscript (PDF) — manuscript.pdf · the formal manuscript submission.
• APA paper version — seven-section Paper page: Introduction · Literature Review · Methodology · Results · Discussion · Data · Sources.
• Slide deck — Presentation with APA in-text citations and a References slide.
• Code walkthrough — Notebook with main.py and all seven supporting modules embedded inline.

How to cite

Baral, K. (2026). Predictive modeling of typhoid incidence in Nepal under extreme climate change scenarios using machine learning algorithm [Manuscript]. Department of Health Informatics, Kathmandu University, Dhulikhel, Nepal.

@article{baral2026manuscript,
  author      = {Baral, Kritika},
  title       = {Predictive Modeling of Typhoid Incidence in Nepal
                 Under Extreme Climate Change Scenarios Using
                 Machine Learning Algorithm},
  year        = {2026},
  institution = {Department of Health Informatics,
                 Kathmandu University},
  address     = {Dhulikhel, Nepal},
  url         = {https://baralsamrat.github.io/publications/manuscript/}
}