5 — Discussion
5.1 — Interpretation of climate–typhoid relationships
This study provides robust evidence that climate variability — and particularly temperature — plays a significant and quantifiable role in shaping typhoid fever dynamics in Nepal. At the district-month level, mean temperature is the strongest individual climate driver (Pearson r = 0.50 – 0.63), followed by precipitation (r = 0.33 – 0.37), the monsoon-season indicator (r = 0.25 – 0.29), and flood-event frequency (r = 0.19). One-month-lagged versions of these variables — precipitation r = 0.313, humidity r = 0.180, flood frequency r = 0.122 — together with the autoregressive cases_lag1 term (r = 0.731) and lagged temperature (r = 0.599) support a predictive model that explains R² = 0.87 of district-month variance on a strictly chronological held-out test window. These findings reinforce the classification of typhoid as a climate-sensitive waterborne disease whose transmission is strongly modulated by both gradual climate trends (temperature) and hydro-meteorological extremes (floods), interacting with persistent endemic foci (Levy et al., 2023). The dominant influence of cases_lag1 (normalised importance ≈ 0.67) — well above any single climate predictor — additionally underscores that typhoid in Nepal is partly endogenous to the spatial distribution of unsafe water sources, not driven by climate alone.
The structural dominance of lagged climate exposure — precipitation, humidity, and flood frequency one month before the case window — highlights the importance of extreme events as acute disease amplifiers. Floods disrupt WASH systems by breaching latrines, contaminating shallow groundwater, and overwhelming drainage infrastructure — particularly in the densely populated Terai, where open defecation rates remain high and tube wells are shallow (Baker et al., 2025; Central Bureau of Statistics and UNICEF, 2019). The observed one-month lag between climate exposure and peak typhoid incidence is biologically plausible and aligns with both the Salmonella Typhi incubation period and the delayed pathway from environmental contamination to clinical presentation. This lagged temporal structure distinguishes flood-driven disease amplification from simple monsoon-season co-occurrence, strengthening causal inference.
Climatic modifiers further conditioned this relationship. High monsoon rainfall (> 300 mm / month) amplified typhoid risk by accelerating fecal runoff into surface water sources, while elevated relative humidity (> 80%) likely extended bacterial survival in aquatic environments, consistent with prior evidence (Brubacher et al., 2020; Levy et al., 2023). Temperature exerted a secondary but synergistic role in lowland Terai districts, where warmer conditions may enhance bacterial replication rates (Trájer et al., 2022). Together, these findings demonstrate that typhoid risk in Nepal is governed by compound climate stresses in which flood events act as acute transmission triggers while background climate conditions modulate baseline transmission efficiency. Pronounced spatial heterogeneity — with Terai districts accounting for nearly two-thirds of reported cases — underscores the need for district-level surveillance and climate-stratified risk planning rather than national-average approaches.
5.2 — Predictive performance and methodological contribution
Four model architectures — Random Forest (R² = 0.8563), XGBoost (R² = 0.8614), a Multilayer Perceptron (R² = 0.8635), and a Weighted Ensemble of all three (R² = 0.8675) — converged to within 0.012 R² of each other on a strictly chronological held-out test window covering the most recent 12 months of data. This tight clustering is not a coincidence: it indicates that the predictive ceiling is set by the signal in the climate × autoregressive features, not by the choice of algorithm. The Weighted Ensemble is the headline performer on every metric (RMSE = 126.20, MAE = 68.07, MAPE = 41.01 %), validating the use of structurally distinct learners whose errors are partially uncorrelated (Dixon et al., 2023). When the operational decision is to deploy a single model — for example, into Nepal’s DHIS2 surveillance dashboard — XGBoost is the recommended pick on the basis of its low MAE, interpretable gain-based feature importance, graceful missing-value handling, and reproducibility under fixed random seeds (Chen & Guestrin, 2016; Choi et al., 2023). Together, the four models explain roughly 87 % of district-month typhoid variance using only freely-available climate inputs and the previous month’s case count — a meaningful improvement on linear and ARIMA-class baselines that earlier Nepal-focused studies have relied on, and one that captures the joint interaction between flood frequency, lagged precipitation, monsoon seasonality, and persistent endemic foci that those approaches cannot model simultaneously.
Methodologically, this study contributes a reproducible hybrid climate–health framework — integrating passive HMIS surveillance, DRR disaster records, and reanalysis climate grids — that can be adapted to other climate-sensitive enteric diseases (cholera, shigellosis) or applied in comparable LMIC settings with similar data infrastructure.
5.3 — Implications for public health, early warning, and governance
The attribution of approximately 87 % of district-month typhoid variance to a feature set composed entirely of lagged climate exposures and the previous month’s case count carries a direct policy implication: disease control strategies that focus solely on vaccination or clinical management are insufficient without parallel investment in climate-resilient WASH infrastructure and early warning capacity. Flood-prone Terai districts — already carrying two-thirds of national burden while facing the greatest structural vulnerabilities and highest poverty rates — represent the highest-priority zones for anticipatory public health action (Asian Development Bank, 2024; Central Bureau of Statistics and UNICEF, 2019).
Operationalisation in DHIS2
The predictive models presented here could be operationalized within Nepal’s existing digital health infrastructure. Integration with the District Health Information Software 2 (DHIS2) platform would allow real-time risk estimation based on incoming flood alerts and climate forecast data, enabling automated pre-monsoon trigger mechanisms for targeted interventions such as:
- Emergency chlorination of contaminated water sources.
- Distribution of water purification materials.
- Mobile health camp deployment.
- District-level hygiene communication.
For this integration to be effective, institutional coordination between the Department of Hydrology and Meteorology (DHM), the Ministry of Health and Population (MoHP), and the National Disaster Risk Reduction and Management Authority (NDRRMA) is essential. Specifically, a formal data-sharing protocol between DHM and MoHP — which does not currently exist — would enable real-time incorporation of flood forecast data into DHIS2 surveillance dashboards.
Governance and capacity challenges
Practical implementation of climate-health early warning systems in Nepal faces several governance and capacity challenges that must be acknowledged:
- Limited district capacity. District-level health offices in the Terai often have limited technical capacity to interpret climate risk outputs; operationalization of any early warning tool would therefore require dedicated training and decision-support materials for district health officers.
- Data quality variability. Data quality in rural HMIS remains inconsistent, with some districts reporting only 60–70% of expected monthly records, creating gaps that can reduce model accuracy in the communities most at risk.
- Federalism transition. Political and administrative decentralization since 2017 has created ambiguity about health emergency coordination responsibilities between federal, provincial, and local governments — a governance gap that climate-health early warning systems must navigate explicitly.
Implications for TCV strategy
While 85% national TCV coverage has reduced burden in some districts, climate-driven transmission risk persists where coverage is lowest — notably in Karnali Province (< 70% coverage), which also faces increasing flood exposure (Coalition Against Typhoid, 2024). Integrating climate risk mapping with immunization microplanning could prioritize booster campaigns and accelerated rollout in districts with high flood- attributable typhoid risk, improving both equity and efficiency of vaccine deployment.
Projected increases of 25–40% in typhoid burden under mid-century climate scenarios highlight the health co-benefits of climate adaptation investment more broadly: flood-resilient WASH infrastructure, improved urban drainage, and safe water access function simultaneously as disaster risk reduction and long-term infectious disease prevention measures (Intergovernmental Panel on Climate Change, 2022; Ministry of Forests and Environment, Government of Nepal, 2021).
Typhoid incidence may also serve as a useful indicator of climate- related vulnerability in water and sanitation systems — reflecting how effectively communities are protected from flood-driven contamination events. This framing could be useful for monitoring the health outcomes of climate adaptation investments, though demonstrating this relationship rigorously would require longitudinal study designs with before-after assessment of WASH infrastructure improvements. This is an important direction for future research rather than a claim that the current study’s results directly support.
5.4 — Limitations and future research directions
This study has several limitations:
- Ecological design. The study relies on aggregated district-level data, precluding individual-level causal inference and potentially obscuring within-district heterogeneity in exposure and access.
- Surveillance capture fraction. HMIS typhoid data are based on clinical diagnosis rather than laboratory confirmation, introducing potential misclassification; the surveillance capture fraction is estimated at approximately 36%, meaning the true burden and its climate sensitivity may differ from observed patterns.
- Flood event detail. Flood data were coded as frequency counts, without information on event severity, duration, or spatial extent. More granular flood characterization (e.g. inundation depth, affected population) would likely strengthen the climate–disease attribution.
- Climate input resolution. ERA5-Land and CHIRPS grids, while state-of-the-art, may under-represent micro-climates and complex Himalayan terrain at the district level.
- Short training window. Nine years of monthly observations (108 time steps nationally) is short for LSTM training; a longer panel would likely improve sequential model performance.
- Projection stability assumption. The 2050 projections assume stationarity of the climate–disease relationship, holding WASH infrastructure, TCV coverage, urbanization, and pathogen characteristics constant — assumptions unlikely to hold over a 25-year horizon.
The original draft of this section continues with further limitations and a future-work list. Re-paste the remaining content into this file once available; the structure here is already in place to receive it.
Future research directions
- Pathogen surveillance integration. Coupling HMIS clinical records with environmental wastewater surveillance and laboratory-confirmed serotyping would refine both burden estimates and climate-attributable variance.
- District-specific lag tuning. Re-fitting per-zone models with region-specific lag structures to capture differential transmission dynamics across Terai, mid-hill, and Himalayan zones.
- Operational DHIS2 pilot. Packaging the trained ensemble as an API consumed by Nepal’s DHIS2 deployment, with weekly risk-tier alerts pushed to district health officers in two pilot Terai districts as a proof-of-concept for national scale-up.
- WASH adaptation evaluation. Longitudinal before-after designs measuring typhoid incidence in districts that receive flood-resilient WASH investments under Nepal’s Country Partnership Strategy (2025–2029) — testing whether climate-resilient infrastructure produces the predicted health co-benefits.