5 — Discussion
Thesis · APA section — integrated discussion, conclusions, policy recommendations
Thesis · Section 5 of 7 · Paper landing
5.1 Summary of the study
This thesis presents the first comprehensive, national-scale, district-level, multi-model machine-learning study of typhoid incidence across Nepal’s 77 administrative districts for January 2015 – December 2023. Four validated institutional and open-access data sources are integrated: HMIS, DRR Portal, CHIRPS, and ERA5-Land. The analytical dataset comprises 7,327 valid district-month observations.
Eleven predictors were engineered. Four ML models — Random Forest, XGBoost, MLP, and a Weighted Ensemble — were trained on the 2015–2021 dataset and evaluated on a temporally held-out 2022–2023 test set. The Weighted Ensemble achieves the strongest predictive performance across all metrics (R² = 0.8675; RMSE = 126.20; MAE = 68.07; MAPE = 41.01 %). XGBoost feature-importance analysis identifies cases_lag1 as the dominant predictor (normalized importance = 0.670). Seasonal analysis confirms August as the peak typhoid month (~3× dry-season baseline).
5.2 Confirmation of the climate–typhoid nexus
The convergent evidence from correlation analysis, lag-correlation ranking, feature-importance analysis, seasonal distribution, and geographic variation collectively provides the strongest quantitative confirmation to date of the climate–typhoid relationship in Nepal at the national district level. Temperature is the strongest climatic predictor — exceeding precipitation in both univariate correlation and multivariate feature importance — a pattern reflecting Nepal’s extraordinary altitudinal range and temperature’s influence through multiple simultaneous biological pathways (pathogen biology, vector dynamics, food safety) (Amin et al., 2021; Bao et al., 2023; Crump et al., 2004; Hashizume et al., 2010).
5.3 Implications for climate-informed early warning
The Weighted Ensemble model’s R² of 0.8675 on the held-out 2022–2023 test set indicates that 86.75 % of district-month variance can be explained using the prior month’s HMIS case count and current-month climate data. This level of accuracy is sufficient for operational early warning — the goal is not perfect individual prediction but reliable identification of districts and months at elevated risk relative to seasonal baselines, enabling pre-positioned resource deployment before outbreak peaks.
Data latencies (HMIS 4–6 weeks, CHIRPS 2–3 weeks, ERA5-Land 2–4 weeks) suggest predictions could be operationally available four to six weeks in advance of peak incidence — meaningful lead time for district health officers to pre-position oral rehydration therapy, antibiotics, water-purification tablets, and community health-worker deployment.
5.4 Policy alignment with the National Adaptation Plan
Nepal’s National Adaptation Plan 2021–2050 (Government of Nepal, 2021) calls for disease early-warning systems linked to climate monitoring and WASH investment in climate-vulnerable communities. The quantitative findings here directly answer that call: the data infrastructure for operational climate-health early warning already exists, requiring only an analytical integration framework. The geographic concentration of highest burden in Madhesh Terai districts — precisely the areas with lowest WASH coverage and highest flood exposure — provides quantitative justification for targeted dual-benefit climate-adaptation WASH investments.
5.5 Comparison with global performance benchmarks
The Weighted Ensemble R² of 0.8675 exceeds comparable benchmarks: Bao et al. (2023) achieved XGBoost R² of 0.72–0.81 for typhoid in Chinese provinces (climate-only features); Amin et al. (2021) reported Random Forest R² of 0.78 for diarrheal disease in Bangladesh. The advantage attributable to: (a) the autoregressive cases_lag1 feature absent from climate-only studies; (b) the nine-year 77-district panel; (c) the ensemble combination approach. Studies that have explicitly incorporated autoregressive features in analogous contexts report R² in the 0.85–0.92 range (Nsoesie et al., 2014), confirming the present study’s performance is methodologically credible.
5.6 Principal conclusions
- The climate–typhoid relationship in Nepal is robust, statistically significant, and consistent across nine years and 77 districts. Monsoon season is the primary epidemiological amplifier — seasonal peak ~3× the dry-season baseline.
- Temperature is the strongest climatic predictor, surpassing precipitation in both correlation and feature importance.
- ML ensemble models can achieve R² > 0.86 for district-level typhoid forecasting; Weighted Ensemble outperforms all individual components.
- Autoregressive case dynamics account for the dominant fraction of model-explained variance (importance = 0.670).
- Flood events show a modest positive association (r = 0.19; importance = 0.006) — episodic and spatially concentrated, but operationally significant in affected communities.
- The integrated open-access dataset and ML framework are technically reproducible and operationally feasible as a near-real-time early-warning tool.
5.7 Policy recommendations
Six targeted recommendations for MoHP, MoFE, NDRRMA, and development partners.
5.7.1 Operationalize a district-level typhoid early-warning system
MoHP and DHM should invest in operationalizing the ML framework as a district-level early-warning system: (a) ingest monthly HMIS data on completion; (b) retrieve CHIRPS and ERA5-Land estimates via APIs; (c) run the Weighted Ensemble for one-month-ahead predictions; (d) issue automated alerts to district health officers when predicted counts exceed seasonal risk thresholds. Modest computational requirements (laptop-class hardware) and freely available data sources make this system feasible within Nepal’s existing digital health infrastructure.
5.7.2 Prioritize pre-monsoon WASH interventions in high-burden districts
Concentration of monsoon-season burden in Madhesh Terai districts (Table 4.5) provides quantitative justification for geographically and temporally targeted WASH investment. The April–May pre-monsoon window is the optimal intervention opportunity: chlorination and maintenance of water-supply systems, household water-treatment and covered-storage promotion, handwashing facility installation at schools and health posts, community-led sanitation monitoring strengthening.
5.7.3 Expand typhoid conjugate vaccine (TCV) coverage in endemic Terai districts
MoHP’s Expanded Program on Immunization should programmatically expand TCV in Madhesh, eastern Koshi, and Lumbini Terai districts. Annual pre-monsoon TCV campaigns targeting children 6 months – 15 years, timed for March–May delivery, could substantially attenuate the monsoon-season epidemic peak. Economic modeling using district-level predictions identifies highest-return target districts.
5.7.4 Strengthen HMIS data quality and diagnostic capacity
MoHP should prioritize: (a) blood-culture diagnostic capacity at all 77 district hospitals; (b) automated HMIS data-quality audit procedures; (c) performance incentives for timely complete reporting; (d) training of district health management teams in HMIS interpretation and early-warning response.
5.7.5 Integrate climate-health monitoring into NAP implementation
MoFE should formally incorporate climate-sensitive disease monitoring into NAP health-sector adaptation. Specifically: (a) a national climate-health monitoring dashboard consolidating CHIRPS, ERA5-Land, and HMIS data streams; (b) annual district-level climate-health impact assessments; (c) a formal inter-ministerial working group (MoHP, MoFE, DHM, NDRRMA) for data sharing and outbreak response coordination.
5.7.6 Standardized post-flood disease surveillance protocol
MoHP and NDRRMA should jointly develop a protocol mandating: (a) enhanced active HMIS reporting from flood-declared districts within 72 hours of emergency declaration; (b) deployment of rapid-response teams with typhoid RDTs and treatment stockpiles within 5 days of major events; (c) pre-positioned azithromycin and ceftriaxone in NDRRMA district caches; (d) mandatory post-flood water-quality testing within one week of flood recession.
Table 5.1 — Policy recommendations mapped to empirical findings and lead institutions
| Recommendation | Key empirical finding | Lead institution(s) |
|---|---|---|
| 5.7.1 Operationalize early warning | R² = 0.8675; monthly data on regular cycles | MoHP, DHM, WHO, UNICEF |
| 5.7.2 Pre-monsoon WASH | 3× monsoon amplification; Madhesh lowest WASH | MoHP, MoWSSS, UNICEF, World Bank |
| 5.7.3 Expand TCV | Predictable August peak; Madhesh, Koshi, Lumbini highest burden | MoHP, EPI, Gavi |
| 5.7.4 HMIS data quality | Under-reporting acknowledged; quality lifts model accuracy | MoHP, DoHS, WHO |
| 5.7.5 NAP integration | NAP Tier-1 priority; 86.75 % accuracy demonstrated | MoFE, MoHP, DHM, NDRRMA |
| 5.7.6 Post-flood surveillance | r = 0.19 flood-typhoid; 48/77 districts flood-affected | MoHP, NDRRMA, Nepal Red Cross |
5.8 Limitations and future research directions
- HMIS clinical-diagnosis data rather than laboratory-confirmed introduce potential misclassification. Future work should incorporate blood-culture sentinel surveillance from selected Kathmandu, Patan, and Terai hospitals.
- The ecological design precludes individual-level inference. Multilevel models incorporating individual-level vaccination, household WASH, and socioeconomic indicators would provide more mechanistically complete estimates.
- The study does not project burden under future climate scenarios. Future work should generate district-level projections under SSP2-4.5 and SSP5-8.5 using downscaled CMIP6 outputs.
- Monthly temporal resolution is coarser than ideal; weekly or biweekly predictions via HMIS expansion to weekly reporting would provide more operationally timely alerts.
- The study does not model spatial autocorrelation between neighboring districts. Spatial-panel models or graph neural networks could improve outbreak-propagation accuracy.
- LSTM recurrent neural networks merit systematic comparison against the tree-based and MLP models employed here.
5.9 Final reflections
This thesis demonstrates that the systematic integration of Nepal’s national health surveillance infrastructure with globally recognized, freely available climate reanalysis and remote-sensing products — processed through a rigorous multi-model machine-learning framework — can produce accurate, interpretable, and policy-relevant predictions of typhoid incidence at the district level across an entire country. The R² of 0.8675 achieved on genuinely held-out data establishes that the approach is not merely theoretically sound but practically feasible with existing data resources and modest computational infrastructure.
Nepal stands at a pivotal intersection of vulnerability and opportunity. Climate change is intensifying the environmental conditions that drive typhoid transmission precisely as the country is making historic investments in WASH infrastructure, health system strengthening, and climate adaptation. Evidence-based predictive tools that enable proactive rather than reactive public-health responses represent not merely an academic contribution but an investment in the equity and resilience of Nepal’s health system.
← Back: 4 — Results · → Next: 6 — Data