5 — Discussion
Manuscript · APA section
Manuscript · Section 5 of 7 · Paper landing
5.1 Confirmation of the climate–typhoid nexus
The convergent evidence from correlation analysis, lag-correlation ranking, feature-importance analysis, and seasonal distribution collectively provides robust quantitative confirmation of the climate–typhoid relationship in Nepal at the national district level. Across nine years of data and 77 districts, the study confirms that:
- Typhoid incidence follows a consistent and pronounced seasonal cycle driven by monsoon precipitation and temperature dynamics.
- Temperature is the strongest climatic predictor, operating with a one-month lag through multiple simultaneous biological pathways.
- Precipitation, flooding, and humidity provide supplementary but statistically meaningful predictive signals.
- Autoregressive epidemiological dynamics account for the largest proportion of model-explained variance.
These conclusions are consistent with and extend regional literature from Bangladesh (Hashizume et al., 2010), China (Bao et al., 2023), and India (Amin et al., 2021).
The dominance of temperature over precipitation as the leading climate predictor — which contrasts with the Bangladesh and Vietnamese literature where precipitation was the primary driver — reflects Nepal’s extraordinary altitudinal range (60–8,848 m asl) and the multiple simultaneous biological pathways through which temperature influences S. Typhi transmission (pathogen biology, vector dynamics, food safety) (Crump et al., 2004).
5.2 Why XGBoost is the operational model
The selection of XGBoost as the operational headline rests on four criteria, each of which has practical consequence for field deployment:
- Interpretability for public-health stakeholders. Gain-based feature importance and tree-traversal explanations are accessible to non-technical decision-makers — a requirement for stakeholder engagement in DHIS2-integrated workflows.
- Robustness under missing data. Real-world DHIS2 ingestion is routinely incomplete; XGBoost’s built-in missing-value handling degrades gracefully where MLP would require imputation.
- Reproducibility under fixed seed. Tree ensembles are deterministic given the random seed; MLP training trajectories vary with batch order — a property that complicates regulatory and audit settings.
- Lowest MAE (69.50 cases / district-month). Absolute error in case-count units is the most operationally relevant metric; XGBoost wins this metric while sitting within 0.002 R² of the MLP fallback.
5.3 Implications for climate-informed early-warning systems
The attribution of approximately 86 % of district-month typhoid variance to a feature set composed entirely of freely available climate inputs 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.
An operational system would ingest monthly HMIS case data (4–6 week reporting delay), current-month CHIRPS precipitation (2–3 week delay), and current-month ERA5-Land temperature and humidity (2–4 week delay), generating district-specific one-month-ahead predictions approximately four to six weeks in advance of seasonal peaks — sufficient 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) explicitly calls for climate-sensitive disease surveillance and early-warning systems integrated with climate-monitoring platforms. The 86 % R² achieved here using only freely available CHIRPS and ERA5-Land products demonstrates that the data infrastructure for operational climate-health early warning already exists and requires only an analytical integration layer to become operational.
5.5 Projections under climate change
| Scenario | Floods | Precip. (mm) | Temp. (°C) | RH (%) | Cases | Δ |
|---|---|---|---|---|---|---|
| Baseline (2015–2023) | 97 | 127 | 14.3 | 72.8 | 375,000 | — |
| SSP2-4.5 (moderate, 2050) | 116 | 146 | 15.3 | 76.4 | 469,000 | +25 % |
| SSP5-8.5 (high, 2050) | 136 | 165 | 15.8 | 78.0 | 525,000 | +40 % |
Climate perturbations from CMIP6 regional projections, propagated by Monte Carlo (n = 1,000). Indicative — not a precise forecast.
5.6 Limitations
- HMIS under-reporting — likely most severe in Karnali and Sudurpashchim Provinces (Bhatt et al., 2023).
- Diagnostic misclassification — HMIS clinical diagnosis may conflate typhoid with malaria, dengue, or other febrile illnesses (Parry et al., 2002).
- Spatial aggregation — district-level CHIRPS and ERA5-Land aggregation smooths over intra-district heterogeneity.
- Flood-event underregistration — DRR thresholds may exclude small-scale local events.
- Ecological design precludes individual-level inference.
For the full institutional, NAP-alignment, and limitations discussion, see the Thesis Discussion chapter.
← Back: 4 — Results · → Next: 6 — Data