2 — Literature Review
Thesis · APA section
Thesis · Section 2 of 7 · Paper landing
2.1 Introduction
This section critically reviews the scientific literature relevant to three core domains: (a) the epidemiology and global burden of typhoid fever, with attention to South Asia and Nepal; (b) climate-driven modulation of typhoid transmission through precipitation, temperature, humidity, and flooding; and (c) the application of machine-learning methods to climate-sensitive disease prediction.
2.2 Epidemiology and global burden of typhoid fever
2.2.1 Global incidence and mortality
The GBD Typhoid and Paratyphoid Collaborators estimated 10.9 million incident cases of typhoid fever globally in 2017, with approximately 68 % of the burden concentrated in South Asia, followed by sub-Saharan Africa (15 %) and Southeast Asia (12 %) (GBD 2023 Causes of Death Collaborators, 2025). Global mortality in 2017 was estimated at 116,800 deaths. Children aged 5–14 years bear the highest age-specific incidence globally (Parry et al., 2002; World Health Organization, 2018). Multidrug-resistant (MDR) typhoid has become endemic in several South Asian countries, and extensively drug-resistant (XDR) strains have been reported from Pakistan and Nepal (Bhatt et al., 2023; World Health Organization, 2018).
2.2.2 Typhoid in Nepal
Nepal has historically recorded among the highest per-capita typhoid incidence rates in South Asia. Bhatt et al. (2023) found that enteric fever accounts for 5–8 % of all OPD visits during peak monsoon season, with rates in highly endemic Terai districts exceeding 1,000 cases per 100,000 population per year. Hospital-based blood-culture studies confirm the predominance of S. Typhi (~80–90 %) (Bhatt et al., 2023; Parry et al., 2002). The HMIS represents the primary national source of systematic district-level typhoid surveillance (Ministry of Health and Population, Government of Nepal, 2023).
2.3 Climate drivers of typhoid transmission
2.3.1 Precipitation and water-contamination
Heavy rainfall generates surface runoff that carries fecal matter into open water sources, shallow wells, and the intake points of inadequately protected piped water systems (Bao et al., 2023; Hashizume et al., 2010). Hashizume et al. (2010) documented a 2–3 week lag between precipitation and diarrheal disease in Dhaka. Bao et al. (2023) identified one-month lagged precipitation as the strongest contemporaneous climate predictor of typhoid in Jiangsu, China. The precipitation–typhoid relationship is not strictly linear: at very high rainfall intensities, a threshold-saturation pattern is observed.
2.3.2 Temperature and pathogen biology
Laboratory studies establish that S. Typhi growth in liquid culture follows a characteristic thermal response curve, with optimal replication at 37 °C, substantially reduced growth below 15 °C, and heat inactivation above 60 °C (Crump et al., 2004). Amin et al. (2021) found that a 1 °C increase in mean monthly temperature was associated with a 3.2 % increase in monthly diarrheal disease cases in Bangladesh. Temperature also regulates mechanical fly-vector populations (Crump et al., 2004). The high positive correlations between mean temperature and typhoid incidence observed in this study (Pearson r = 0.50–0.63; §4.4) are biologically plausible and methodologically consistent with this international literature.
2.3.3 Humidity and co-factor mechanisms
Relative humidity influences typhoid transmission through its regulation of S. Typhi survival in aerosols and moist environmental matrices, the viability of pathogen-contaminated food, and the activity of fly vectors. High humidity prolongs survival of S. Typhi in moist fecal matter, soil surfaces, and uncovered food items. The moderate association observed here (r = 0.22) is consistent with the hypothesis that humidity operates primarily as a co-factor rather than as an independent primary driver (Bao et al., 2023).
2.3.4 Flood events and post-flood clusters
Flooding amplifies typhoid risk through simultaneous contamination of multiple water sources, displacement of populations into degraded sanitary conditions, disruption of healthcare access, and immune stress from nutritional deprivation (NPDRR Nepal, 2025; World Health Organization, 2018). NDRRMA (2025) documented 267 flood events in 2022 alone, affecting 48 of 77 districts. Post-flood typhoid cluster investigations conducted by MoHP rapid-response teams have confirmed elevated typhoid incidence in affected districts within two to six weeks of major flood events (Ministry of Health and Population, Government of Nepal, 2023).
2.4 Climate change, Nepal, and health outcomes
2.4.1 Climate-change projections for Nepal
Nepal is recognized as among the countries most vulnerable to climate change despite contributing < 0.027 % of global greenhouse gas emissions (Ministry of Forests and Environment, Government of Nepal, 2021). The IPCC (2022) projects mountain regions will warm at 1.5–2 times the global average rate due to elevation-dependent warming amplification. Under SSP2-4.5, Nepal’s mean temperature is projected to increase 1.5–2.5 °C by 2060 and 2.0–3.0 °C by 2100. Under SSP5-8.5, increases of 3.0–5.0 °C by 2100 are within the range of plausible outcomes. Climate-model ensembles project 10–40 % increases in the return frequency of extreme rainfall events by mid-century (Government of Nepal, 2021; Ministry of Forests and Environment, Government of Nepal, 2021).
2.4.2 Health policy and climate adaptation in Nepal
Nepal’s National Adaptation Plan 2021–2050 (Government of Nepal, 2021) identifies the health sector as a Tier 1 priority for climate adaptation, calling specifically for strengthening of climate-sensitive disease surveillance, development of disease early-warning systems integrated with climate-monitoring platforms, climate-proofing of health infrastructure, integration of WASH investments with adaptation planning, and national research capacity development. The Nepal Multiple Indicator Cluster Survey (Central Bureau of Statistics and UNICEF, 2019) confirms that approximately 30 % of households in Madhesh Province lack access to improved sanitation facilities — concentrated precisely in the districts most exposed to climate-driven flood and precipitation increases.
2.5 Machine learning in climate-health modeling
2.5.1 Overview and comparative capabilities
Machine learning offers several demonstrated advantages over traditional parametric regression for climate-health modeling: handling of high-dimensional predictor spaces with multicollinear variables; modeling of non-linear, threshold, and interaction effects without manual specification; robustness to violations of distributional assumptions; and built-in feature-importance quantification (Breiman, 2001; Chen & Guestrin, 2016; Nsoesie et al., 2014).
2.5.2 Random Forest
Random Forest (Breiman, 2001) constructs decision trees using bootstrap aggregation (bagging), with each tree trained on random subsets of observations and features. RF is inherently resistant to overfitting, handles missing data and multicollinear features without preprocessing, and provides Gini-based feature-importance estimates.
2.5.3 XGBoost
XGBoost (Chen & Guestrin, 2016) is a scalable, regularized implementation of gradient boosting that iteratively constructs decision trees to minimize a regularized loss function. It incorporates L1/L2 regularization, second-order Taylor expansion, column subsampling, and parallel tree construction. Bao et al. (2023) demonstrated XGBoost’s superiority over RF, SVR, and ridge regression for monthly typhoid prediction in China.
2.5.4 Multilayer Perceptron
The MLP is a feedforward neural network composed of an input layer, one or more hidden layers with non-linear activations, and an output layer. MLPs are universal function approximators given sufficient depth (Nsoesie et al., 2014). Disadvantages include larger training-sample requirements, hyperparameter sensitivity, and lower interpretability.
2.5.5 Ensemble methods — Weighted Combination
Model ensembling — combining predictions from multiple independently trained models — consistently outperforms individual component models in prediction tasks by reducing variance, capturing complementary patterns, and providing robustness to misspecification (Nsoesie et al., 2014). Ensemble approaches have demonstrated performance gains of 3–12 % in R² over the best individual component in influenza prediction challenges and dengue forecasting competitions. Weighted ensemble methods, which combine predictions using optimized weights rather than simple averaging, typically outperform equal-weight averaging when component models exhibit meaningfully different performance levels. In this thesis, ensemble weights are optimized on the held-out validation subset of the training data by minimizing RMSE using SLSQP.
2.6 Key prior studies
| Citation | Location | Method | Period | Key findings |
|---|---|---|---|---|
| Bao et al. (2023) | Jiangsu, China | XGBoost, RF, SVR | 2005–2020 | XGBoost R² = 0.81; temperature (lag-1) and precipitation (lag-1) strongest |
| Amin et al. (2021) | Bangladesh | RF, ANN | 2007–2018 | RF R² = 0.78 for diarrheal disease; monsoon indicator added value |
| Hashizume et al. (2010) | Dhaka, Bangladesh | Time-series regression | 2000–2008 | 2–3 week precipitation lag; temperature and humidity independently significant |
| Bhatt et al. (2023) | Nepal | Systematic review | multiple | Monsoon-season peak confirmed; substantial HMIS under-reporting acknowledged |
| GBD 2023 Causes of Death Collaborators (2025) | Global | GBD modeling | 2017 | 10.9 M typhoid cases globally; South Asia 68 %; children 5–14 highest incidence |
| Crump et al. (2004) | Global (review) | Narrative review | multiple | Thermal regulation of S. Typhi survival; fly-vector dynamics; food/water mechanisms |
| Nsoesie et al. (2014) | Global (review) | SR of ML methods | multiple | ML outperforms regression; ensemble approaches most robust |
| Ministry of Health and Population, Government of Nepal (2023) | Nepal | HMIS surveillance | 2021/2022 | Typhoid consistently in top-5 OPD diseases; monsoon peak; Terai highest burden |
| NPDRR Nepal (2025) | Nepal | DRR Portal | 2015–2023 | 267 flood events in 2022; 48/77 districts affected |
2.7 Data sources in climate-health research
2.7.1 HMIS as a research resource
National HMIS systems represent the primary source of routine disease surveillance data in most LMICs (Amin et al., 2021; Bao et al., 2023). The Nepal HMIS has been operational in its current form since 1994 and has been progressively digitized, with web-based data entry standard at district level from approximately 2015 onwards (Ministry of Health and Population, Government of Nepal, 2023). Bhatt et al. (2023) estimated overall HMIS reporting rates at approximately 70–85 % of registered health facilities nationally.
2.7.2 CHIRPS precipitation dataset
CHIRPS (Funk et al., 2015) combines satellite-derived Cold Cloud Duration estimates with quality-controlled rain-gauge measurements to produce blended precipitation at 0.05° (~5 km) spatial resolution and monthly temporal resolution from January 1981 to near-present. CHIRPS performs particularly well in regions with complex topography and sparse gauge networks; validation studies for Nepal report Pearson r = 0.75–0.92 against gauge measurements (Funk et al., 2015).
2.7.3 ERA5-Land reanalysis dataset
ERA5-Land (Muñoz-Sabater et al., 2021) provides global land-surface data at 0.1° (~9 km) spatial resolution and hourly temporal resolution from January 1950 to near-present, with 4D-Var data assimilation from radiosondes, surface stations, aircraft, and satellite platforms. Independent validation confirms mean absolute errors typically below 1.5 °C for monthly mean temperature over South Asia.
2.8 Research gap and original contribution
No published study has: (a) applied a multi-model ML framework to predict typhoid incidence across all 77 districts of Nepal; (b) integrated HMIS, CHIRPS, ERA5-Land, and DRR Portal datasets within a unified analytical framework; (c) explicitly modeled the lagged climate–disease relationship and autoregressive case dynamics at the district-month scale across a nine-year horizon; or (d) evaluated ensemble model performance against individual model baselines on a temporally held-out test set in the Nepali context. This thesis fills each of these gaps.
← Back: 1 — Introduction · → Next: 3 — Methodology