A Machine Learning Approach to Defining and Predicting the Scale of Typhoid Fever Outbreaks

This study utilizes machine learning on historical outbreak data to establish standardized criteria for defining typhoid fever outbreaks and identifies key sanitation and demographic risk factors to predict and prepare for large-scale epidemics in high-risk regions like Africa and South Asia.

The Gist

Imagine typhoid fever as a mischievous, invisible fire that can start in a kitchen and spread through a neighborhood. Sometimes, it's just a small spark that dies out quickly. Other times, it becomes a raging wildfire that consumes entire cities.

The problem is that health officials often don't know the difference between a "spark" and a "wildfire" until it's too late. They also lack a clear rulebook for when to call in the fire trucks (vaccines) to stop the fire from spreading.

This paper is like a team of digital detectives using Machine Learning (super-smart computer programs) to solve two big mysteries:

1. What exactly counts as an "outbreak"? (When do we sound the alarm?)
2. Which countries are most likely to face a "wildfire" instead of just a "spark"?

Here is the breakdown of their investigation, explained simply:

1. Defining the Alarm: The "Spark" vs. The "Fire"

The researchers looked at 34 past typhoid outbreaks that had detailed records (like a diary of every day the fire burned). They wanted to find a pattern to decide: When is a cluster of sick people actually an "outbreak"?

The Discovery:
They found that every single real outbreak shared two common traits:

• It lasted at least 7 days.
• It involved at least 6 total sick people (or just 2 if they were 100% sure via lab tests).

The Analogy:
Think of it like a smoke alarm. If you see one wisp of smoke for an hour, it might be toast. But if you see smoke for 7 days and it fills the whole house with 6 people coughing, you know it's a fire. The authors propose this as the official rule: If you see 6+ cases in a week, or 2 confirmed lab cases, sound the alarm.

2. The Two Buckets: Small Puddles vs. Big Lakes

Next, they used a computer technique called Unsupervised Learning. Imagine you have a pile of mixed-up marbles (outbreaks) and you want to sort them into buckets without knowing the colors beforehand. The computer looked at the size and duration of the outbreaks and naturally sorted them into two distinct groups:

• Bucket A (The Small Puddles): Outbreaks with fewer than 191 cases. These were short and manageable.
• Bucket B (The Big Lakes): Outbreaks with 288+ cases. These were massive and long-lasting.

To make it easier to use, they drew a line in the sand at 250 cases.

• Under 250? It's a "Small Outbreak."
• Over 250? It's a "Large Outbreak."

3. Predicting the Future: The Crystal Ball

Now that they could tell a "Small" from a "Large" outbreak, they asked: Can we predict which one a country will face before it even happens?

They trained Supervised Learning models (like a student studying for a test) using data from 215 past outbreaks. They fed the computer information about the countries where these outbreaks happened, specifically looking at:

• Water & Sanitation: Do people have clean tap water? Do they have toilets that don't leak?
• Urbanization: Are people living in crowded cities?
• Travel: Do people fly in from places where typhoid is common?

The Result:
The computer learned a clear pattern. Large, dangerous outbreaks are most likely to happen in countries where:

• Clean water is hard to find.
• Toilets are unsafe or non-existent.
• Many people live in crowded cities.

4. The Map of Danger Zones

Using these smart models, the team projected what might happen in 2023 across 192 countries.

• The "Safe Zones" (Blue): North America, Europe, and Australia were predicted to mostly have small, manageable outbreaks (or none at all).
• The "High-Risk Zones" (Red): Large parts of Sub-Saharan Africa and South Asia were predicted to face massive, large-scale outbreaks if the disease strikes.

The Analogy:
Imagine a weather forecast. Instead of predicting rain, these models are predicting "Typhoid Storms." The map shows that while some places might get a light drizzle (small outbreak), countries like the Democratic Republic of Congo, Nigeria, and Bangladesh are in the path of a hurricane (large outbreak).

5. Why This Matters

This research is like giving public health officials a flashlight and a map.

• The Flashlight (The Definition): Now they know exactly when to turn on the light and say, "This is an outbreak, we need to act."
• The Map (The Prediction): They can see which countries are most vulnerable to a "wildfire."

The Bottom Line:
By knowing where the big fires are likely to start, health organizations (like the WHO) can pre-stock vaccines (Typhoid Conjugate Vaccines) in those specific regions. Instead of waiting for the fire to burn out of control, they can have the fire trucks ready to roll the moment the alarm sounds.

This study turns guesswork into a strategic plan, potentially saving thousands of lives by getting the right tools to the right places before the disaster hits.

Technical Summary

1. Problem Statement

Typhoid fever remains a significant global health burden, causing an estimated 10–20 million cases and over 100,000 deaths annually, primarily in low- and middle-income countries (LMICs). While the World Health Organization (WHO) recommends Typhoid Conjugate Vaccines (TCVs) for outbreak response, there is no standardized definition for triggering such a response. Current surveillance is often inconsistent, and the lack of a clear threshold for what constitutes an "outbreak" versus sporadic cases hinders timely public health interventions. Furthermore, there is a need to predict which countries are susceptible to large-scale epidemics to facilitate proactive preparedness (e.g., vaccine stockpiling).

2. Methodology

The study employed a two-stage machine learning pipeline: Unsupervised Learning to define outbreak characteristics and Supervised Learning to predict outbreak scale based on country-level features.

A. Data Sources

• Training Data (Unsupervised): A dataset of 34 distinct typhoid outbreaks (2000–2022) from the Koh et al. study, featuring detailed time-series data (case counts, start/end dates, interventions).
• Training Data (Supervised): A merged dataset of 215 outbreaks (2001–2022) from Koh et al., Kim et al., and Appiah et al.
• Predictors: Country-level data from the World Bank (WaSH coverage, demographics, urbanization) and modeled air travel data (proportion of travelers from high-incidence countries).

B. Unsupervised Learning (Defining Outbreaks)

• Goal: Identify natural clusters in outbreak data to establish a threshold for "large" vs. "small" outbreaks.
• Preprocessing: Variables with >20% missing data were excluded. Remaining missing values were imputed using Iterative Imputation (Random Forest). Highly collinear variables were removed.
• Algorithm:
  • Dimensionality Reduction: Uniform Manifold Approximation and Projection (UMAP) was used instead of PCA to preserve local and global non-linear structures.
  • Clustering: K-means clustering.
  • Optimization: The optimal number of clusters ($k$) was determined using the Silhouette Score and the Elbow Method (Within-Cluster Sum of Squares).
  • Robustness: Consensus clustering was performed over 1,000 iterations with different random seeds to ensure stability.

C. Supervised Learning (Predicting Scale)

• Labeling: Outbreaks were labeled as "Large" ($\ge$ 250 total cases) or "Small" ($<$ 250 total cases). The threshold of 250 was derived from the clustering gap (Cluster A min: 288; Cluster B max: 191) plus a 30% margin for underreporting.
• Models: Five classifiers were trained: Logistic Regression, k-Nearest Neighbors (kNN), Support Vector Machine (SVM), Random Forest, and XGBoost.
• Handling Imbalance: The dataset was imbalanced (~75% small, ~25% large). SMOTE (Synthetic Minority Oversampling Technique) was applied to the training set only.
• Validation: Stratified 5-fold cross-validation was used.
• Threshold Optimization: Instead of the default 0.5 probability threshold, Youden's J statistic was used to find the optimal classification threshold for each model to maximize sensitivity and specificity.
• Explainability: SHapley Additive exPlanations (SHAP) were applied to Random Forest and XGBoost models to interpret feature importance.

3. Key Contributions

1. Operational Definition of an Outbreak: The study proposes a standardized, evidence-based definition for the onset and conclusion of a typhoid outbreak:
   • Onset: $\ge$ 6 total cases (suspected/probable/confirmed) OR $\ge$ 2 laboratory-confirmed cases within a 7-day period.
   • Conclusion: Fewer than 6 total cases and fewer than 2 confirmed cases for two consecutive weeks.
2. Bimodal Outbreak Classification: Demonstrated that typhoid outbreaks naturally segregate into two distinct clusters: "Small" (median 75 cases, max 191) and "Large" (median 879 cases, min 288).
3. Predictive Framework: Developed a robust machine learning framework to predict the scale of potential outbreaks based on pre-outbreak WaSH and demographic indicators.
4. Global Risk Mapping: Generated high-resolution predictions for 192 countries, identifying specific regions at high risk for large-scale epidemics.

4. Results

Clustering Analysis

• Optimal Clusters: Both Silhouette and Elbow methods confirmed $k=2$ as the optimal number of clusters.
• Cluster Characteristics:
  • Cluster A (Large): Median 879 cases, median duration 181 days.
  • Cluster B (Small): Median 75 cases, median duration 60 days.
• Validation: High silhouette score (0.61) and low Davies-Bouldin index (0.51) confirmed strong cluster separation.

Predictive Modeling Performance

• Best Performers:
  • SVM achieved the highest accuracy (0.72) in cross-validation.
  • Logistic Regression achieved the highest AUC (0.72) and F1-score, indicating superior overall discriminatory power.
• Threshold Calibration: Optimal thresholds varied significantly by model (e.g., XGBoost: 0.04; kNN: 0.60), highlighting the necessity of calibrating thresholds rather than using 0.5.
• Sensitivity: Models showed high sensitivity in predicting large outbreaks, minimizing false negatives (critical for public health).

Geographic Predictions (2023 Projections)

• High-Risk Regions: Consistent predictions of large-scale outbreaks were identified in Sub-Saharan Africa (e.g., DRC, Nigeria, Rwanda, Burundi, Ethiopia) and South Asia (e.g., India, Bangladesh, Pakistan, Nepal).
• Low-Risk Regions: North America, Europe, Australia, and New Zealand were consistently predicted to face only small-scale outbreaks.
• Uncertainty: Greater model disagreement was observed in Oceania and parts of Central/South America.

Feature Importance (SHAP Analysis)

• Top Predictors: The most influential factors driving large outbreak predictions were:
  1. Prevalence of open defecation.
  2. Proportion of the population with safely managed sanitation services.
  3. Urban population percentage.
  4. Time since 2000 (temporal trend).
• Interpretation: Poor WaSH infrastructure is the primary driver of outbreak magnitude.

Out-of-Sample Validation

• The model was validated against 5 reported outbreaks in 2023–2025. It correctly predicted the scale (Large vs. Small) for 4 out of 5 cases (100% sensitivity for large outbreaks, 66% specificity for small).

5. Significance and Implications

• Policy Impact: The proposed definition provides a clear, actionable trigger for public health authorities to initiate TCV campaigns and other interventions, addressing the current ambiguity in outbreak response.
• Resource Allocation: By identifying countries with high probabilities of large-scale outbreaks, the study enables targeted TCV stockpiling and pre-positioning of resources in high-risk zones (e.g., DRC, Rwanda, Bangladesh).
• Methodological Advancement: The study demonstrates the efficacy of combining unsupervised learning (for definition) and supervised learning (for prediction) in infectious disease epidemiology, particularly when baseline surveillance data is sparse.
• Limitations & Future Work: The study relies on country-level data (masking subnational heterogeneity) and assumes stable air travel patterns. Future work should incorporate seroincidence data, climatic variables, and subnational metrics to refine predictions.

In conclusion, this research bridges the gap between surveillance data and actionable public health policy by defining typhoid outbreaks rigorously and providing a predictive tool to anticipate their scale, ultimately supporting more effective and timely responses to typhoid fever.