Pediatric HIV Hotspots in Kenya: Machine Learning and Geostatistical Analysis for Enhanced Case Finding

This study developed and validated a framework combining machine learning and geostatistical analysis to predict pediatric HIV incidence and identify statistically significant hotspots across Kenya, thereby enabling more equitable and data-driven resource allocation for case finding.

The Gist

The Big Picture: Finding the "Hidden Spots"

Imagine Kenya's fight against HIV in children is like a massive game of hide-and-seek. Health workers know the game is being played in certain neighborhoods (counties), but they don't always know exactly where the kids are hiding or how many there are in each specific spot. Sometimes, the official reports are like a blurry photo—they show the general area, but miss the fine details.

This paper is about a team of researchers who built a smart digital detective to sharpen that photo. They combined two powerful tools:

1. Machine Learning (The Crystal Ball): A computer program that learns from past data to guess where new cases might be.
2. Geostatistics (The Heat Map): A way of looking at the map to see where cases are "clumping" together like magnets.

Their goal was to create a clearer picture of where children with HIV are living, so health resources (like tests and medicine) could be sent exactly where they are needed most.

---

How They Did It: The Recipe

1. Gathering the Ingredients
The researchers didn't just look at HIV numbers. They gathered a huge bowl of ingredients from two main sources:

• The Test Results: Data from actual HIV tests given to children between October 2022 and June 2023.
• The Context: Data from a national survey (the 2022 Kenya Demographic and Health Survey) about things like:
  • How many pregnant women got tested for HIV?
  • How many children are stunted (not growing well)?
  • How many people have had multiple partners?
  • How much malaria medicine (Fansidar) was used?

2. Training the Detective (Machine Learning)
They fed this data into a computer and asked it to learn the patterns. They tried three different "algorithms" (mathematical recipes) to see which one was the best guesser.

• The Winner: A method called Lasso Regression. Think of this as a very strict editor that looks at all the clues and says, "Okay, these three things matter the most; ignore the rest."
• The Result: The computer predicted 3,160 new cases. The actual official reports said 3,092. That is a very close match (like guessing 3,160 jellybeans in a jar when there are actually 3,092).

3. Drawing the Map (Geostatistics)
Once the computer made its predictions, the researchers didn't just look at the raw numbers. They adjusted for population size.

• Analogy: If County A has 1 million kids and County B has 10,000 kids, finding 50 cases in County A isn't as scary as finding 50 cases in County B.
• They calculated the "incidence" (cases per 10,000 kids) to make a fair comparison.
• Then, they used a special statistical tool (Getis-Ord Gi*) to find Hotspots.
  • Hotspots: Areas where cases are clumped together significantly more than by random chance (like a pile of hot coals).
  • Coldspots: Areas where cases are surprisingly low (like a cool breeze).

---

What They Found: The Map Revealed

The "Usual Suspects"
The map confirmed what health officials already suspected: Western Kenya (specifically Homa Bay, Siaya, and Kisumu) is a major hotspot. These areas have high rates of HIV, and the computer agreed with the human reports.

The "Surprises"
The computer found something the human reports missed. In some areas, the computer predicted high rates, but the official reports were low.

• Analogy: Imagine a smoke detector that goes off in a room where you don't see smoke yet. The computer is saying, "Something is brewing here; check this out."
• Isiolo (in the north) showed the highest rate of infection per child.
• Tana River, Lamu, and Vihiga were flagged by the model as having higher risks than the current reports suggested. This might mean these areas are missing cases because they aren't testing enough children yet.

The "Clumping" Effect
The study proved that HIV cases aren't scattered randomly like raindrops. They cluster. If a child in one village has HIV, it's statistically more likely that a child in the next village over also has it. This helps explain why resources need to be targeted at specific regions rather than spread evenly everywhere.

---

The "Uncertainty" Check

The researchers were careful not to just give a single number. They built a "safety net" around their predictions.

• The Analogy: Instead of saying "There are exactly 50 cases," they said, "We are 95% sure the number is between 40 and 60."
• They found that for almost every county, the real number fell inside their safety net.
• Two Exceptions:
  1. Homa Bay: The real number was higher than the safety net. This suggests their testing programs there are working so well they are finding more cases than the model expected.
  2. Siaya: The real number was lower than the safety net. This suggests they might be missing cases, or the model overestimated the risk there.

---

The Bottom Line

This paper didn't invent a new drug or a new test. Instead, it built a better map.

By combining a smart computer guess with a detailed look at the geography, the researchers created a framework that helps health leaders see the "hidden spots" of pediatric HIV. It allows them to say, "We know the big clusters in the West, but let's also go check these other areas where the computer thinks there might be hidden cases."

The study concludes that using this mix of Machine Learning (to predict) and Spatial Analysis (to map) is a powerful way to make sure no child is left behind in the fight against HIV.

Technical Summary

1. Problem Statement

Despite Kenya's robust national HIV program and prioritization of high-burden counties, a critical evidence gap exists in predictive surveillance for pediatric HIV. Current strategies rely heavily on retrospective, aggregated data, which limits the ability to:

• Objectively predict pediatric HIV burdens in near real-time.
• Identify statistically significant spatial clusters (hotspots) that may be missed by traditional reporting.
• Equitably compare disease burden across counties with vastly different population sizes.

There is a need for an integrated analytic framework that combines Machine Learning (ML) for prediction with Geostatistics for spatial validation to guide precision interventions and resource allocation.

2. Methodology

The study employed a dual analytical approach integrating supervised machine learning with geostatistical modeling across Kenya's 47 counties.

Data Sources

• Outcome Data: National HIV Testing Services (HTS) data from October 1, 2022, to June 30, 2023, specifically tracking newly diagnosed children (aged 0–14 years).
• Predictor Data: County-level indicators from the 2022 Kenya Demographic and Health Survey (KDHS).
• Variables: 19 candidate predictors spanning maternal health (e.g., PMTCT seropositivity, ANC attendance), child health (e.g., stunting, malaria prophylaxis), sexual behavior, gender-based violence, and socioeconomic factors.
• Population Denominator: 2023 Kenya National Bureau of Statistics population projections for children aged 0–14.

Machine Learning Workflow

• Preprocessing: Missing values were imputed using Predictive Mean Matching (PMM). Variables with high multicollinearity (VIF > 5) were excluded.
• Algorithms: Three penalized regression models were trained and compared: Ridge (L2), Lasso (L1), and Elastic Net. These were chosen for their ability to handle multicollinearity and perform variable selection in high-dimensional data.
• Training: The dataset was split 70:30 (training/testing). Hyperparameters were tuned via 10-fold cross-validation using grid search.
• Performance Metrics: Models were evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).

Geostatistical Analysis

• Incidence Calculation: Predicted case counts were converted to incidence per 10,000 children to normalize for population size.
• Spatial Autocorrelation: Moran's I statistics were used to quantify global spatial clustering.
• Hotspot Detection: Getis-Ord Gi* statistics were applied to identify statistically significant clusters of high (hotspots) and low (coldspots) case density.
• Uncertainty Quantification: A residual-based bootstrap procedure (2,000 iterations) was used to generate 95% prediction intervals for county-level estimates.

Validation

• Frequentist: Welch's two-sample t-test and Hedges' g effect size compared predicted vs. reported distributions.
• Bayesian: Sensitivity analysis using a Jeffreys–Zellner–Siow (JZS) Cauchy prior to calculate the Bayes Factor ($BF_{01}$) supporting the null hypothesis of no difference.

3. Key Results

Model Performance

• Best Model: The Lasso regression model demonstrated superior predictive accuracy with the lowest error metrics (RMSE = 0.122, MAE = 0.099).
• Predictive Accuracy: The model predicted 3,160 new pediatric HIV cases, closely matching the 3,092 cases reported nationally (a 2.2% difference).
• Statistical Validation:
  • No significant difference was found between predicted and reported distributions (Welch's t = 0.11, p = 0.911).
  • Bayesian analysis yielded a Bayes Factor ($BF_{01}$) of ~4.57, providing moderate evidence in favor of the null hypothesis (no difference).
  • Uncertainty Calibration: 45 out of 47 counties (95.7%) had reported case counts falling within the model's 95% bootstrap prediction intervals, indicating well-calibrated uncertainty bounds.

Key Predictors

Feature importance analysis identified PMTCT HIV seropositivity as the strongest positive predictor. Other significant factors included:

• Positive associations: Severe stunting, number of Fansidar doses, and multiple sexual partners.
• Negative associations: Antenatal care (ANC) visits, male partner HIV testing, and women's property ownership.

Spatial Findings

• Clustering: Significant spatial autocorrelation was confirmed for both reported (Moran's I = 0.22, p = 0.001) and predicted (Moran's I = 0.37, p < 0.001) data.
• Hotspots: 13 counties were identified as statistically significant hotspots (95–99% confidence), predominantly in Western Kenya (Migori, Siaya, Homa Bay, Busia, Kisii, Bungoma, Kakamega, Kisumu, Bomet, Vihiga, Nandi, Kericho, Trans Nzoia).
• Coldspots: Laikipia and Nyeri were identified as coldspots.
• Incidence Disparities:
  • Highest Reported Incidence: Isiolo (11.2/10k), Homa Bay (7.7/10k).
  • Predicted vs. Reported Gaps: The model identified potential under-detection in counties like Tana River (Predicted 4.2 vs. Reported 1.0) and Lamu (Predicted 4.2 vs. Reported 2.8), suggesting areas for targeted testing expansion.

Outlier Analysis

Two counties fell outside their 95% prediction intervals, offering distinct programmatic insights:

1. Homa Bay: Reported cases (399) exceeded the upper bound (380). This suggests intensified programmatic success (e.g., effective index testing) detecting cases beyond demographic expectations.
2. Siaya: Reported cases (148) fell below the lower bound (189). This suggests under-detection or gaps in testing coverage relative to the underlying burden.

4. Key Contributions

1. Methodological Framework: Successfully validated a hybrid framework combining penalized ML (Lasso) with geostatistical hotspot analysis for pediatric HIV surveillance.
2. Population-Adjusted Incidence: Demonstrated the necessity of converting raw case counts to incidence per 10,000 children to enable equitable comparisons across counties of varying sizes, preventing the masking of high-risk, low-population areas.
3. Uncertainty Quantification: Introduced bootstrap-derived prediction intervals to operational decision-making, allowing health authorities to distinguish between random variation and genuine programmatic outliers (e.g., Homa Bay vs. Siaya).
4. Predictive Surveillance: Shifted the paradigm from retrospective reporting to predictive epidemiology, identifying potential "hidden" hotspots in Eastern and Northern regions where routine surveillance may be lagging.

5. Significance and Implications

• Resource Optimization: The framework enables health authorities to move from broad, retrospective targeting to precision interventions, directing testing and treatment resources to statistically validated hotspots and under-detection zones.
• Scalability: The approach is replicable and scalable to other infectious diseases and sub-Saharan African contexts where data is aggregated at the administrative level.
• Policy Integration: The study advocates for integrating these predictive models into national systems like DHIS2, the National Data Warehouse (NDW), and Case-Based Surveillance (CBS) to support adaptive, real-time epidemic response.
• Equity: By highlighting disparities in detection (e.g., Siaya's under-detection), the model supports more equitable allocation of resources to ensure no high-burden community is left behind.

Conclusion: This study provides a robust, data-driven tool for enhancing pediatric HIV surveillance in Kenya, proving that integrating machine learning with spatial statistics can significantly improve the accuracy of burden estimation and the efficiency of public health interventions.