Estimating Plasmodium falciparum Parasite Rate using Test Positivity Rate from 2016-2024: Health Management Information Systems in Uganda

This study demonstrates that Health Management Information System data, specifically test positivity rates, can be effectively triangulated with survey data using multi-level logistic regression to generate high-resolution, monthly estimates of *Plasmodium falciparum* parasite rates in Uganda, offering a cost-effective and resilient alternative for malaria surveillance as community survey frequency declines.

The Gist

Imagine Uganda is a giant, bustling city where a sneaky thief named Malaria is constantly trying to break into homes. The city's health officials (the National Malaria Control Program) need to know exactly where the thief is hiding, how many people he has caught, and when he is most active so they can send the right police force to the right neighborhoods.

For a long time, the only way to find the thief was to send out a massive, expensive team of detectives to knock on every door in the city. These were the Surveys (like the DHS and MIS). While these detectives were very accurate, they were slow, expensive, and only showed up once every few years. By the time they reported back, the thief might have moved to a different neighborhood.

The Problem:
The health officials needed a way to track the thief in real-time, every single month, without waiting for the expensive detectives to return. They had a different source of information: the Health Clinic Logs (HMIS). Every day, thousands of sick people go to clinics, get tested, and the results are recorded. This data is updated constantly, but it's messy. It's like looking at a security camera feed that only shows people who chose to walk into the station. It doesn't show everyone in the city, and sometimes the camera glitches or the lighting is bad.

The Solution: The "Translator" Model
This paper is about building a clever translator that turns those messy clinic logs into a clear map of where the malaria thief is hiding.

Here is how they did it, using a simple analogy:

1. The "Test Positivity Rate" (The Clue)

Every time someone goes to a clinic, they take a malaria test. The Test Positivity Rate (TPR) is simply the percentage of those tests that come back positive.

• Analogy: Imagine a fishing net. If you cast your net and catch 10 fish, and 5 of them are the "thief" (malaria), your catch rate is 50%. If you only catch 1 fish and it's the thief, your rate is 100%.
• The Catch: This rate changes based on how many people show up to the clinic, not just how many thieves are in the city. If everyone stays home because they are scared of the clinic, the rate looks weird.

2. The "Severe Cases" (The Intensity Meter)

The researchers realized that just looking at the catch rate wasn't enough. They added a second clue: How many of the caught patients were very sick?

• Analogy: If you catch a few small fish, it's one thing. But if you start catching giant, dangerous sharks, you know the ocean is in trouble. The number of "severe" malaria cases acts like a warning siren that tells the model, "Hey, the situation here is getting serious!"

3. The "Training" (Teaching the AI)

The researchers took their "Translator" model and taught it using the data from the expensive, accurate Detective Surveys (the gold standard). They showed the model: "Here is what the clinic logs looked like in 2016, 2017, and 2018. And here is what the real survey said the malaria rate was. Learn the pattern."

They found that if they smoothed out the data (looking at a 6-month average instead of just one day) and looked at the "severe cases," the model could guess the real malaria rate with 79% accuracy. That's like a weather forecaster who is right almost 8 out of 10 times!

What Did They Discover?

Once the model was trained, they let it run on data from 2016 to 2024. Here is what the "Translator" revealed that the old surveys missed:

• The Thief Moves Fast: The model showed that malaria spikes and dips every month, often following the rainy seasons (like a tide coming in and out). The old surveys, which only happened once a year, completely missed these monthly waves.
• The "2018 Dip": The model spotted a massive drop in malaria in 2018 across the whole country. The surveys were too sparse to see this clearly, but the clinic logs showed it instantly.
• The "West Nile" Success Story: In the West Nile region, the model showed a huge drop in malaria in 2024. Why? Because the government launched a massive mosquito-spraying campaign. The model proved it worked in real-time, allowing officials to celebrate and move resources elsewhere.
• Hidden Hotspots: Sometimes a whole region looks "okay" on average, but the model showed that one specific district inside that region was actually a hotbed of malaria. It's like a classroom that looks fine on average, but one student is failing miserably. The model found that student.

Why Does This Matter?

Think of this model as a GPS for Malaria.

• Before: Health officials were driving blind, waiting for a map that was 2 years old.
• Now: They have a live GPS that updates every month.

This is crucial because big surveys are becoming harder to do (due to cost and political issues). This paper proves that we don't need to wait for the expensive detectives anymore. We can use the data we already have every day in the clinics to make smart, fast decisions.

In a nutshell: The researchers built a smart calculator that turns messy clinic test results into a clear, monthly map of malaria. This helps Uganda fight the disease faster, cheaper, and more effectively, saving lives by catching the "thief" before he can strike again.

Technical Summary

1. Problem Statement

Malaria transmission in Uganda is highly heterogeneous, requiring high-resolution, frequent data to support Sub-National Tailoring (SNT) of interventions. However, the current standard for measuring transmission intensity—the Plasmodium falciparum parasite rate (PfPR)—relies on population-based surveys (e.g., DHS, MIS). These surveys are:

• Infrequent and Sparse: Conducted every few years, leaving gaps in temporal monitoring.
• Age-Restricted: Typically limited to children under five, missing transmission dynamics in older populations.
• Costly: Difficult to scale for routine, real-time decision-making.

Conversely, Health Management Information Systems (HMIS) provide routine, passive surveillance data (Test Positivity Rate or TPR) with high spatial and temporal resolution. However, TPR is inherently biased by healthcare-seeking behavior, diagnostic testing practices, and non-malarial febrile illnesses, making it an unreliable direct proxy for population prevalence without calibration.

The Core Challenge: How to translate biased, routine HMIS TPR data into accurate, high-resolution estimates of PfPR to enable real-time malaria program management in the absence of frequent surveys.

2. Methodology

The study developed a statistical modeling framework to triangulate HMIS data with survey data to predict monthly PfPR at the district level.

• Data Sources:

  • Training Data (Ground Truth): Individual-level data from three nationally representative surveys (2016 DHS, 2018 MIS, and 2017–2019 LLINEUP) covering 45,396 children under five.
  • Predictor Data: Routine HMIS data from ~4,000 health facilities (2016–2024), including total tests, confirmed cases, severe cases, and outpatient visits.
  • Covariates: Environmental data (rainfall, temperature), population density, and median walk time to facilities were explored.

• Data Preprocessing:

  • Imputation & Cleaning: An algorithm removed implausible values and imputed missing data for HMIS variables.
  • Filtering: Only facilities reporting >50% of the time period were included.
  • Smoothing: A 180-day rolling average was applied to TPR to reduce noise from short-term fluctuations in testing rates.
  • Alignment: Data were matched by district, month, and year.

• Statistical Model:

  • Type: Multi-level (mixed-effects) logistic regression.
  • Outcome: Binary PfPR status for children under five (0/1).
  • Predictors:
    • Primary: 180-day smoothed TPR (all ages).
    • Secondary: Proportion of severe malaria cases (inpatient/total confirmed).
    • Exploratory: Temperature and rainfall (included in initial models but excluded from the final operational model due to minimal gain in predictive power vs. complexity).
  • Random Effects: Clustering at the district and region levels.
  • Selection Criteria: Model selection was based on Akaike Information Criterion (AIC), prioritizing models that balanced predictive accuracy with the feasibility of routine data updates.

• Validation:

  • Predicted PfPR was validated against observed survey PfPR using Pearson's correlation.
  • Confidence intervals were generated via bootstrapping (1,000 simulations).
  • National and regional estimates were derived by population-weighting district-level predictions.

3. Key Results

• Model Performance:

  • The final model (Smoothed TPR + Proportion of Severe Cases) achieved a strong positive correlation with observed survey PfPR (Pearson's rho = 0.79, p < 0.001).
  • This performance was nearly identical to a more complex model including environmental covariates (rho = 0.81), suggesting that TPR alone captures most seasonal transmission dynamics.
  • The model successfully predicted monthly PfPR for all 112+ districts from 2016 to 2024.

• Temporal and Spatial Insights:

  • Seasonality: The model captured seasonal peaks corresponding to rainy seasons (March–May, September–November) which are often missed by sparse survey data.
  • Intervention Impact:
    • A significant drop in prevalence in 2018 was observed across regions.
    • In West Nile, prevalence dropped from >30% to 8.7% in 2024, correlating with intensified vector control (IRS) starting in 2022.
    • In Lira, prevalence increased after IRS efforts ceased in 2022.
  • Heterogeneity: The model revealed significant intra-regional variation. For example, while the Lango region had a modest regional average (26.4%), district-level estimates ranged from 9.2% to 51% within the same year.

• Survey Comparison:

  • National estimates derived from the model aligned well with survey estimates for 2017–2019 and 2024.
  • Discrepancy: The 2016 DHS estimate (30.7%) was nearly double the model's prediction (17.1%). The authors attribute this to poor HMIS data quality during the system's infancy in 2016 and potential RDT overestimation in the survey compared to microscopy.

4. Key Contributions

1. Operationalizing HMIS for Prevalence Estimation: The study demonstrates that routine facility data, when calibrated, can serve as a proxy for population-level parasite rates, filling the temporal gaps between surveys.
2. Bias Mitigation Strategy: By incorporating the proportion of severe cases as a covariate, the model accounts for variations in care-seeking behavior and the presence of non-malarial febrile illnesses, improving the robustness of TPR as a predictor.
3. High-Resolution Surveillance: The approach generates monthly, district-level estimates, enabling the National Malaria Control Program (NMCP) to detect local outbreaks or declines in transmission that are invisible in annual regional or national aggregates.
4. Cost-Effective Alternative: The study provides a scalable, low-cost methodology for malaria surveillance that does not rely on expensive, logistically complex household surveys.

5. Significance and Implications

• Programmatic Decision Making: The model enables "near real-time" monitoring, allowing programs to adapt interventions (e.g., IRS, LLIN distribution) dynamically based on current transmission levels rather than historical survey data.
• Resilience in Resource-Limited Settings: As political instability and funding constraints threaten the continuity of large-scale surveys (like DHS/MIS), this HMIS-based modeling offers a resilient alternative to maintain surveillance continuity.
• Data Quality Improvement: The reliance on HMIS data incentivizes the identification of facilities with unreliable reporting, creating a feedback loop to improve overall data quality.
• Global Applicability: The methodology is applicable to other high-burden, high-transmission settings where HMIS infrastructure exists but survey frequency is declining.

Conclusion: The study validates that a simple, updateable statistical model using routine HMIS data (TPR and severe case proportions) can accurately estimate malaria prevalence. This shifts the paradigm from sporadic, static survey-based planning to continuous, dynamic, and data-driven malaria control.