A Rule-Based Machine Learning Model for Predicting Virological Failure Among Children Living With HIV in Malawi

This study presents a machine learning framework combining Random Forest feature selection, association rule mining, and k-prototype clustering to identify key drivers and distinct risk profiles of virological failure among children living with HIV in Malawi, offering a solution to overcome delays in the country's resource-constrained treatment monitoring system.

The Gist

Imagine Malawi is running a massive, life-saving operation to keep children with HIV healthy. They give them special medicine (ART) every day to stop the virus from multiplying. The goal is to keep the virus "asleep" (suppressed). But sometimes, the virus wakes up and starts fighting back. This is called Virological Failure.

Right now, doctors have to wait months to run a lab test to see if the virus has woken up. It's like waiting for a smoke alarm to go off after the house is already on fire. By the time they know, the damage is done, and the child gets sicker.

This paper is about building a digital "early warning system" using Artificial Intelligence (AI) to predict which children are at risk before the virus wakes up.

Here is how the researchers did it, explained with simple analogies:

1. The Problem: The "Blind Spot"

In Malawi, there aren't enough experts to manually check every single patient's file to find patterns. Plus, the lab tests are slow. The researchers wanted to use a computer to look at thousands of patient records and say, "Hey, look at this group of kids. They share these specific traits. They are very likely to have the virus wake up soon."

2. The Tools: Two Different Detective Styles

The researchers used two different types of AI "detectives" to solve the mystery:

Detective A: The "Rule Finder" (Association Rule Mining)

Think of this detective as a super-observant librarian who notices that certain books are always checked out together.

• How it works: The computer looked for specific combinations of facts that almost always lead to a problem.
• The Discovery: It found "recipes" for failure. For example, it found a rule that says:

  "If a child is between 10 and 14 years old, has been on medicine for more than 5 years, has Tuberculosis (TB), and is on a specific drug mix, there is a 92% chance their virus will wake up."

• The Analogy: It's like a weather forecast that says, "If it's windy, cloudy, and the barometer is dropping, it will rain 9 out of 10 times." The computer found these "rainy day" patterns for HIV treatment.

Detective B: The "Grouping Expert" (Clustering)

Think of this detective as a teacher trying to sort students into different study groups based on how they behave.

• How it works: Instead of looking for specific rules, the computer looked at all the kids who failed treatment and asked, "Who looks like who?" It sorted them into two distinct "tribes" or profiles.
• The Discovery: It found two main types of "at-risk" kids:
  1. The Younger Group: These are younger kids (around 12) who are often underweight but surprisingly have high immune counts. They are on specific older medicines.
  2. The Older Group: These are teenagers (around 17) who are struggling more. They have lower immune counts, are on newer medicines, and often have TB.
• The Analogy: It's like realizing that in a school, the kids who fail math fall into two groups: those who just need a tutor (Group 1) and those who are skipping class entirely (Group 2). You need to help them in different ways.

3. The Clues: What Makes the Virus Wake Up?

The computer analyzed the data and found the biggest "red flags" that predict failure. Imagine these as the ingredients in a "bad recipe":

• BMI (Body Weight): This was the #1 clue. If a child is too thin or has weight issues, the medicine doesn't work as well.
• Time on Medicine: The longer a child has been on the same medicine (over 5 years), the higher the risk. It's like a car engine wearing out after too many miles; the virus might be getting used to the medicine.
• Age (10–14 years): This is the tricky "teenage transition" phase. Kids this age are trying to be independent, might forget pills, or feel embarrassed about their status, leading to missed doses.
• Tuberculosis (TB): Having TB is like having a second enemy fighting inside the body, making it harder for the HIV medicine to do its job.
• The Drug Type: Some specific combinations of drugs (like the "13A" regimen) were linked to higher failure rates.

4. Why This Matters

Before this study, doctors had to wait for a lab test to confirm a child was failing treatment. By then, the child might be very sick.

With this new "Early Warning System":

• Doctors can act fast: If a child fits the "Rule" (e.g., a 12-year-old with TB on a specific drug), the doctor knows immediately to switch their medicine or give them extra support before the virus gets worse.
• Targeted Help: They can treat the "Younger Group" differently than the "Older Group." Maybe the younger ones need better nutrition, while the older teens need counseling to help them remember their pills.
• Saving Lives: It stops the virus from spreading to others and prevents the child from getting sick.

The Bottom Line

This paper is about teaching a computer to be a super-doctor's assistant. It looks at the messy, complicated details of a child's life (age, weight, other diseases, how long they've taken meds) and says, "Be careful, this specific combination of factors usually leads to trouble."

It's a shift from reacting to a crisis (waiting for the fire alarm) to preventing the crisis (smelling the smoke and putting it out early). This is a huge step forward for keeping Malawi's children healthy.

Technical Summary

1. Problem Statement

Malawi's HIV treatment monitoring system faces critical bottlenecks, including a shortage of medical experts and a reliance on viral load testing intervals of 3 to 12 months. This delay in detecting virological failure (VF) (defined as viral load $\ge$ 1000 copies/mL) leads to increased risks of disease progression, transmission, and mortality.

• The Gap: While Malawi has achieved high viral suppression rates overall (94%), there is a significant disparity between adults (96.9%) and children/adolescents (81%).
• The Challenge: Manual analysis of patient data to identify those at risk of VF is impractical due to the volume of cases. Existing systems lack predictive models to identify complex, non-linear combinations of risk factors before a patient fails treatment.

2. Methodology

The study employed a Design Science Research (DSR) methodology, utilizing a hybrid machine learning framework that combines supervised feature selection, unsupervised clustering, and association rule mining.

A. Data Source and Preprocessing

• Dataset: Retrospective data from 2022–2023 involving 5,137 Children Living with HIV (CLHIV) across 260 health facilities in Malawi's southern region.
• Ethics: Data was fully anonymized; the study received ethical approval from MUSTREC.
• Preprocessing:
  • Cleaning: Removal of duplicates and empty rows.
  • Binning (Discretization): Continuous variables were converted to categorical for rule mining:
    • CD4 Count: Low vs. High.
    • BMI: Underweight, Normal, Overweight, Obese.
    • ART Duration: <12 months, 1–3 years, 4–5 years, >5 years.
    • Adherence: Poor (<80%) vs. Good (>80%).
    • Age: Developmental groups (<5, 5–9, 10–14, 15–17).
  • Feature Selection: A Random Forest classifier was used to rank feature importance and reduce dimensionality.

B. Model Architecture

1. Feature Importance (Random Forest): Identified the most predictive variables for VF.
2. Association Rule Mining (Apriori Algorithm):
   • Transformed data into a boolean matrix (one-hot encoding).
   • Parameters: Minimum Support = 0.00095; Minimum Confidence = 90%.
   • Goal: To discover rules of the form $Antecedent \implies Consequent$ (where Consequent = Virological Failure).
   • Metrics: Support, Confidence, and Lift (to measure association strength).
3. Clustering (k-Prototypes Algorithm):
   • Selected because it handles mixed data types (numeric and categorical).
   • Goal: To identify distinct risk profiles (clusters) among patients experiencing VF.
   • Optimization: Used the Cao method for initial centroids.
   • Evaluation: Silhouette Score and Cluster Purity.

3. Key Results

A. Feature Importance

The Random Forest analysis identified Body Mass Index (BMI) as the single most important predictor (Score: 0.23), followed by:

1. Duration on ART (0.197)
2. CD4 Count (0.173)
3. TB Status (0.120)
4. Age (0.113)
   Other significant factors included ART regimen, gender, and adherence.

B. Association Rule Mining

The model generated 28 high-confidence rules (100% of rules met the 90% confidence threshold).

• Top Performance: The top 10 rules showed an Average Confidence of 90.9% and an Average Lift of 4.81 (Max Lift: 4.89).
• Interpretation: A lift of ~4.9 indicates that the co-occurrence of specific risk factors and virological failure happens nearly 5 times more often than by random chance.
• Key Rules Identified:
  • Rule 1 & 3: CLHIV aged 10–14, on ART >5 years, with Confirmed TB, on specific regimens (13A or TDF/3TC/DTG) have a 92.3% probability of VF.
  • Rule 5 & 6: CLHIV with CD4 200–349 cells/mm³, on ART 1–3 years, using Regimen 15PA (ABC/3TC), have a 92.3% probability of VF.

C. Clustering Analysis (Risk Profiling)

The k-prototypes algorithm successfully identified two distinct risk profiles among patients with VF:

• Cluster 0 (61 cases, 6.3%):
  • Profile: Younger children (~12 years), Underweight, High CD4 counts (Median 1000), No TB suspicion, High adherence, on 13A/ABC regimens.
  • Insight: Represents a unique subgroup where VF occurs despite high immunity and good adherence, potentially linked to malnutrition (low BMI).
• Cluster 1 (909 cases, 93.7%):
  • Profile: Older adolescents (~17 years), Lower CD4 counts (Median 250), Mixed TB status (includes all TB cases), Mixed adherence, on 15A/TDF/3TC/DTG regimens.
  • Insight: Represents the broader, heterogeneous adolescent population facing standard challenges of long-term treatment and co-infection.
• Performance: The clustering achieved 100% purity (all cases in clusters were VF cases) and a Silhouette Score of 79%, indicating excellent cluster separation.

4. Key Contributions

1. Hybrid Framework: Successfully integrated Random Forest (for feature selection), Apriori (for rule extraction), and k-Prototypes (for profiling) to tackle a complex public health problem.
2. High-Confidence Predictive Rules: Established specific, actionable clinical rules with >90% confidence, moving beyond general risk factors to precise patient scenarios (e.g., specific age + TB status + ART duration).
3. Discovery of Anomalous Profiles: The clustering revealed a counter-intuitive profile (Cluster 0) where children with high CD4 counts and good adherence still experienced VF, highlighting the critical role of BMI/Nutrition which is often overlooked in standard monitoring.
4. Contextual Relevance: Provided the first machine learning-driven risk stratification specifically for CLHIV in Malawi, addressing the specific gap in pediatric viral suppression rates.

5. Significance and Limitations

Significance:

• Clinical Decision Support: The model offers a tool for clinicians to identify high-risk patients before the next scheduled viral load test, enabling early intervention.
• Resource Optimization: Helps prioritize limited medical resources for the most vulnerable subgroups (e.g., adolescents with TB co-infection or underweight children).
• Policy Impact: Highlights the need to monitor BMI and nutritional status alongside viral loads and CD4 counts for children on ART.

Limitations:

• Data Scope: The study relied solely on clinical data from health facilities. It lacked social determinants of health (e.g., household food security, caregiver education, economic stability) which are known to impact adherence and outcomes.
• Retrospective Nature: As a retrospective study, it identifies associations rather than proving causality.
• Generalizability: The model was trained on data from the southern region of Malawi; external validation in other regions or countries is required.

Conclusion

The study demonstrates that machine learning, specifically the combination of association rule mining and clustering, is highly effective in predicting virological failure in Malawian children. By identifying that BMI, TB status, and ART duration are critical drivers, and by uncovering distinct risk profiles, the framework provides a quantitative, evidence-based approach to improving HIV care and achieving the UNAIDS 95-95-95 targets for children.