Risk Stratification for Early Detection of Diabetes and Hypertension in Resource-Limited Settings: Machine Learning...

Area of Science:

• Public Health
• Medical Informatics
• Machine Learning

Background:

• Noncommunicable disease (NCD) screening programs are scaling up in low- and middle-income countries (LMICs).
• Limited health resources in LMICs necessitate highly accurate risk stratification for NCD screening.
• Effective identification of high-risk individuals is crucial for resource optimization.

Purpose of the Study:

• To develop machine learning (ML)-based risk stratification algorithms for diabetes and hypertension.
• To tailor algorithms for at-risk populations in community-based screening programs within low-resource settings.
• To enhance the accuracy of identifying individuals at high risk for diabetes and hypertension.

Main Methods:

• Trained and tested ML models using data from 2278 patients in urban slums of Hyderabad, India.
• Utilized data collected by community health workers via door-to-door and camp-based screenings.
• Compared ML model performance against established US and UK risk scores using Area Under the Curve (AUC) and false negatives.

Main Results:

• Random forest models demonstrated superior prediction accuracy for both diabetes and hypertension.
• ML models achieved a 35.5% higher AUC for diabetes and 13.5% for hypertension compared to existing scores.
• Models reduced false negatives by 620/1000 for diabetes and 220/1000 for hypertension, decreasing costs by up to 35% per screening.

Conclusions:

• ML models can significantly improve risk stratification in NCD screening programs.
• Leveraging ML enables more effective utilization of limited health resources in LMICs.
• This approach enhances the efficiency and cost-effectiveness of identifying individuals at high risk for NCDs.