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

Justin J. Boutilier, Timothy C. Y. Chan, Manish Ranjan, Sarang DeoJournal of Medical Internet Research 2021 · DOI

Abstract

In the next decade, health systems in many countries are planning to spend significant resources on noncommunicable disease screening programs and our study demonstrates that machine learning models can be leveraged by these programs to effectively utilize limited resources by improving risk stratification.

Introduction

Noncommunicable diseases (NCDs) like diabetes and hypertension now account for over 75% of NCD deaths globally, with a massive burden falling on low- and middle-income countries (LMICs) like India. Health systems in these regions face a two-fold challenge: extreme resource scarcity and a reliance on "one-size-fits-all" Western risk scores.

Existing tools, such as the American Diabetes Association (ADA) or Framingham scores, were developed using Western populations to predict long-term (e.g., 10-year) risk. These models often fail to generalize to the social, genetic, and lifestyle realities of urban slums and are not optimized for the immediate, actionable "short-term risk" that community health workers (CHWs) need to prioritize doctor referrals.

Methodology and Data

Researchers developed machine learning models specifically tailored to predict the short-term (2-month) risk of diabetes and hypertension in low-resource urban settings. The study utilized a dataset of 2,278 patients collected by CHWs in the urban slums of Hyderabad, India. This cohort was derived from an initial screen of 51,474 individuals, specifically selecting those who visited a physician within 2 months to provide a "ground truth" diagnosis.

The authors compared five common machine learning classifiers: Decision Trees, Regularized Logistic Regression, K-Nearest Neighbors (KNN), Random Forest, and AdaBoost. A key feature of the study was incremental feature analysis, testing performance across five levels of data availability, from simple questionnaires to the inclusion of blood glucose measurements. They also performed a cost analysis to calculate the financial impact of incorrect risk stratification (false positives and false negatives).

Key Results

The Random Forest model consistently outperformed other classifiers, likely due to its ability to handle complex, non-linear relationships in health data. It achieved an AUC of 0.910 for diabetes and 0.792 for hypertension.

Comparison of AUC for diabetes prediction: Random Forest vs. traditional and retrained scoring algorithms.

As illustrated in the performance data, localized ML models significantly outperform traditional clinical scores, even when those scores are retrained on local data. For diabetes, the Random Forest model provided a 35.5% improvement over the best-performing baseline (retrained UK risk score).

Comparison of AUC for hypertension prediction: Random Forest vs. original and retrained Framingham scores.

Economic and Clinical Impact

  • Reducing Missed Cases: At a fixed specificity of 0.9, the model reduced the expected number of false negatives by 620 per 1000 screenings for diabetes and 220 per 1000 for hypertension.
  • Financial Savings: Improved stratification reduces the cost of incorrect diagnoses by US $1.99 (35%)** per diabetes screening and **US $1.60 (21%) per hypertension screening. Scaled nationally, this represents billions in potential savings.

Discussion and Limitations

While the results are promising, the study notes a few limitations. The dataset only included individuals who followed up with a doctor, which may introduce selection bias. Furthermore, the models were developed for urban slums and may require further adjustment for rural populations. The authors emphasize that the next hurdle is the "last mile" problem: ensuring that identified high-risk patients actually receive and adhere to treatment through better mHealth integration.

Conclusion

By moving away from generalized Western models toward localized machine learning, health systems in resource-limited settings can effectively leverage limited resources to improve NCD risk stratification and early detection.

Summary updated

Abstract

In the next decade, health systems in many countries are planning to spend significant resources on noncommunicable disease screening programs and our study demonstrates that machine learning models can be leveraged by these programs to effectively utilize limited resources by improving risk stratification.
BibTeX
@article{boutilier2021risk,
  author = {Justin J. Boutilier and Timothy C. Y. Chan and Manish Ranjan and Sarang Deo},
  title = {{Risk Stratification for Early Detection of Diabetes and Hypertension in Resource-Limited Settings: Machine Learning Analysis}},
  journal = {Journal of Medical Internet Research},
  year = {2021},
  doi = {10.2196/20123},
}
Figure