A mobile platform for non-invasive diabetes screening

Abstract

The global burden of diabetes is profound, with over 400 million cases worldwide. Middle-income countries, such as those in Southeast Asia, are especially affected as diabetes becomes a major public health issue. In particular, India has a record number of people diagnosed with diabetes, 8.8% of the adult population. Despite the benefits of early prevention and treatment, awareness of the disease remains low, with health workers unable to keep up with the demands of the population. In addition to simply screening for diabetes, it is also important to assess a person's severity of diabetes so that the proper intervention and therapy can be delivered. Current screening techniques rely on blood tests, often fingerprick blood glucose tests, but even though portable glucose meters are fairly inexpensive ($10 USD), they require a reliable stock of glucose test strips which are relatively costly ($0.05 USD) and are often in short supply.

As a possible alternative for diabetes screening, our group has designed a mobile platform that integrates various non-invasive tests for diabetes including clinical questionnaires, thermal imaging, iris imaging, retina imaging and photoplethysmography. For the purposes of evaluation, we have defined six stages of diabetes progression: stage 0 (no diabetes), stage 1-3 (prediabetes), stage 4 (diabetes), and stage 5 (advanced diabetes). The data collection and assessment platform includes an Android mobile application and a server to store and process the measurements and return the results to the Android client and web client. In this thesis, I describe the development of the server API for this platform, as well as the development of a Bayesian network model that is used to process the data and predict the specific stage of diabetes.

As a sample real-world implementation of this platform, our team has begun a large-scale diabetes study in two dierent sites in India, one in Mumbai and one in the Bangalore area. Based on data from this field study, I developed models for three dierent stages of diabetes pathogenesis: stage 0-3 (no diabetes to prediabetes), stage 4 (diabetes) and stage 5 (advanced diabetes). The performance of each model was evaluated using the area under the ROC curve (AUC). The best performing model was a Bayesian network that integrated questionnaire and iris data. Preliminary results for this model show high differentiation for each stage, with AUC scores of 1.0, sensitivity scores for each stage above .67 and specicity scores for each stage above .68. While data collection is still ongoing, these early results are encouraging and show a promising path for future large-scale diabetes screening.