Smartphone-based Identification of Critical Levels of Glycated Hemoglobin A1c using Transdermal Optical Imaging

Hudda Salih, Si Wu, Evgueni Kabakov, Kang Lee, Weihong ZhouUTSC 2021 · DOI

Abstract

The global prevalence of diabetes is rising rapidly, necessitating improved management strategies for optimal glycemic control. Uncontrolled diabetes leads to severe complications, including cardiovascular disease and increased mortality. Current monitoring techniques, while effective, face limitations in accessibility and patient compliance. This study investigates the viability of Transdermal Optical Imaging (TOI) technology, specifically the smartphone-based application Anura, combined with advanced machine learning for non-invasive diabetes screening. We compared TOI-derived glycated hemoglobin A1c (HbA1c) classifications against results from FDA-approved blood immunoassays in a cohort of 513 participants.

Introduction

Effective diabetes management is often hindered by the invasive nature and cost of current monitoring tools. Daily blood glucose tests are volatile and require frequent needle pricks, while the "gold standard" HbA1c test requires laboratory blood draws. For the 463 million adults living with diabetes worldwide, these barriers frequently lead to poor compliance.

This research, published in the Journal of Natural Sciences, explores a "contactless" alternative. By utilizing Transdermal Optical Imaging (TOI) to capture subtle changes in facial blood flow, researchers developed a machine learning model to identify critical HbA1c levels without invasive sampling.

Methods

The study employed the Anura smartphone application to perform TOI.

Data Acquisition and Processing

Using an iPhone 7, researchers recorded 1-minute videos of participants' faces. The technology utilizes the translucency of human skin, allowing smartphone sensors to detect red light re-emitted from hemoglobin beneath the dermis. The team used "bit-plane slicing" to isolate high signal-to-noise ratio (SNR) data, focusing on facial blood flow patterns associated with cardiovascular activity.

Machine Learning Approach

A Random Forest (RF) classifier (a "kitchen sink" model) was trained to analyze these patterns. The model classified participants as "diabetic" or "non-diabetic" based on a clinical HbA1c threshold of 5.7%. Ground truth was established using FDA-approved blood immunoassays.

Study Population

The dataset comprised 513 participants (385 males, 128 females) recruited during annual physical examinations at the Health Management Center of The Affiliated Hospital of Hangzhou Normal University, China. Model validation involved 20 trials of training and testing, with "pristine testing" performed on 400 pseudo-randomly selected participants.

Results

The RF classifier demonstrated a proof-of-concept for non-invasive screening with the following performance metrics:

  • Classification Accuracy: 66%
  • Area Under the Curve (AUC): 0.69
  • True Positive Rate: 0.65
  • False Positive Rate: 0.34

While an AUC of 0.69 is considered "fair" in clinical diagnostics, it significantly outperforms random chance (0.50), indicating that facial blood flow contains detectable "optical signatures" correlated with HbA1c levels.

Discussion

The study suggests that facial blood flow is tightly coupled with cardiovascular health. Given that HbA1c accounts for approximately 97% of total hemoglobin, the researchers hypothesized that total hemoglobin concentration extracted via TOI serves as a viable proxy for diabetes classification.

Implications for Global Health

By utilizing standard consumer hardware like the iPhone 7, this technology suggests that the infrastructure for global health screening is already widely available. This could democratize diagnostics, providing an inexpensive "early warning system" that prompts at-risk individuals to seek professional medical intervention.

Limitations and Future Directions

Several challenges remain before clinical deployment:

  • Environmental Factors: The study was conducted under controlled lighting and positioning. Real-world performance with variable lighting and head movement requires further testing.
  • Demographic Calibration: The participant pool was primarily of East Asian descent. Future models must be calibrated across diverse skin tones to ensure accuracy and avoid algorithmic bias.
  • Granularity: Current models provide binary classification; future iterations aim to predict specific HbA1c percentages and identify prediabetic states.

Conclusion

This study provides evidence for the potential of TOI technology for contactless, non-invasive, and inexpensive assessments of diabetes risk. While not a replacement for clinical diagnosis, it offers a promising tool for ubiquitous screening and improved chronic disease management.

Summary updated

Abstract

Worldwide, the prevalence of diabetes has continued to increase rapidly. This gives rise to concerns regarding appropriate diabetes management to ensure optimal glycemic control. Untreated or uncontrolled diabetes can lead to a host of complications, such as cardiovascular diseases, an increased likelihood of morbidity and mortality (Deshpande, Harris-Hayes, & Schootman, 2008). A challenging problem which arises in diabetes management is the limitations of current blood glucose monitoring techniques. Electronic medical devices can potentially overcome the persistent problems in the healthcare industry. Thus, for this study, it was of interest to investigate whether advanced machine learning methods and Anura, a smartphone-based transdermal optical imaging technology (TOI) that assess health markers, can be a viable solution for diabetes management. Objectives: To examine the validity of TOI and a novel machine algorithm for diabetes prediction (i.e diabetes and non-diabetes). We compared the diabetes classification from TOI’s obtained glycated hemoglobin A1c (HbA1c) concentrations against data obtained from FDA approved blood immunoassay. Methods: In the present study, we used a kitchen sink random forest machine algorithm for diabetes prediction. The data set was obtained from 513 participants recruited during their annual physical examination at the Health Management Center of The Affiliated Hospital of Hangzhou Normal University, China. This included participant’s TOI and blood immunoassay determined HbA1c concentrations. To validate the model, pristine testing was done on 400 pristine participants pseudo randomly selected during 20 trials of training/testing. Results: The confusion matrix found TOI to have a classification accuracy of 66%, and the ROC curve of the RF classifier found TOI to have a ROC AUC of 0.69. Conclusions: The present study provides evidence for the potential use of the TOI technology, Anura, for contactless, non-invasive, and inexpensive assessments of diabetes.
BibTeX
@inproceedings{salih2021smartphone,
  author = {Hudda Salih and Si Wu and Evgueni Kabakov and Kang Lee and Weihong Zhou},
  title = {{Smartphone-based Identification of Critical Levels of Glycated Hemoglobin A1c using Transdermal Optical Imaging}},
  booktitle = {UTSC},
  year = {2021},
  doi = {10.33137/jns.v2i1.34645},
}
Figure