Pioneering diabetes screening tool: machine learning driven optical vascular signal analysis

Area of Science:

• Biomedical Engineering
• Optical Sensing Technology
• Machine Learning Applications

Background:

• Diabetes mellitus prevalence is rising, necessitating non-invasive screening methods.
• Current diagnostic methods for diabetes are invasive, highlighting the need for accessible alternatives.
• Optical sensing and machine learning offer potential for analyzing physiological signals related to glycemic status.

Purpose of the Study:

• To develop and validate a non-invasive optical method integrated with machine learning for classifying individuals into normal, prediabetic, and diabetic categories.
• To engineer a novel device for capturing real-time optical vascular signals.
• To assess the efficacy of machine learning algorithms in differentiating glycemic states using optical signal features.

Main Methods:

• A novel device captured optical vascular signals from participants across three glycemic states.
• Signal quality assessment and preprocessing ensured data reliability.
• Time-domain analysis and wavelet scattering techniques were used for feature extraction.
• Ensemble bagged trees and random forest classifiers were trained and validated.

Main Results:

• The ensemble bagged trees classifier with wavelet scattering features achieved 86.6% accuracy.
• The random forest classifier with time-domain features achieved 80.0% accuracy.
• Both models demonstrated effectiveness in differentiating normal, prediabetic, and diabetic individuals.

Conclusions:

• The non-invasive optical-based approach combined with machine learning shows potential as a screening tool for diabetes mellitus.
• The achieved classification accuracy supports further investigation and validation in larger populations.
• This method offers a promising non-invasive alternative for initial diabetes detection.