Noninvasive Diabetes Detection through Human Breath Using TinyML-Powered E-Nose

Area of Science:

• Biomedical Engineering
• Computational Biology
• Analytical Chemistry

Background:

• Volatile organic compounds (VOCs) in exhaled breath are key biomarkers for disease detection.
• Acetone in breath is a significant noninvasive biomarker for diabetes mellitus.
• Current electronic nose (e-nose) systems for diabetes detection often require complex computational environments.

Purpose of the Study:

• To develop an embedded system for real-time, noninvasive diabetes mellitus detection.
• To integrate electronic nose technology with Tiny Machine Learning (TinyML).
• To evaluate the diagnostic performance of the integrated system.

Main Methods:

• Developed an embedded system combining an e-nose with Metal Oxide Semiconductor (MOS) sensors.
• Integrated TinyML algorithms for real-time data analysis.
• Tested the system on 44 participants (22 healthy, 22 with diabetes).
• Employed XGBoost and deep learning algorithms (DNN, 1D-CNN).

Main Results:

• The XGBoost algorithm achieved 95% accuracy in diabetes detection.
• Deep learning algorithms (DNN, 1D-CNN) achieved 94.44% detection efficacy.
• The embedded system demonstrated effective real-time analysis of breath biomarkers.

Conclusions:

• Combining e-noses with TinyML in embedded systems offers a potent noninvasive approach for diabetes mellitus detection.
• The developed system shows promise for accessible and efficient diabetes screening.
• Real-time biomarker analysis via integrated e-nose and TinyML systems is feasible for medical diagnostics.