Non-invasive estimation of hemoglobin using machine learning algorithms from i-PPG and PPG signals

Area of Science:

• Biomedical Engineering
• Medical Devices
• Machine Learning in Healthcare

Background:

• Non-invasive hemoglobin estimation is crucial for diagnosing and monitoring anemia.
• Photoplethysmogram (PPG) signals offer a promising avenue for non-invasive physiological measurements.
• Existing methods for hemoglobin estimation often require invasive procedures or specialized equipment.

Purpose of the Study:

• To compare the accuracy of hemoglobin estimation using two types of PPG signals: t-PPG (device-based) and i-PPG (smartphone-based).
• To evaluate the performance of various machine learning algorithms for non-invasive hemoglobin estimation.
• To validate the findings against clinical laboratory tests.

Main Methods:

• Collected t-PPG and i-PPG signals from 21 volunteers (aged 20-54).
• Filtered and processed PPG signals to remove noise and extracted optical attenuation-based features.
• Utilized machine learning algorithms including decision trees, boosted trees, bootstrap, NTanH, and XGBoost for hemoglobin estimation.
• Compared estimated hemoglobin values with clinical laboratory results.

Main Results:

• The XGBoost model demonstrated superior performance in estimating hemoglobin levels from both t-PPG and i-PPG signals.
• High R-squared values (0.9862 for t-PPG, 0.9812 for i-PPG) and low RMSE (0.0489 g/dL for t-PPG, 0.0329 g/dL for i-PPG) were achieved.
• Both t-PPG and i-PPG signals, when processed with machine learning, showed strong correlation with clinical laboratory measurements.

Conclusions:

• Non-invasive hemoglobin estimation using PPG signals, particularly with the XGBoost model, is accurate and reliable.
• Smartphone-based i-PPG offers a convenient and accessible alternative for hemoglobin estimation.
• These findings support the development of portable, non-invasive tools for hemoglobin monitoring in diverse healthcare settings.