Non-Invasive Estimation of Hemoglobin Using a Multi-Model Stacking Regressor.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Medical Diagnostics

Background:

• Anemia, particularly iron-deficiency anemia, is a significant global health issue, especially affecting women of childbearing age.
• Accurate and accessible hemoglobin (Hb) measurement is crucial for diagnosis and management.
• Current methods often involve invasive blood sampling, limiting widespread screening.

Purpose of the Study:

• To develop and validate a novel machine learning-based method for non-invasive total hemoglobin estimation using photoplethysmograms (PPGs).
• To assess the feasibility of this method for maternal anemia detection and its potential as a public health screening tool.

Main Methods:

• A study involving 1583 women in Karnataka, India, collected photoplethysmogram (PPG) signals at four wavelengths (590, 660, 810, 940 nm) using a custom finger sensor.
• A novel feature vector was derived from the PPG signals.
• A machine learning model, a two-layer stack of regressors (LASSO, Ridge, Elastic Net, AdaBoost, SVR), was designed and tested.

Main Results:

• The proposed machine learning method achieved a statistically significant Pearson's correlation coefficient (PCC) of 0.81 (p < 0.01) with gold standard Hb values.
• The Root Mean Square Error (RMSE) was 1.353 ± 0.042 g/dL.
• The stacked regressor model outperformed individual regressors, and including pregnant women in training data improved performance.

Conclusions:

• The study demonstrates the feasibility of a machine learning-based non-invasive hemoglobin measurement system.
• This technology holds significant potential for maternal anemia detection and screening.
• The approach could serve as a basis for a public health tool to supplement global health interventions.