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

Abstract

We describe a novel machine-learning based method to estimate total Hemoglobin (Hb) using photoplethysmograms (PPGs) acquired non-invasively. In a study conducted in Karnataka, India, 1583 women (pregnant and non-pregnant) of childbearing age, with Hb values ranging between 1.6 to 14.8g/dL, had their Hb values estimated using intravenous blood samples and concurrently by a finger sensor custom designed and prototyped for this study. The finger sensor collected PPG signals at four wavelengths: 590nm, 660nm, 810nm, and 940nm. A novel feature vector was derived from these PPGs. A machine learning model comprising of a two-layer stack of regressors including Least Absolute Shrinkage and Selection Operator (LASSO), Ridge, Elastic Net, Adaptive (Ada) Boost and Support Vector Regressors (SVR) was designed and tested. We report a statistically significant Pearson's correlation coefficient (PCC) of 0.81 (p < 0.01) between the Hb value estimated by the proposed methodology and gold standard values of Hb, with a Root Mean Square Error (RMSE) of 1.353 ± 0.042g/dL. The performance of the stacked regressor model was significantly better than the performance of individual regressors (low RMSE, and better CC; p < 0.05). Post-hoc analysis showed that including pregnant women in the training data set significantly improved the performance of the algorithm. This article demonstrates the feasibility of a machine learning based non-invasive hemoglobin measurement system, especially for maternal anemia detection. By developing and demonstrating a machine learning approach on a large data set, we have demonstrated that such an approach could become the basis for a public health screening tool to detect and treat maternal anemia and could supplement global health intervention strategies.