A Non-Invasive Approach for Hemoglobin Estimation

To facilitate hemoglobin value estimation noninvasively, a machine learning model is proposed in this paper. Features extracted from the acquired Photo-plethysmography (PPG) signals are used in the estimation of haemoglobin value using a Machine Learning model. Hemoglobin estimation is also done using modified Beer-Lambert's law along with a Random Forest Regression model. This model is chosen since it has the lowest mean absolute error and highest accuracy when compared to other Machine Learning models like LASSO Regression, Support Vector Regression, Ridge Regression, and ADA-BOOST Regression. The PPG dataset was obtained from 3000 subjects. The signals were recorded from a fingertip at 125Hz and pre-processed for baseline wander minimization, noise, and motion artifact removal. The PPG signal database consists of a cell array of matrices. Matlab helped in extracting 21 features from the PPG signal. Accuracy is enhanced by identifying how each feature is related to each other using a Correlation matrix and Heatmap, including derivatives and Spectral features of PPG signal. The dataset is divided into a training and testing set to train the Machine Learning model. After training with the training data set, the Hb values are estimated with test data as input. The random forest regression ML model used here provides an accuracy of 81% with a Mean Absolute Error of 0.88, Mean Squared Error of 1.76, variance score of 0.61, and root MSE of 0.43.