Towards a Multi-model Comparative Study for Non-invasive Hemoglobin Level Prediction from Fingertip Video

Abstract

Hemoglobin is one of the most significant hematological parameters, bearing valuable diagnostic information on potential diseases. Early detection of severely low hemoglobin levels can save a life, more so for women and children. Advancements in technology have brought several hemoglobin level measurement techniques to the fold and increasing prediction accuracy has become paramount to address the looming needs of a large part of the world population. Considering that, we have performed an exploratory multi-model comparative study using seven different models. Understanding the appropriate usage of generative AI in the field of mobile health can drastically progress the healthcare ecosystem. We leverage the Generative AI’s power to generate synthetic data and augment it with the actual dataset of real patients to improve the performance of models. We define it as the ‘combined’ dataset since it joins the real patient’s data with the generated data. We propose a Stacked Regression with hyperparameters that notably improves prediction on the combined dataset. The Improved Stacked Regression model, our best model, has the lowest Mean Absolute Error (MAE) of 0.68 g/dL and 6% Mean Absolute Percentage Error (MAPE), indicating significantly better performance on the combined data set. It also had an RMSE of 0.96 g/dL, while the second-best model, Lasso Regression, had an RMSE of 1.05 g/dL on the baseline data. Our proposed system can estimate hemoglobin levels and serve as an alternative to the costly, time-consuming, and inconvenient gold standard tests.