A Data-Driven Approach for Non-Invasive Hemoglobin Prediction and Anemia Classification

Introduction: Anemia is a global health concern that affects millions of people worldwide and is primarily diagnosed through invasive blood tests. These methods, while accurate, are costly, time-consuming, and inaccessible in many remote areas. Non-invasive solutions using machine learning and physiological signals such as Photoplethysmography (PPG) offer a promising alternative. Objective: This study aims to develop a machine learning-based approach for non-invasive hemoglobin estimation and anemia classification. The goal is to evaluate the feasibility of using vital signs (SpO₂, heart rate) and PPG signals to provide an accessible and cost-effective solution for real-time anemia detection. Methods: The research utilized a dataset comprising patient vitals and hemoglobin levels. Empirical formulas were applied to compute PPG signals, and machine learning models were trained on these features. Two approaches were analyzed: one based on vital signs alone and another incorporating PPG signals. Models such as Gradient Boosting, Random Forest, and Linear Regression were evaluated using metrics like R² score, RMSE, and MAE. Results: The PPG-based approach achieved near-perfect accuracy with an R² score of 1.000, while the vitals-based approach showed high practical applicability with an R² score of 0.971 using Gradient Boosting. Anemia classification models achieved up to 99.54% accuracy, demonstrating the effectiveness of machine learning for non-invasive diagnostics. Conclusion: This study highlights the potential of machine learning in non-invasive hemoglobin estimation and anemia detection. While PPG-based models offer superior accuracy, vitals-based models are more practical for real-world applications due to sensor limitations. The findings pave the way for integrating non-invasive anemia detection into wearable devices and healthcare systems.