Hemoglobin Estimation and Anemia Severity Classification via PPG Signal Extracted from Mobile Fingertip Videos

Anemia is a severe health issue of worldwide concern with children and women in low-resource environments being disproportionately impacted. The traditional diagnostic tools, i.e., Complete Blood Count (CBC) tests, are accurate but invasive, expensive, and unavailable to many patients. This study provides a non-invasive method of hemoglobin measurement and anemia severity classification based on the analysis of smartphone fingertip videos, offering a low-cost solution for early screening in underserved areas. This study addresses major limitations of prior works, such as the low quality of signals, the lack of severity categorization, and imbalance of datasets, with a robust pipeline that combines photoplethysmography (PPG) signal extraction, advanced preprocessing, data augmentation, physiologically-aware feature engineering, and machine learning models. Leveraging a publicly available dataset of 150 fingertip videos, collected under institutional ethical approval with informed consent, the values of hemoglobin were manually classified into anemia classes in accordance with the WHO guidelines. The data augmentation (amplitude scaling, jittering, signal wandering) and dynamic frame selection algorithms have been integrated into the optimal methodology to increase signal fidelity. The XGBoost model performed better on hemoglobin regression (RMSE = 0.3896), and the ANN model performed better on anemia severity classification (accuracy = 0.9900). These findings outperform the results of the existing studies regarding this dataset. Future efforts will be aimed at increasing demographic diversity, real-world testing and development of mobile application to increase accessibility and impact.