Key result
A non-invasive hemoglobin measurement system using multiwavelength PPG and a BPNN model achieved a mean absolute error of 0.421 g/dL, 0.624 g/dL, and 1.57 g/dL across three datasets.
Why the study?
Continuous hemoglobin monitoring helps diagnose conditions such as polycythemia, anemia, and cardiovascular diseases, but current measurement techniques remain invasive and cause discomfort and pain.
A novel non-invasive hemoglobin measurement system using multiwavelength PPG and a BPNN machine learning model demonstrated low mean absolute error and was successfully implemented on an embedded microcontroller for real-time execution.
Noninvasive multiwavelength PPG Hb estimation remains investigational; leaves open validation before cardiovascular use.
Continuous hemoglobin (Hb) monitoring can diagnose conditions such as polycythemia, anemia, and various cardiovascular diseases. Current Hb measurement techniques are still invasive, often causing discomfort and pain for the user. Non-invasive Hb measurement is a newer technology that is gaining popularity among researchers. This paper proposes a new non-invasive Hb measurement approach using multiwavelength PPG signals. The transmissive mode PPG sensor and the reflective mode PPG sensor collected real-time PPG signals at four wavelengths: 560 nm, 660 nm, 880 nm, and 940 nm. A novel twenty-six feature vector such as perfusion density (PD), PPG absorbance (ABR), molar extinction coefficient (MEC), magnitude of prominent peak (PPM), frequency of prominent peak (PPF), and absorbance ratio are first time computed in this work from the real-time acquired PPG signal. The performance of the proposed framework is analyzed by utilizing the optimized four different machine learning (ML) regression models such as the AdaBoost model, XGBoost model, BPNN model, and the CatBoost model. Further, this work utilized Shapley Additive Explanations (SHAP) values to visualize the importance of each feature in Hb estimation, enabling a clear understanding of how different features contribute to the prediction process. This SHAP value helps to select the most relevant features for improving the accuracy and interpretability of the Hb estimation model. Results indicate that the BPNN model exhibited superior performance, yielding a mean absolute error (MAE) of 0.421 g/dL, 0.624 g/dL, and 1.57 g/dL for Dataset-1, Dataset-2, and In-house dataset respectively. Additionally, the proposed framework is implemented on an embedded platform featuring an STM32F67ZI microcontroller, showcasing its effectiveness for real-time execution.
No takes yet. Share an insight, caveat, or question.
Pankaj et al. (2025) studied Hemoglobin estimation. Multiwavelength PPG with BPNN model vs. Other ML models (AdaBoost, XGBoost, CatBoost) was evaluated on Mean absolute error (MAE) for hemoglobin estimation. A non-invasive hemoglobin measurement system using multiwavelength PPG and a BPNN model achieved a mean absolute error of 0.421 g/dL, 0.624 g/dL, and 1.57 g/dL across three datasets.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: