Non-Invasive Anemia Detection: A Multichannel PPG-Based Hemoglobin Estimation with Explainable Artificial Intelligence

Anemia is a prevalent hematological disorder that requires frequent hemoglobin monitoring for early diagnosis and effective management. Conventional hemoglobin assessment relies on invasive blood sampling, limiting its suitability for large-scale or continuous screening. This paper presents a non-invasive framework for hemoglobin estimation and anemia screening using multichannel photoplethysmography (PPG) signals and explainable artificial intelligence. Four-wavelength PPG signals (660, 730, 850, and 940nm) are processed to extract optical and cross-wavelength features, which are aggregated at the subject level to avoid data leakage. A gradient boosting regression model is employed to estimate hemoglobin concentration, followed by post-regression anemia screening using World Health Organization (WHO) thresholds. Model interpretability is achieved using SHapley Additive explanations (SHAP), enabling both global and subject-specific analysis of feature contributions. Experimental evaluation on a publicly available dataset demonstrates a mean absolute error of 8.50 plus minus 1.27 and a root mean squared error of 8.21g/L on unseen test subjects, indicating the potential of the proposed approach for interpretable, non-invasive hemoglobin monitoring and preliminary anemia screening.

💡 Research Summary

The paper addresses the pressing need for a non‑invasive, scalable solution to monitor hemoglobin (Hb) levels and screen for anemia, a condition affecting over two billion people worldwide. Conventional Hb measurement relies on venous blood draws, which are uncomfortable, carry infection risk, and require laboratory infrastructure, making frequent or large‑scale screening impractical. Leveraging the wavelength‑dependent absorption properties of hemoglobin, the authors propose a comprehensive framework that combines multichannel photoplethysmography (PPG), machine‑learning regression, and explainable artificial intelligence (XAI).

Four wavelengths—660 nm, 730 nm, 850 nm, and 940 nm—were selected to capture both visible and near‑infrared absorption peaks of oxy‑ and deoxy‑hemoglobin. Raw PPG recordings from a publicly available dataset containing 152 adult subjects were segmented into 500‑sample windows. Signal quality was quantified using signal‑to‑noise ratio (SNR) and a Welch‑based signal quality index (SQI) before and after applying a third‑order Butterworth band‑pass filter (0.5–5 Hz) to preserve the physiological cardiac band while suppressing baseline drift and high‑frequency noise.

From each filtered segment, a rich set of features was extracted: time‑domain statistics (mean, standard deviation, RMS, peak‑to‑peak amplitude), pulsatile components (AC, DC, AC/DC ratio), logarithmic attenuation, and frequency‑domain descriptors (dominant frequency, band power within the heart‑rate range, spectral entropy). Crucially, cross‑wavelength ratios (e.g., 660 nm/940 nm) and attenuation ratios were computed to encode relative optical absorption differences that are directly linked to Hb concentration. After cleaning the feature matrix (removing NaN‑heavy and constant columns), the segment‑level features were aggregated at the subject level using robust statistics (mean and median), thereby eliminating label duplication and aligning the input data with the single Hb reference value per subject.

The regression problem was explored with several tree‑based ensembles: Random Forest, CatBoost, XGBoost, and LightGBM. An 80:20 subject‑wise split ensured that no individual appeared in both training and test sets, preventing data leakage. LightGBM emerged as the best performer, achieving a mean absolute error (MAE) of 8.50 ± 1.27 g/L and a root mean squared error (RMSE) of 8.21 g/L on the held‑out test subjects—error margins that are clinically acceptable for preliminary screening.

To make the model transparent, SHapley Additive exPlanations (SHAP) were applied. Global SHAP summary plots identified AC/DC ratio, specific cross‑wavelength ratios, and spectral entropy as the most influential features. Local SHAP waterfall charts for individual subjects revealed how particular features (e.g., a high peak‑to‑peak amplitude at 850 nm) pushed the prediction upward or downward, offering clinicians intuitive insight into each decision. Dependence plots further illustrated non‑linear relationships, such as a threshold effect where AC/DC ratios above a certain value sharply increase predicted Hb.

Anemia screening was performed as a post‑regression step using World Health Organization (WHO) thresholds: Hb < 130 g/L for males and Hb < 120 g/L for females. No separate classification model was trained, and anemia labels were not required during model training, avoiding circular inference. Although ground‑truth anemia annotations were unavailable, the authors examined the robustness of the screening by varying the thresholds and observing the resulting classification stability.

The study’s contributions are threefold: (1) a rigorously processed multichannel PPG pipeline that mitigates motion artifacts and tissue scattering; (2) a subject‑level feature aggregation strategy that aligns machine‑learning inputs with clinical measurements and prevents overfitting; (3) an integrated XAI layer that delivers both global and patient‑specific interpretability, fostering trust and facilitating future sensor design refinements. Limitations include the modest sample size, lack of external validation, and the absence of diverse skin tones or activity conditions that could affect PPG quality in real‑world wearable scenarios. Future work should expand the dataset across demographics, incorporate real‑time wearable implementations, and explore hybrid models that combine PPG with demographic or contextual data to further improve accuracy and generalizability. Overall, the paper demonstrates that multichannel PPG coupled with explainable gradient‑boosted regression can provide a viable, low‑cost pathway toward continuous, non‑invasive hemoglobin monitoring and large‑scale anemia screening.