FedHemo: A Federated CTAB-GAN+ Framework for Privacy-Preserving Non-Invasive Hemoglobin Prediction using Smartphone-based PPG Signals

Abstract

Non-invasive hemoglobin (Hb) prediction from smartphone-captured fingertip photoplethysmography (PPG) signals possesses transformative promise for point-of-care anemia screening. However, the fairness and generalizability of regression models trained for the task can be severely limited by small, class-imbalanced clinical datasets and demographic representational bias. Our previous research indicated that vanilla GANs amplified bias across most regression models, whereas conditional GANs (cGAN) improved class balance but ignored demographic factors and privacy risks. We introduce FedHemo, a federated CTAB-GAN+ framework that tackles these three challenges in a single pipeline. Jensen-Shannon Divergence (JSD) is a widely used statistical measure for quantifying the disparity between probability distributions. The lower JSD values indicate that the synthetic model generates more representative samples. Pearson Pairwise Correlation (PPC) was utilized to assess feature dependency. Subsequently, our proposed system’s higher PPC values reflect stronger preservation of feature relationships, ensuring that FedHemo accurately captures the complex physiological inter-dependencies inherent in PPG signals. Furthermore, FedHemo consistently surpassed vanilla GAN and cGAN across every fidelity metric. It most notably reached a mean JSD of 0.267, which is more than doubling the distributional accuracy of the cGAN (0.682), while maintaining a superior PPC of 0.654. Downstream regression results show FedHemo-augmented training achieves a best MAE of 0.912 g/dL with SVR (RBF), an improvement over the 0.92 g/dL baseline from our prior work on the same model. The privacy of the training data is further confirmed by a Distance to Closest Record (DCR) of 0.193. We also provide a three-panel SHapley Additive exPlanations (SHAP)-driven visualization and analytics dashboard to facilitate health data analysis.