Conditional Diffusion–Augmented Cascaded Multi-View Attention BiLSTM for Non-Invasive Blood Glucose Estimation from Photoplethysmography

Photoplethysmography (PPG)-based blood glucose estimation is attractive for non-invasive monitoring, but its development is constrained by limited paired PPG–glucose data and the weak representation of glucose-related waveform variations. This study proposes a framework combining conditional diffusion-based data augmentation with a cascaded multi-view attention bidirectional long short-term memory (BiLSTM) network. Blood glucose level, heart rate, and body mass index were used as physiological conditions to generate synthetic PPG segments, while channel and temporal attention together with cascaded BiLSTM layers were used to extract discriminative spatiotemporal features. A moth–flame optimization algorithm was employed to tune key hyperparameters. Experiments were conducted on a public dataset containing 67 PPG recordings from 23 participants. The framework was evaluated using subject-wise five-fold cross-validation, ensuring that all recordings from the same participant remained within a single fold. In the primary seed-42 analysis, the proposed method achieved a recording-level RMSE of 0.59 mmol/L, an MAE of 0.38 mmol/L, a MARD of 5.48%, and Clarke Zone A and A + B proportions of 97.0% and 100%, respectively; repeating the complete evaluation with seeds 2026 and 3407 produced closely similar recording-level results. These results suggest that physiologically conditioned generative augmentation may improve PPG-based blood glucose estimation in small-sample settings.