Transformer-Based Deep Learning for Cardiovascular Disease Classification Using Raw Photoplethysmography Signals

Abstract

Cardiovascular disease remains the foremost cause of death worldwide, and elevated blood pressure is its largest modifiable contributor, yet many hypertensive adults remain undiagnosed because sphygmomanometry requires clinical contact. Photoplethysmography offers an inexpensive optical alternative already embedded in wearables, but published classifiers depend on handcrafted descriptors or time-frequency images, and most are validated with splitting schemes that allow recordings from one participant to appear in both training and test partitions. This work introduces CPA-Former, a cardiac-phase-aware Transformer consuming the raw pulse together with its first and second derivatives, combining a multi-derivative tokenizer, a positional encoding defined on cardiac phase rather than absolute time, cross-derivative co-attention, and segment-level attention pooling. It is evaluated on the public PPG-BP database of 219 participants under strict subject-independent cross-validation, alongside four deep baselines and classical references trained identically. The proposed model attains 45.60% accuracy on four-stage staging and 71.66% on binary screening. The central finding is negative and arguably more useful than a headline number: architectures spanning two orders of magnitude in parameter count prove statistically indistinguishable on this cohort, and none convincingly exceeds handcrafted morphology features. Repeating the identical experiment with folds formed over recordings rather than participants inflates accuracy substantially, offering a parsimonious explanation for the far higher figures reported elsewhere on this dataset.