Real-time classification of blood pressure changes using photoplethysmography and deep learning

Blood pressure (BP) fluctuations are key indicators of cardiovascular health, yet most non-invasive monitoring approaches focus on estimating absolute BP values. This study introduces a novel classification framework for detecting directional BP changes—systolic (SBP), diastolic (DBP) and mean (MBP)—using continuous photoplethysmography (PPG) signals and a single initial BP calibration. Data from 1,626 patients in the Vital Signs Database (VitalDB) with synchronised PPG and BP recordings were analysed. BP changes were categorised as ‘Spike’ (increase above a threshold), ‘Stable’ (within ± threshold), or ‘Dip’ (decrease below a threshold). Four deep learning models were evaluated: multi-layer perceptron, convolutional neural network, residual network, and Encoder. A subset of 1,000 patients was used for training and validation with balanced label distributions. Two independent test datasets were compiled: Test-I (n=600) with balanced classes and Test-II (n=26) reflecting natural BP fluctuation distributions. The Encoder model, leveraging PPG and second-derivative signals with Softmax-based temporal attention, achieved the highest accuracy: 77.8% ± 2.8% (Test-I) and 80.0% ± 0.8% (Test-II) at thresholds of 30 mmHg (SBP), 15 mmHg (DBP) and 20 mmHg (MBP), with corresponding F1-scores of 78.2% ± 2.8% and 83.1% ± 2.2%. These findings demonstrate the feasibility of real-time, non-invasive BP change detection from PPG, offering a promising alternative to regression-based BP estimation for continuous health monitoring. • Novel framework to classify systolic, diastolic, and mean BP changes from PPG. • Alternative to regression models for real-time clinical alerting. • F1-score > 80% on imbalanced, clinically realistic datasets using optimised Encoder. • Publicly available models, training code, and preprocessing scripts.