Predicting hypotension from a pulse waveform

An 85-million-parameter transformer was pretrained by masked reconstruction on intraoperative waveforms from VitalDB, then fine-tuned to predict hypotension. Pretraining used two NVIDIA B200 GPUs for 38.2 minutes. On patients held out of training, 30 seconds of photoplethysmography predicted whether mean arterial pressure would fall below 65 mmHg within 5 minutes (test AUROC 0.693), compared with 0.676 for gradient-boosted trees on hand-crafted pulse features and 0.641 for the same transformer trained from scratch.