Can a Transformer Predict Blood Pressure Without a Cuff?

A research team submitted a 21-page paper to arXiv on 24 August 2026. The paper proposes a hybrid Transformer framework for cuffless continuous blood pressure prediction.

The framework does not ingest raw waveforms. It models 10-step sequences of six physiological descriptors and two demographic covariates. Those descriptors come from ECG and PPG signals. The covariates are age and sex.

What the system actually ingests

The source pool comprised 28,486 waveform segments from 203 subjects in the MIMIC-III Waveform and Clinical Databases. Feature generation retained 53,621 observations from 166 subjects. The held-out test set contained 2,431 segment-level windows. Each window represents a short temporal slice. Six physiological descriptors capture heartbeat morphology and pulse-wave timing. Two demographic covariates anchor the prediction to the patient. The sequence length is fixed at ten steps.

How the three branches work together

A Multi-Source Temporal Encoder Module runs three parallel branches. A Transformer branch captures temporal dependencies. A Kolmogorov-Arnold Network branch captures nonlinear relationships. An XGBoost branch captures tabular structure. A Dynamic Conditional Fusion-Decoder then merges the branches. It applies differential multi-head attention. It uses token-weighted aggregation. It adds gated residual correction. A robust composite objective jointly optimizes diastolic and systolic targets.

What the numbers show on MIMIC-III

On the 2,431 test windows the mean error for diastolic pressure was 0.41 mmHg with a standard deviation of 3.74 mmHg. The 95 percent limits of agreement spanned -6.93 to 7.74 mmHg. Ninety-eight point forty-eight percent of predictions fell within 10 mmHg. For systolic pressure the mean error was -1.60 mmHg with a standard deviation of 5.95 mmHg. The 95 percent limits of agreement spanned -13.25 to 10.06 mmHg. Ninety-four point thirty-six percent of predictions fell within 10 mmHg.

Where the validation stops

This retrospective analysis is not formal device validation. Subject-disjoint and external evaluation remain necessary before clinical use. The paper does not report performance on data from other hospitals. It does not report performance across different device manufacturers. It does not report performance during motion artifact or arrhythmia. The authors note five figures and six tables in the manuscript. Code and data links appear on the arXiv page. The preprint carries version number v1.