Recordings and reference blood pressure AI
This project uses the public bed-based ballistocardiography dataset, described by Carlson et al. (2021). The maintained analysis includes 40 participants and the first 100 detected cycles after five seconds of each recording. Excerpts are fixed by this rule, not selected for predictive performance.
Signal preparation
The released processed signals are sampled at 1 kHz. The release applies ECG filtering at 1–40 Hz, BCG at 1–10 Hz, and low-pass filtering below 10 Hz to PPG and pressure. Its pressure-delay correction is already incorporated. Display traces are antialias-resampled to 100 Hz; measurements use the original sampling rate.
The reference is Finometer-reconstructed brachial pressure, supplied as reBAP. Its acquisition voltage is multiplied by 100 to obtain mmHg. Each retained ECG cycle supplies its pressure maximum and minimum; see systolic and diastolic pressure. This is a reconstructed reference, not direct arterial-line pressure.
Quality and missing data
Reference cycles must be finite, within 20–300 mmHg, have pressure range at least 5 mmHg, and duration 0.35–1.8 seconds. Feature validity is assessed separately. Missing landmarks retain their original cycle positions rather than joining nonadjacent beats.
The matched prediction comparison additionally requires valid original Film0 features and both pressure targets. All 40 people contribute, with 16–26 valid test windows each, totaling 1,030. The windows overlap in time; they are not independent participants. See calibration and evaluation and validation context.