40 participants from the public bed-based ballistocardiography dataset. I use short excerpts of 100 detected heartbeats per person.
A small research demonstration
Estimating blood pressure from heartbeat waveforms
Can changes in ECG, finger pulse and bed-sensor waveforms track a person's blood pressure after an initial period of measurements?
After learning from a person's early waveforms and reference pressure, does a model improve on simply repeating their early average?
A personally calibrated, same-recording pilot. It does not test a device on a new person without calibration or establish clinical accuracy.
What did the sensors predict?
A personally calibrated regression using 42 waveform measurements achieved mean participant absolute errors of 3.84 mmHg for systolic pressure and 2.56 mmHg for diastolic pressure. Sharing a common slope pattern across the cohort while allowing personal deviations gave errors of 3.53 / 2.41 mmHg (systolic / diastolic). Each person's error is averaged first, then the 40 participants are weighted equally. These observed average gains remain uncertain in this small cohort.
The model learns from the first 70 heartbeats and predicts later five-beat averages ending at cycles 75–100. It improves on repeating each person's early reference-pressure average and on a four-feature BCG regression evaluated on the same windows. Some participants still have large errors, particularly during changes outside their calibration range.
Explore prediction accuracy and individual pressure traces →
From heartbeat to pressure estimate
ECG records cardiac electrical activity, PPG measures the finger's optical pulse, and bed-mounted BCG sensors record small movements associated with the heartbeat. Timing, amplitude and shape measurements summarize these signals. Personal ridge and a model with shared cohort slopes estimate the maximum (systolic) and minimum (diastolic) pressure within a heartbeat, retaining personal calibration in both cases.
Synchronized signalsECG, PPG and two BCG channels
→42 waveform featuresTiming, magnitude and shape
→Personal calibrationEarly features paired with reference BP
→Later predictionsHeld-out reference BP scores errors
Research context
This demonstration builds on Carlson et al.'s public bed-based BCG study, extending its four-feature regression to multiple waveform channels. The reference is Finometer-reconstructed brachial pressure. Results concern brief later excerpts from the same recordings used for calibration; they do not establish performance on new people without calibration or during long-term monitoring.