Cardiovascular sensing · physiology models · forecasting

Multimodal digital patient research

We compare several ways of measuring cardiovascular physiology with computer models and heart-rate forecasts. The research used real recordings; the interactive signals below are synthetic, and real-data results appear separately.

Static bundle · no model runtimeSYNTHETIC DATAOffline precomputed outputs · synthetic inputResearch use · not a clinical tool

The research used real physiological measurements and modelling experiments. The interactive signals on this public page are synthetic demonstration data, created so the work can be explored without publishing the underlying participant recordings. Results from the real-data experiments are shown separately as aggregate findings.

Synthetic exercise-recovery demonstration. These generated traces are not participant recordings. Generic RGB modulation exercises the rPPG processing code, but is not a validated skin, camera or optical model. CHROM, POS and luminance estimates are the research implementations running on synthetic inputs.

Synthetic RR-derived heart-rate reference

Synthetic RR-derived estimate

A rate derived from the synthetic ECG-like beat times, used as the reference series in the forecast comparison.

Synthetic RGB channel means (display model)

RedGreenBlue

Generated values from a generic channel-modulation display model; not a camera or skin model.

Stored rate estimates

CHROMPOSY

Estimator outputs are separate from the synthetic reference.

Band-limited acoustic envelope · 20–400 Hz

Channel 1Channel 2

One-second RMS features describe acoustic energy. Cross-device timing is approximate; ECG beat synchrony is unavailable.

Acoustic spectral-rate features

Blind dominantSecond-ranked blindReference-guided diagnostic

Blind peaks are spectral components. The reference-guided candidate is diagnostic and reference-informed; none identifies S1/S2 or valve events.

How to read these sensor examples

No single sensor is treated as the patient. Each observes physiology through a different measurement process, so comparing them helps separate cardiovascular changes from effects introduced by the sensor.

Electrical activity · ECG

The real study used a Polar H10 chest strap as an independent ECG-derived heart-rate reference and an AD8232 ECG board connected to an ESP32 as a second electrical measurement. These help establish timing and heart-rate agreement; they do not describe a complete cardiovascular state. The trace above is synthetic and RR-derived.

Polar H10 chest strap laid flat, with the sensor module at its centre — Polar H10 chest strap used as an ECG-derived heart-rate reference.

Contact optical pulse · PPG

MAX30102 sensors measured changes in red and infrared light at the skin. Contact PPG provides an optical view of the pulse and helps distinguish limitations in camera sensing from changes in the underlying cardiovascular signal.

ESP32 breakout board beside a small MAX30102 sensor board and jumper wires — ESP32 breakout and MAX30102 sensor board for contact optical pulse sensing.

Remote camera pulse · rPPG

An iPad camera recorded facial red, green and blue intensity. CHROM, POS and luminance-based methods looked for cardiac-rate information in those colour signals. Cardiac-rate-consistent information can remain present even when the strongest spectral component is not cardiac, making it hard to identify blindly.

Cardiovascular sound · PCG-like audio

A stethoscope coupled to a lavalier microphone recorded acoustic energy. The study examined spectral and envelope features, while avoiding claims about individual valve events without adequate synchronisation and validation.

Purple stethoscope with its chestpiece connected to a lavalier microphone — Stethoscope and lavalier microphone used to record cardiovascular sound.

Electrical activity → ECG·Optical pulse change → PPG·Camera colour variation → rPPG·Acoustic energy → PCG-like audio

Definition of terms

Short explanations of the technical language used in this tab.

Synthetic / generated trace: An invented signal created for this demonstration. It is not a recording from a person.
Exercise recovery episode: The period after activity ends, when heart rate moves back toward its resting level.
ECG / RR interval: An electrocardiogram records cardiac electrical activity. An RR interval is the time between successive detected heartbeats in that signal.
RR-derived heart rate / bpm: Heart rate calculated from beat-to-beat intervals; bpm means beats per minute.
PPG / contact PPG: Photoplethysmography measures changes in light at the skin that are related to the pulse. Contact PPG uses a sensor touching the skin.
RGB / rPPG: RGB means red, green and blue camera channels. Remote PPG estimates pulse-related changes from their variation in video without touching the skin.
CHROM, POS and Y: Three ways to turn camera colour variation into candidate pulse estimates. CHROM and POS combine colour channels; Y uses brightness. Their outputs are estimates, not direct measurements of heart rate.
Reference / estimator: A reference is the series used for comparison. An estimator is software that produces a candidate value from a signal.
Blind / reference-guided: Blind analysis selects a signal component without using the ECG-derived rate. Reference-guided analysis uses that rate to inspect a nearby component and is diagnostic, not an independent estimate.
Spectral component / rate: A repeating pattern found by analysing a signal by frequency. Its rate need not equal the heartbeat rate.
Acoustic envelope / RMS: An envelope tracks the strength of recorded sound over time. Root mean square (RMS) is a summary of signal energy within a time window.
20–400 Hz / events per minute: Hertz (Hz) counts sound-wave cycles per second; this range is the analysed audio band. Events per minute reports the rate of a repeating acoustic pattern.
PCG-like audio / S1 and S2: Phonocardiogram (PCG) refers to recorded heart sounds. S1 and S2 are the usual first and second heart sounds; the plotted features do not identify them.
Device synchronisation: Aligning recordings from different devices to the same time. Approximate timing limits beat-by-beat comparisons.
Polar H10, AD8232, MAX30102 and ESP32: The ECG chest strap, ECG board, optical pulse sensor and microcontroller used in the research setup. Their recordings are not published here.

Separate model episode. Pulse and OpenBF outputs come from a simulated cardiac-cycle export. Its phase axis is independent of the synthetic recovery demonstration.

Flow waveforms

Pulse · aortic inflowOpenBF · external carotid flow

Pulse aortic inflow and OpenBF external carotid flow, shown in native flow units.

Pressure waveforms

Pulse · aortic pressureOpenBF · external carotid pressure

Pulse aortic and OpenBF external carotid pressures, shown in mmHg.

Pulse heart-rate state

Pulse HR

The model-cycle phase is independent of the synthetic recovery demonstration.

Optical observation boundary

A validated vascular-to-cutaneous observation model is not included in this episode. The research record includes conditional contact-reflectance and visible-camera control experiments using generic phantoms and declared transfer assumptions; these do not establish participant-specific facial rPPG. No synthetic camera trace is shown here.

Pulse state → OpenBF vascular outputs → cutaneous transfer not established → no synthetic rPPG claim

What the model episode represents

Pulse · whole-body physiology

Pulse represents whole-body physiological state and transitions. Here it supplies an inspectable mechanistic projection under declared assumptions, rather than acting as a black-box predictor.

OpenBF · arterial flow

OpenBF models one-dimensional arterial blood flow. Coupled downstream of Pulse, it provides higher-resolution vascular pressure and flow waveforms.

Where the modelling stops

This work does not include a validated link from an OpenBF carotid waveform through facial microcirculation and tissue optics to camera RGB. The model episode therefore makes no claim to generate synthetic rPPG.

Definition of terms

Short explanations of the technical language used in this tab.

Model episode / simulated cardiac cycle: One saved computer-model run showing a heartbeat cycle. Its time axis is separate from the synthetic recovery episode.
Pulse Physiology Engine: Software that simulates whole-body physiological states and their changes under specified assumptions.
OpenBF: Software that simulates pressure and blood flow along an arterial network.
Coupling / model output: Coupling passes a declared output from one model into another. A model output is a simulated value, not a patient measurement.
Aorta / aortic inflow: The aorta is the large artery leaving the heart. Aortic inflow is the blood entering it per unit of time in the model.
External carotid artery: A branch of the carotid circulation supplying parts of the head and face. Its modelled flow is an upstream vascular reference, not a skin measurement.
Flow waveform / mL/s: How simulated blood-flow rate changes over a heartbeat; mL/s means millilitres per second.
Pressure waveform / mmHg: How simulated arterial pressure changes over a heartbeat; mmHg means millimetres of mercury.
Heart rate / bpm: The number of heartbeats per minute represented by the model.
Cycle phase: Time within a single simulated heartbeat, rather than time in the separate recovery recording.
Mechanistic projection: A model calculation of how its represented physiology evolves from specified conditions. It is not automatically an individual prediction.
Observation model / vascular-to-cutaneous transfer: A proposed link from vascular activity to a sensor reading at the skin. That link has not been validated for facial camera signals here.
Phantom / reflectance: A phantom is a simplified test material or computer representation. Reflectance is the light returned from it. Such controls do not reproduce a participant’s face.
Camera RGB / rPPG: Red, green and blue camera measurements and a pulse estimate derived from them. This model episode does not generate a validated camera signal.

30 s synthetic context → held-out synthetic future. Forecasts use only generated values through 30 s. Later synthetic values are shown for retrospective scoring and were not passed to TimesFM or Pulse weighting.

Synthetic context, held-out values, and forecasts

Synthetic contextHeld-out syntheticTimesFM-3 forecastPulse scenario projectionPersistence baselineLinear-trend baseline

Pulse is an HR-constrained projection from a frozen 72-candidate exercise–recovery family, not an individual state estimate. TimesFM-3 was run offline on the synthetic RR-derived rate series.

Forecast evaluation

Mechanistic context compatibility

Frozen Pulse 4.3.2 StandardMale baseline family: exercise intensity fractions 0.05–0.30, durations 30/45/60 s, cessation-to-origin delays 0/5/10/15 s, equal prior, no extra recovery parameters. Candidate weights use six synthetic context points at 5 s intervals and a 10 bpm model-discrepancy scale. Held-out values are excluded from conditioning. The ensemble interval is descriptive, not calibrated.

Forecasting data versus projecting physiology

TimesFM-3 receives a heart-rate history and forecasts that observed data series. Pulse projects how a represented cardiovascular system evolves under stated assumptions and interventions. Those are different kinds of output.

This chart uses synthetic values: only the first 30 seconds condition the forecasts, and later synthetic points are used for scoring. In the separate real-data recovery experiment, both approaches received the first 30 seconds of H10 heart-rate information and were compared with the following 30 seconds and with persistence and linear-trend baselines. Under the frozen Pulse exercise-recovery assumptions, the candidate family did not adequately represent the observed recovery trajectories. It was left unchanged rather than tuned against the held-out future.

Definition of terms

Short explanations of the technical language used in this tab.

Context / held-out future: Context is the first 30 seconds supplied to a forecaster. Held-out values are the later points kept aside until scoring.
Synthetic observation / RR-derived rate: An invented example series. Its heart rate is calculated from generated beat intervals rather than recorded from a person.
Forecast / projection: A forecast extends a measured data series. A mechanistic projection extends a simulated physiological state under stated assumptions.
TimesFM-3: A machine-learning model that predicts future values of a time series from its earlier values.
Pulse scenario projection: A future heart-rate trajectory produced by weighting a fixed family of Pulse simulations using the early context. It is not a personalised state estimate.
Persistence / linear-trend baseline: Simple comparison forecasts: persistence repeats the last known value; linear trend extends a recent straight-line fit.
Candidate / ensemble / prior: A candidate is one simulated trajectory. An ensemble is the set of candidates. An equal prior gives each candidate the same starting weight before context is considered.
Conditioning / candidate weight: Using early observations to give more weight to simulations that agree better with that early context. Later points are excluded.
Model discrepancy scale / 10 bpm: A chosen tolerance for disagreement between a candidate and the context, measured in beats per minute. It affects the weights.
Effective ensemble size: A summary of how many candidates contribute materially after weighting. It can be much smaller than the number generated.
MAE / RMSE: Mean absolute error averages the size of forecast errors. Root mean square error also averages error but gives larger misses more weight. Lower is better for both.
Context RMSE: Root mean square error measured against the early points used for conditioning, before the future forecast is judged.
Forecast interval / calibration: A range shown around a forecast. Calibration would require checking whether such ranges contain future observations at their stated frequency; the Pulse interval here is descriptive, not calibrated.
Frozen family / retrospective scoring: The 72 simulations and rules were fixed before scoring the later points. Retrospective scoring compares a forecast with observations already collected, while withholding those future values from the forecast calculation.
Exercise intensity / cessation-to-origin delay: Settings in the Pulse simulations: the requested exercise effort and the time between stopping exercise and the start of the displayed recovery period.

Research results · aggregate findings.

Real-data held-out forecast error by recovery (MAE, bpm)

Recovery	TimesFM-3	Pulse projection	Persistence	Linear trend

These aggregate findings come from the real-data experiments. They are separate from the synthetic traces and forecast example in the other tabs; those did not produce these research results.

Definition of terms

Short explanations of the technical language used in this tab.

Aggregate finding / real data: A summary across observations from the real research recordings. This tab does not expose the underlying recordings or participant-level traces.
Recovery episode: One measured period after exercise. The listed recoveries are repeated episodes, not independent participants.
Held-out forecast error: The difference between forecast values and later observations that were kept out of the forecasting step.
MAE / bpm: Mean absolute error is the average size of those differences; bpm means beats per minute. Lower MAE means closer predictions in these episodes.
TimesFM-3: A machine-learning model forecasting the future of the measured heart-rate series.
Pulse projection: A forecast from a fixed family of simulated exercise-recovery trajectories, weighted using only the earlier heart-rate context.
Persistence / linear trend: Simple comparison forecasts: carrying the last value forward or extending a recent straight-line trend.
Synthetic demonstration: The generated example in the other tabs. Its plotted signals and forecast do not produce the real-data results in this table.

Research repository and demo architecture

Research record · htn-coupling

The research repository contains protocols, frozen assumptions, model configuration, Pulse/OpenBF coupling code, analysis scripts, tests and aggregate research results.

Public artefact · daviddasa.com

This site contains a static interactive demonstration and synthetic data bundle. It does not publish private physiological recordings or run the physiology and machine-learning models in your browser.

What the browser does

Pulse represents whole-body physiology; OpenBF models arterial flow; TimesFM-3 forecasts a measured time series. The browser charts present precomputed signals and outputs. The renderer displays scientific data; it does not generate the physiology.