Validation Results | ViFi

4.15 bpm cross-session HR mean absolute error against a Polar H10 chest strap, on $44 of commodity ESP32-S3 hardware, single subject, single room, 4 paired captures totaling ~8 minutes of real-hardware data.

The model never saw the held-out session during training. Both cross-session evaluations are on completely independent recordings.

Hardware

Setup

• Subject seated in a chair with the TX board behind and the RX board in front (~1.5–2 m apart)
• Both antennas vertical at chest height
• Chest centered on the line between the two antennas
• No motion; normal breathing; ~2 minutes per capture
• Polar Beat phone app closed during capture (BLE single-connection)
• All captures in the same room, same time-of-day

Logging

CSI: a custom serial reader records raw bytes from the RX board's UART for exactly 120 seconds, alongside a metadata sidecar containing the measured packet rate.

HR: the Polar H10 streams 1 Hz HR notifications over BLE; auto-reconnects on Windows BLE drops (which happen every ~30 seconds and are handled transparently).

Both run as PowerShell background jobs in parallel; total wall-clock per capture: ~2.5 minutes.

Pipeline (per 10-second window)

1. Subcarrier selection. Compute variance across 192 subcarriers in the 0.1–3 Hz physiological band. Select top 8 by variance.
2. 1-D envelope. Standardize each selected subcarrier (zero mean, unit variance), then average across them.
3. Bandpass filter. Butterworth, 0.1–3 Hz.
4. FFT. 4× zero-padded, parabolic peak interpolation in respiratory (0.15–0.6 Hz) and cardiac (0.9–1.8 Hz) bands.
5. Feature extraction. 9-dim vector of peak frequencies, peak ratios, envelope statistics, and total band energy.
6. XGBoost regression. Two regressors (HR, RR), trained on stacked feature matrix.

Sample-rate handling

When ESP-IDF doesn't emit per-line timestamps (the default configuration), the parser synthesizes a uniform timestamp grid. Default 100 Hz was wrong — the watchdog throttles real packet rate to ~67–77 Hz. The capture tool writes a metadata sidecar with the measured rate; the parser auto-loads the sidecar to use the correct synthesized fs. This single correction is the difference between MAE 24 bpm (wrong assumed rate) and MAE 4 bpm (correct rate, after retraining).

Cross-session evaluation

The retraining tool accepts multiple paired captures and trains an XGBoost regressor on the union of all of them. To produce an honest generalization estimate, one session is held out from training, the model is retrained on the remaining sessions, then evaluated on the holdout.

The XGBoost regressor never sees the held-out session during training. Predictions on the holdout are pure generalization, not memorization.

These are the boundaries of the current dataset and the milestones queued up to extend each one. The pipeline and tooling are designed to scale into all of them — the work ahead is data collection and validation.

• Multi-subject coverage. Captures so far are from the founder. Multi-subject validation across 10+ participants is the next milestone — see the roadmap.
• Multi-room coverage. All captures in one room. Expanding to additional rooms is on the immediate roadmap.
• Subjects in motion. Validated for stationary subjects today. Walking, fidgeting, and repositioning are next on the test plan.
• Heart-rate range. Subject HR was 80–100 bpm. Resting (~60), elevated (~130), and arrhythmia ranges are upcoming test conditions.
• Cross-subject held-out validation. Sessions are independent recordings of the same subject; cross-subject holdout arrives with the multi-subject dataset.
• Regulatory grade. This is a research-grade result. The path to FDA 510(k) Class II clearance is mapped on the roadmap.

See the roadmap for the full sequence.