# HealthCam: analyzer evaluation and source notes

This note accompanies the HealthCam 3.0 case study. The research results are from the earlier HealthCam 2.0 thesis, not a benchmark of the current platform or all analyzers.

## Scope and attribution

HealthCam is the capture, orchestration, service integration, ingestion, analytics, and product platform built by Diego Turri. Replaceable analyzers are an architectural goal. The heart-rate analyzer integrates pyVHR, a research framework from PHuSe Lab at the University of Milan associated with the thesis supervisor's research. The blink analyzer uses MediaPipe landmarks with custom blink-processing logic. The underlying pyVHR algorithms and MediaPipe models are not Diego's original work.

- pyVHR: https://github.com/phuselab/pyVHR
- Research paper: https://doi.org/10.7717/peerj-cs.929

## Method

The thesis explored whether webcam heart-rate estimates could resolve differences between desktop applications in everyday use. Reference data came from a Polar H10; an Arduino sensor collected ambient brightness. Analysis included timestamp alignment, measurement error, agreement, and lighting comparisons.

## Presentation sources

Source: Diego Turri, “Presentazione Tesi 7.pptx”, HealthCam 2.0 thesis presentation, supplied by the author.

- Slide 8: historical workstation photo. The website uses the existing image extracted from the corresponding thesis presentation PDF; this is not a screenshot of the current 3.0 interface.
- Slide 9: reference correlations reported as 0.09 (MJPG) and −0.01 (lossless).
- Slides 10–11: about 11.6 BPM mean absolute error versus about 3 BPM differences between applications. These magnitudes are visualized on the page; the graphic is not a significance test or a new experiment.
- Slide 15: raw data, GitHub Actions, and derived insights. Current scheduling was checked against the 3.0 codebase, not inferred from the older slide.

The tested configuration did not reliably distinguish these application-level differences. This limits the evaluated estimator's use in those conditions. It does not establish that all pyVHR methods fail in real-world use or that the HealthCam software pipeline failed. Conversely, a working software pipeline does not validate physiological measurement accuracy.

## Reproducibility

The saved H10 notebook is a different run: 12.07 BPM MAE, 14.47 BPM RMSE, and 443 matched observations. These should not be combined with the presentation into one benchmark. The notebook references local CSV datasets absent from the inspected checkout, so the original numerical results were not rerun for this page. The hosted demo uses synthetic telemetry and is not measurement-validation evidence.

## Current data jobs

GitHub Actions schedules incremental summary updates every five minutes and a daily rebuild at 02:10 UTC. The daily window starts two days before the latest summary date. Scheduled execution may be delayed; this is not an exact freshness guarantee. Summary queries reduce repeated scans of raw telemetry, but no measured scalability or latency improvement is claimed.