Finding a pulse in the camera's noise

Vastmindz came to It-Jim for computer vision and signal processing at research level, combined with the ability to ship code that runs on a phone. The founder worked directly with our CEO, Ievgen Gorovyi, who holds a PhD and leads an R&D team that works from published research to production code.

The project started with a technical baseline. Within a month the team had tested several face detectors, tracking approaches and channel-mixing methods against a public benchmark with contact-sensor ground truth. The work then ran in sprints, with a written report for every research step: over 200 pages by month 7.

We sorted every failed scan in client iPhone 11 recordings by cause: bright light in front, backlight, artificial light. Those causes became quality metrics. The library measures illumination, sharpness, face size and orientation, and the signal-to-noise ratio of the pulse itself. When the instant SNR drops below about 5 dB, that stretch of the reading is marked unreliable and left out of the result.

Phones deliver frames at uneven intervals, so the library stopped assuming a fixed frame rate. It reads the real timestamp of every frame and computes the vitals from the actual timing.

In side-by-side testing, both Vastmindz and the market-leading rPPG app returned a heart rate for a photo held in front of the camera. We added liveness detection based on eyelid motion: the library detects blinks from the speed of eye-landmark changes and rejects a static photo.

Each real-world problem and its fix:

• Movement and changing light: landmarks weighed against optical flow keep the regions in place; illumination, sharpness, face size and orientation are scored on every scan.
• Uneven frame rate: vitals are computed from the real timestamp of every frame.
• Backlight or a lamp in front: stretches where the pulse SNR drops below about 5 dB are marked unreliable and left out.
• A photo held to the camera: the blink-based liveness check rejects a static photo.

The team collected its own data where public datasets fell short: recordings with a DSLR and an iPhone filmed in parallel, with a finger oximeter and a blood pressure cuff as reference; a controlled desaturation test for SpO2; recordings at rest, after exercise and in different poses.

Before the investor demo in month 9, a 10-day SpO2 sprint ran experiments daily on two datasets: one recorded in near-ideal conditions to validate ideas, one from real users to show results.

We checked the camera against the devices doctors and athletes already trust: a finger pulse oximeter, medical monitors in hospitals, smartwatches and the market-leading camera app. Each test was harder than the one before. The camera kept up.

Vastmindz does not sell to patients directly. It licenses the vitals scan as an SDK: insurers, telemedicine providers and corporate wellness programmes build it into their own apps, so their customers and employees can check their vitals with the phone they already have.

In its Clutch review, Vastmindz says the work let it accelerate and scale: the platform can now handle millions of users, and the company consistently sees accuracy rates of 98% or above on its own metrics.

The technology also moved into clinical and enterprise use. The Canadian public company Ehave licensed it for G20 countries. Leeds Teaching Hospitals NHS Trust began a study of the Vastmindz app on adult and child in-patients, and Vastmindz launched Visix inside Microsoft Teams, where employees check their vitals with the laptop camera in under a minute.