Non-invasive hemoglobin estimation
Loading model...
Analyzing frames…
Overhead room light or indirect daylight from a window. The model was trained under clinical lighting (~4000–5500 K). Avoid flash, direct sunlight, colored LEDs, or dim rooms.
Fill the frame with the fingernail. The nail bed should cover most of the guide box.
The model reads the natural nail bed color. Polish, gel, or acrylic nails will produce wrong results.
The 30-frame capture averages over ~2.5 seconds. Movement causes blur and high variance.
Cold hands have reduced blood flow to the nail bed, which can affect color. Avoid pressing down hard — let the finger rest naturally.
Performance degrades on very dark or very pale skin (undertested in training data). Blue-ish or yellow-ish artificial lighting shifts the critical red-green color axis and causes hallucinated readings.
Frame-to-frame predictions varied too much (high std dev or range). This usually means the fingernail was not clearly visible, lighting changed, or the input is out-of-distribution. Do not trust this result.
Multi-frame median of 30 captures with IQR outlier rejection.
Deployed web model MAE is higher on single frames; multi-frame improves stability. Not a substitute for a blood test.
Gray = invalid frame · Red outline = IQR outlier (excluded from median)