Upload an ATR-FTIR infrared spectrum from a saliva sample and get an instant machine-learning screening estimate for Type 2 diabetes risk.
Accepted formats: CSV (wavenumber, absorbance columns) or a JSON array of absorbance values across 399–4000 cm⁻¹.
The neural network outputs a single number between 0 and 1 for every spectrum — called the model score. It is produced by a sigmoid activation on the final layer, so it always falls in this range. A score of 1 means the spectrum looks maximally similar to diabetic samples in the training data; 0 means it looks maximally non-diabetic.
A fixed threshold (set to 0.60 based on the team's model evaluation) acts as the classification boundary. Scores at or above the threshold are labelled elevated risk; scores below are labelled low risk.
The confidence percentage measures how far the score is from that boundary, rescaled to 0–100 %. A score right at the threshold gives 0 % confidence (maximally uncertain). A score of 1.0 (elevated) or 0.0 (low) gives 100 % confidence. This is not a calibrated medical probability — it is a measure of how decisive the model's output is.
From saliva sample to screening result in four steps.
A small saliva sample is collected after fasting.
The sample is scanned with infrared spectroscopy.
Export the spectrum as CSV or JSON and upload it.
The neural network outputs a risk confidence score.