GlySight Kharazmi University · Explainable Machine Learning

Early Type 2 Diabetes Risk Screening — Live Demonstrator

An interactive companion to our study on 253,680 CDC BRFSS 2015 survey responses. The model runs entirely inside your browser: nothing you enter is transmitted or stored.

This is not a medical device and not a diagnostic system. It is an academic research prototype for early risk screening. A high score means only that the entered characteristics resemble those of survey respondents who reported diabetes or prediabetes — it is not a diagnosis. It has undergone no clinical validation. Any health concern must be evaluated by a qualified healthcare professional.

Respondent profile

All 21 model inputs. The result updates as you change any value.
Estimated probability
—%
—
0%threshold100%

Why this score?

Effect of each entered value versus the typical value, in log-odds. Red raises the estimate, teal lowers it.
How to read this These are counterfactual contributions — the change in model output when a single value is replaced by the population median, everything else held fixed. They describe the model, not the human body, and they are not TreeSHAP values. Exact SHAP analysis is reported in the written study.

Decision threshold

The model outputs a probability; the threshold turns it into a decision. 0.50 is only a convention.
0.115
—Recall
—Precision
—Specificity

Model performance

Measured once on a held-out test set of 50,736 records at threshold 0.115.
MetricValue
Recall (sensitivity)0.8383
Specificity0.6616
Precision0.2863
F1-score0.4268
ROC-AUC0.8276
PR-AUC0.4236
Context for these numbers Precision is low by design. An alert here means “worth a confirmatory blood test”, not “has diabetes”. The threshold was deliberately lowered to reduce missed cases, which necessarily increases false alarms.