One Hundred People Like You

Step 1

What is true of you today

Ten things you already know. No laboratory value among them. These feed a risk model fitted on 253,680 responses to the CDC's 2015 behavioural risk survey.

Age

Body mass index

Weight in kg divided by height in metres, squared.

Your general health

Sex

Has a health professional ever told you that you have…

High blood pressure High cholesterol Coronary heart disease or a heart attack

And about daily life

Physical activity in the past 30 days Serious difficulty walking or climbing stairs Smoked at least 100 cigarettes in your life

Step 2

The test, and how good it actually is

No test is perfect, and published estimates disagree. The defaults below are editable on purpose: a tool that hides this behind a single number is hiding the part that matters.

Screening test

Sensitivity

49%

Of people who do have diabetes, the share the test catches.

Specificity

96%

Of people who do not, the share correctly cleared.

Step 3

The result you are holding

Before the test

your risk from the ten answers above

After the test

How much to trust the first number

The whole chain rests on the pre-test probability. If that is wrong by a factor of two, the answer is wrong by roughly the same factor. So the question worth asking of the model is not whether it ranks people well. It is whether, when it says twenty percent, twenty percent of those people turn out to have diabetes.

Measured on 76,104 held-out responses the model never saw during fitting.
Calibration error	0.013	mean gap between predicted and observed, over ten equal-sized bins
Brier skill score	0.173	improvement over always guessing the base rate
ROC AUC	0.819	ranking ability, reported second because it says nothing about calibration
Base rate	13.9%	diabetes prevalence in the survey sample

predicted by the model observed in the data

Each row is one tenth of the held-out sample, ordered by predicted risk. The pale bar is what the model predicted; the solid bar underneath is what actually happened in that tenth. They track closely, which is the property this tool depends on and the reason it is reported before anything else.

What this is not

This is an educational tool for understanding screening arithmetic. It is not a diagnosis, not medical advice, and not a substitute for talking to a clinician about your own result.
The risk model is fitted on United States survey data from 2015, and self-reported. It will fit some populations better than others, and it has not been validated on any Korean cohort.
Published sensitivity and specificity vary between studies and between populations. That is exactly why both are left adjustable rather than baked in.
The arithmetic assumes the test result is independent of the risk factors, given true disease status. Real tests can break that assumption.