ASMA-18

Inspiration

Two asthma patients can look identical on day one, and ten years later one is fine while the other needs the most intensive treatment. Can we tell which is which from the start? We had a real clinical dataset (~550 patients, ~190 variables) to find out.

What it does

ASMA-18 gives each patient a calibrated risk of progressing to severe asthma, using only baseline factors.

AUC 0.75 (95% CI 0.71–0.81), versus 0.57 for initial severity alone.
Low-risk patients progressed about 1 time in 5. High-risk ones, about 4 times in 5.
When it says 50%, about half really progress.

It supports clinicians. It does not prescribe.

How we built it

Python and scikit-learn: a Random Forest and a Logistic Regression, averaged. Every patient is scored by a model that never saw them (repeated cross-validation on 363 patients, 124 of whom progressed).

Challenges we ran into

Our first model was too good to be true. It scored 0.82. Once we removed variables that secretly describe the patient's current state, it dropped to 0.64. We rebuilt it without them.

The data said steroids make asthma worse. Patients on the strongest drugs had less control. Of course, doctors give those drugs to the hardest cases. With one snapshot per patient, no model can tell which drug works best, so we refused to claim it.

Accomplishments that we're proud of

A model that beats a trivial baseline by 0.18 AUC and whose percentages mean what they say.
We tested a second target, "who stays severe", and it did not beat chance (AUC 0.61, CI 0.47–0.68). So we did not present it.
Side finding: in patients with more severe disease, BMI ≥ 30 goes with worse control (37% vs 52%, p = 0.009).

What we learned

A high score means nothing until you remove leakage and compare it with a simple baseline. And an honest "we can't know" is worth more than a flashy number.

What's next for ASMA-18

Follow-up data (treatment dates, doses, repeated control scores) to study treatments properly, validation in other hospitals, and a simple dashboard for clinicians.