Machine learning to predict progression in adolescent idiopathic scoliosis — the aim is not diagnosis but optimizing how often X-rays are taken (NCT07556042)

1Not knowing where it goes distorts the treatment

The difficulty in scoliosis lies less in the curve than in the inability to see where it is heading. The record lists what follows from that: bracing for 18 hours a day over years, compliance undermined by appearance, function and skin irritation, and split judgments in cases sitting on the border of surgical indication.

Because progression cannot be read, both patients who wear a brace they did not need and patients whose intervention comes late are produced by the same uncertainty. The value of a correct prediction here is not that it widens the options but that it removes treatment nobody needed.

2The target is imaging frequency, not diagnosis

What this study asks of AI is not to find disease in an image. It is to predict progression so that follow-up radiography can be reduced. The record notes that repeated imaging for monitoring raises lifetime cancer risk through cumulative ionizing radiation, and that the risk of breast cancer in girls with AIS is reported to be roughly seven times that of the healthy population.

Stretch the intervals too far, though, and the window for early intervention is missed. The idea is to ease a bind in which imaging more moves away from safety while imaging less moves toward missed deterioration.

3Trials of 100 or fewer make up a third of the field

Planned enrollment is 30. Across the 750 medical-AI clinical trials this site holds as of 2026-08-28, 254 — 34 percent — plan 100 participants or fewer, against a median of 200. Small studies do not generalize well, but the design itself, building a model retrospectively and validating it prospectively, follows a standard sequence in medical-AI development.

Clinical evaluation of medical AI has a thick layer of small exploratory work, and this study sits in that layer. The performance of the model, and whether the predictions hold, are not part of this registry record.

4The more you image, the further from safety

Follow-up in scoliosis carries a bind. What the study asks of AI is not diagnosis but loosening that bind.

Because progression cannot be read, both patients who keep wearing an unnecessary brace and patients whose intervention comes late are produced. The value of an accurate prediction here is not adding treatment options but removing unnecessary treatment and unnecessary imaging. Thirty participants are planned, with a model built retrospectively and validated prospectively.