Measuring port catheter tip position with AI — a prospective study that locks the model first and adds no imaging — a clinical trial (ClinicalTrials.gov)

1A few centimeters that matter

A central venous port is implanted under the skin for chemotherapy or long-term infusion, and where the tip of its thin catheter comes to rest bears directly on safety. Too shallow invites thrombosis or extravasation of the drug; too deep raises arrhythmia and cardiac complications.

In practice the position is checked against the junction of the vena cava and the right atrium, but reading that position off a chest radiograph takes measurement work and varies between readers. This study asks how accurate an AI that automates the measurement actually is, using real cases.

2Lock it first, then measure

1. 1DevelopBuild the AI method that estimates tip position from images
2. 2Test internallyCheck its behavior on data already in hand
3. 3LockFreeze the model before prospective enrollment opens
4. 4Enroll prospectivelyRegister eligible patients within routine care
5. 5CompareMeasure mean absolute error against a prespecified manual method

That order is not a formality. Left unlocked while new cases arrive, a model gets tuned toward the examples it handles badly, and the result is a score that only holds on the data it was tuned to. Freezing before prospective enrollment, with the comparison method fixed in advance, removes the room to adjust conditions after the fact.

The practice is not unusual in AI performance evaluation, but writing it explicitly into the design says something about the quality of this study.

3No extra burden added for research

The other commitment is that no imaging is performed for research purposes. Reference measurements are taken from chest CT scans that were clinically indicated anyway, in cases where both the catheter tip and the cavoatrial junction are visible. That is a decision not to add a radiation-bearing examination on top of routine care, and it lets the evaluation proceed without increasing the burden on participants.

The record also states that the automated output will not guide catheter placement or change clinical decisions.

4Error itself as the primary measure

The primary measure is mean absolute error — on average, how far the automated estimate sits from the actual position. Not a binary right or wrong, but the size of the miss taken directly as the metric. Position is a domain where a difference of a few centimeters carries meaning, and the distribution of error says more about usability than a hit rate does.

The target of 100 participants is not large, but fixing the definition of performance and the validation procedure up front is a useful model for how to measure medical AI before it reaches the clinic.