Two FDA-cleared pulmonary-embolism algorithms showed substantial differences in detection by clot type in a retrospective evaluation of 67,869 CT examinations across 17 facilities. Both produced high specificity, meaning few false positives, but sensitivity fell for smaller, more peripheral and non-acute clots. The study was published in npj Digital Medicine on 21 October.[2][1]
One algorithm examined 30,678 CT pulmonary angiography scans, a dedicated test for pulmonary embolism. It detected 86.8% of report-labelled positive cases, with 99.1% specificity. The second searched 37,191 routine contrast-enhanced CTs for incidental emboli and reached 73.5% sensitivity and 99.8% specificity. The two results describe different clinical tasks rather than a head-to-head test.[2]
Reported real-world performance
| Evaluation | Sensitivity | Specificity | CT scans evaluated |
|---|---|---|---|
| Dedicated PE triage on CTPA | 86.8% | 99.1% | 30,678 |
| Incidental PE detection on routine contrast CT | 73.5% | 99.8% | 37,191 |
Location and acuity change detection
For the dedicated PE algorithm, sensitivity was 99.3% for saddle emboli but only 72.9% for subsegmental emboli. The same model detected 89.7% of acute events and 65.3% of non-acute events. A saddle embolus occupies a large central pulmonary artery junction, whereas subsegmental emboli lie in much smaller peripheral branches. The distinction gives clinical texture to the aggregate sensitivity number.[2]
For triage, the distinction is operationally important. High specificity can limit unnecessary alerts, but a negative output does not rule out an embolism. The lower sensitivity in some subgroups reinforces the need to use the system alongside radiologists’ assessment rather than as a substitute for it.[2]
The answer key came from radiology reports
The reference labels were extracted from existing radiology reports using a separately validated language-model pipeline. That extraction process had reported accuracy of 97% and Cohen’s kappa of 0.94. It made a large retrospective analysis practical, while leaving the evaluation dependent on what the clinical reports recorded.[2]
The authors found lower sensitivity and higher specificity than reported in the products’ FDA-clearance evaluations. Differences in the scan populations, disease characteristics and study designs can affect that comparison. Its value is to show performance in a real clinical population and identify the groups where detection is less reliable.[2]
Hospitals monitoring these systems can therefore ask more useful questions than whether an overall score remains high: which clot locations are missed, whether acute and non-acute cases behave differently, and how changes in the scan mix affect the worklist. That level of monitoring connects model performance to the cases clinicians actually encounter.[2]
Sources & further reading
- Multi-site real-world performance of commercial AI for pulmonary and incidental pulmonary embolism detection npj Digital Medicine / Mansuri et al. · 21 Oct 2026 · Accessed 22 Oct 2026
- Multi-site real-world performance of commercial AI for pulmonary and incidental pulmonary embolism detection Study authors / arXiv · 29 Sep 2026 · Accessed 25 Oct 2026