Auditable CT Phenotyping Through Report-derived Radiological Observations

TL;DR - ACT is an auditable CT phenotyping method that uses report-derived radiological observations to reveal whether predictions rely on clinically valid evidence or diagnostic shortcuts. It outperformed several baselines while showing that high accuracy can mask reliance on unrelated findings.

• Trained on 38,317 patients and 376,194 mined observations, then evaluated on 25,183 held-out patients across 221 EHR phenotypes.
• Outperformed CT-CLIP on unseen CT pulmonary angiography in zero-shot scoring (0.651 vs. 0.572) and linear probing (0.709 vs. 0.662).
• Just 97 observations filled all 221 top-ranked evidence positions; one calcification phrase ranked first for 20 phenotypes, including clinically unrelated conditions.
• Restricting evidence to clinician-specified observations redirected probes toward phenotype-relevant findings in 86 phenotypes without reducing accuracy (0.751 vs. 0.741).