Disentangling Uncertainty from Evidence Validity in Clinical Trial Screening
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Selective clinical trial screening has two independently observable events: whether an eligibility label is decisive and whether its cited evidence was available, patient-bound, and intact at the protocol cutoff. ActAudit represents these events separately: automatic action requires both a singleton split-conformal set and deterministic source, time, patient, span, and integrity checks over a nonempty citation set. To identify the contribution of each condition, the evaluation fixes every retrieved-evidence and label-score tuple and changes only the routing policy. On 21,073 grouped-holdout patient–criterion pairs, nearly equal-coverage policies (67.2% versus 67.3%) yield 0.40% versus 1.27% false-eligible (FE) risk and 0 versus 11 post-cutoff citations. Removing evidence validity from the identical conformal threshold adds 3.4 coverage points, 0.31 FE points, and 14 post-cutoff citations; removing singleton uncertainty adds 12.3 coverage points and 1.31 FE points. Across five prespecified operating points, the joint rule retains at least a 0.65-point nearest-coverage FE advantage over scalar confidence, and frozen transfer to 6,240 FHIR-native pairs reproduces the separation (0.61% versus 1.68% FE). Evidence removal also changes the label or forces deferral for 88.3% of audited note-dependent actions. The results establish structural evidence validity as an independently measurable determinant of selective action in the evaluated clinical cohorts.
- Primary area:
- causal reasoning