Conformal Uncertainty-Guided PET Escalation for Multimodal Alzheimer’s Disease Decision Support: Participant-Level Evaluation in ADNI and OASIS-3

Positron emission tomography (PET) provides informative biomarker measurements for Alzheimer’s disease assessment, but universal PET-inclusive inference is resource-intensive. We evaluated a retrospective, calibrated decision-support analysis that invoked a PET-inclusive classifier only when a no-PET classifier was uncertain. Separate participant-level CN-versus-AD analyses were performed in ADNI (n = 273) and OASIS-3 (n = 932) using 20 repeats of stratified five-fold cross-validation with isolated split-conformal calibration. At the prespecified α = 0.05, PET-inclusive inference was invoked for 76.5% of ADNI and 61.2% of OASIS-3 participants. Macro-F1 was 0.856 and 0.788, respectively, compared with 0.801 and 0.702 without PET and 0.845 and 0.752 with matched-random PET assignment; the corrected repeated five-fold conformal-versus-random test was not significant in ADNI (p = 0.151) but was significant in OASIS-3 (p < 0.001). In this binary setting, exact-budget conformal, entropy, confidence, and margin routing selected the same participants; the contribution of conformal prediction is therefore coverage-calibrated threshold selection rather than a distinct uncertainty ranking. These results describe simulated invocation of an already observed PET-derived feature, not prospective PET referral, clinical safety of withholding PET, or health-economic value.