Abstract
Multimodal models that combine medical images with tabular clinical data are routinely deployed on incomplete inputs: a patient may arrive without a scan, or without labs. Conformal prediction (CP) is the standard way to attach a distribution-free coverage guarantee to such models, but its guarantee holds only under exchangeability of the calibration and test data. We show that calibrating on modality-complete data breaks this assumption for test cases missing an entire modality, and standard (marginal) split CP silently loses conditional validity on the modality-incomplete subgroup. We further show that this failure has two faces: on one dataset marginal CP remains valid but collapses to uninformative prediction sets, while on another it retains informativeness only by violating coverage (dropping to 0.73 against a 0.90 target in the longer-trained regime). We introduce Modality-Conditional Conformal prediction (MC2 ), which stratifies calibration by the modality-availability pattern (a Mondrian taxonomy) and thereby restores valid per-pattern coverage with finite-sample guarantees, requiring no retraining of the predictor. Across two public image+tabular datasets (CBIS-DDSM mammography; OL3I cardiac CT) and a controlled multi-class study, MC2 restores the proper validity–informativeness trade-off in both failure regimes. Through a controlled prevalence-swap experiment we further identify class prevalence, not modality redundancy, as the driver of which face appears, and give a concrete threshold-shift mechanism.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=asy0zD9Urd
BibTex
@InProceedings{AjiAar_Guarantees_MICCAISAT2026,
author = { Ajit, Aaron},
title = { { Guarantees That Survive a Missing Scan: Modality-Conditional Conformal Prediction for Multimodal Medical Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}