Abstract
Contactless respiratory-rate (RR) monitoring enables scalable and unobtrusive vital-sign assessment without wearable sensors, but vision–language models (VLMs) may produce confident estimates when visual evidence is weak, motion-corrupted, or internally inconsistent. We propose EviCal-RR, an evidence-calibrated framework for safe selective RR reporting. For each video window, EviCal-RR constructs a structured profile containing visual and motion summaries, respiratory waveforms, frequency-domain candidates, candidate disagreement, evidence quality, and artifact indicators. An evidence-aware VLM proposes an estimate and rationale; a safety critic reviews evidence sufficiency, cross-channel consistency, and physiological plausibility; and a deterministic evidence contract selects report, flag, or abstain, preventing model confidence from overriding insufficient evidence. Across three respiratory video datasets and multiple frozen medical and general-purpose VLM backbones, EviCal-RR improves RR accuracy, confidence alignment, and evidence-risk recall over direct and evidence-aware prompting while substantially reducing unsupported confident reports. These results support evidence-grounded selective reporting over plausible always-answer prediction for safer contactless physiological monitoring.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SAFER_021.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=u3ZFJmpyG3
BibTex
@InProceedings{SakSha_EviCalRR_MICCAISAT2026,
author = { Sakib, Shadman AND Roy, Nirmalya},
title = { { EviCal-RR: Evidence-Calibrated Reasoning for Contactless Respiratory Monitoring } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}