Reliability-Aware Clinical Relation Learning for Uncertainty-Robust Multi-Label Chest X-ray Classification

Background: Automated multi-label classification of chest radiographs is complicated by overlapping imaging findings, noisy report-derived labels, and changes in disease co-occurrence across clinical settings. Existing clinical-hierarchy methods can encode label relationships, but a rigid hierarchy may conflate categorical membership with empirical association and propagate annotation errors. Methods: We investigated Reliability-Aware Clinical Relation Learning (ReCaRL), a graph-assisted framework for multi-label chest radiograph classification. An image encoder produces initial finding scores, while a graph constructed from training-label associations provides reliability-weighted, confidence-gated updates. Uncertain annotations are handled with observation masks. The evaluation compares flat, weighted, graph-based, and hierarchy-penalty baselines, with internal and external discrimination and calibration assessments. Results: In the internal evaluation, ReCaRL achieved macro AUROC 0.821 and macro AUPRC 0.437, compared with 0.812 and 0.423 for the masked/weighted baseline. Mean AUROC across five common findings was 0.865 versus 0.858. External testing yielded AUROC 0.799 and AUPRC 0.409. Post-hoc calibration reduced the internal Brier score from 0.113 to 0.102 and the external expected calibration error from 0.082 to 0.061, without changing AUROC or AUPRC. Conclusions: Reliability-aware graph propagation and confidence gating were associated with modest improvements in multi-label chest radiograph classification relative to the matched masked/weighted baseline. Post-hoc calibration improved probability reliability in the reported evaluations. The findings do not establish a benefit in prospective clinical decision-making.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-13
DOI
https://doi.org/10.64898/2026.11.11.26367867
Primary Topic
COVID-19 diagnosis using AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Reliability-Aware Clinical Relation Learning for Uncertainty-Robust Multi-Label Chest X-ray Classification

Yifan Jin, Tao Liu, mark stephen
2 links
medRxiv
COVID-19 diagnosis using AI
preprint

Reliability-Aware Clinical Relation Learning for Uncertainty-Robust Multi-Label Chest X-ray Classification

Yifan Jin, Tao Liu, mark stephen
preprint en

Abstract

Background: Automated multi-label classification of chest radiographs is complicated by overlapping imaging findings, noisy report-derived labels, and changes in disease co-occurrence across clinical settings. Existing clinical-hierarchy methods can encode label relationships, but a rigid hierarchy may conflate categorical membership with empirical association and propagate annotation errors. Methods: We investigated Reliability-Aware Clinical Relation Learning (ReCaRL), a graph-assisted framework for multi-label chest radiograph classification. An image encoder produces initial finding scores, while a graph constructed from training-label associations provides reliability-weighted, confidence-gated updates. Uncertain annotations are handled with observation masks. The evaluation compares flat, weighted, graph-based, and hierarchy-penalty baselines, with internal and external discrimination and calibration assessments. Results: In the internal evaluation, ReCaRL achieved macro AUROC 0.821 and macro AUPRC 0.437, compared with 0.812 and 0.423 for the masked/weighted baseline. Mean AUROC across five common findings was 0.865 versus 0.858. External testing yielded AUROC 0.799 and AUPRC 0.409. Post-hoc calibration reduced the internal Brier score from 0.113 to 0.102 and the external expected calibration error from 0.082 to 0.061, without changing AUROC or AUPRC. Conclusions: Reliability-aware graph propagation and confidence gating were associated with modest improvements in multi-label chest radiograph classification relative to the matched masked/weighted baseline. Post-hoc calibration improved probability reliability in the reported evaluations. The findings do not establish a benefit in prospective clinical decision-making.

medRxiv
Sun Yat-sen University (CN)
Good health and well-being
COVID-19 diagnosis using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon