Structured Visual Evidence Decomposition for Evidence-Grounded Multimodal Screening of Obstructive Sleep Apnea-Hypopnea Syndrome

Researchers developed EviOSAHS, an evidence-grounded AI framework that combines visual analysis of facial features with clinical data to screen for obstructive sleep apnea, achieving 94.86% sensitivity and outperforming direct multimodal prompting approaches. The system decomposes facial images into seven anatomical queries before final clinical adjudication, providing a more reliable and auditable screening workflow than traditional foundation model prompting.

EviOSAHS addresses a critical limitation in applying general-purpose multimodal AI models to medical screening tasks. Direct prompting of foundation models for binary medical decisions often produces unstable outputs with poor calibration, creating safety concerns in clinical workflows. This work demonstrates that structured decomposition—breaking complex visual analysis into discrete anatomical components before synthesis—yields significantly more reliable results than end-to-end approaches.

The framework's architecture reflects a broader shift in medical AI toward explainability and auditability. Rather than treating multimodal models as black boxes, EviOSAHS explicitly separates image analysis from clinical reasoning, generating structured evidence cards that clinicians can review independently. This two-stage design achieved 94.86% sensitivity on a 642-subject validation set, with a notably low 5.14% false-negative rate—critical for a screening application where missing cases creates direct patient harm.

The technical validation is rigorous: ablation studies confirmed that seven-question visual decomposition and balanced adjudication were essential to high-sensitivity performance. The 100% structured parse rate and 93.88% high-visibility rate suggest the approach generalizes reliably across diverse facial presentations. However, the authors appropriately position EviOSAHS as a triage assistant rather than diagnostic system, acknowledging that prospective validation and external testing remain necessary before clinical deployment.

This work establishes a replicable template for deploying multimodal foundation models in regulated medical contexts. The emphasis on decomposition, structured outputs, and human-in-the-loop adjudication addresses regulatory and safety expectations that will shape clinical AI adoption moving forward.

• →Structured visual decomposition into seven anatomical queries significantly outperformed direct multimodal prompting for OSAHS screening accuracy
• →The two-stage framework achieved 94.86% sensitivity with only 5.14% false-negative rate, validating the separation of image analysis from clinical reasoning
• →100% structured parse rate demonstrates the reliability of the systematic decomposition approach across diverse patient presentations
• →The authors emphasize EviOSAHS functions as a triage assistant requiring prospective validation before clinical deployment, reflecting appropriate medical AI governance
• →This methodology provides a generalizable template for deploying foundation models in regulated medical screening workflows through structured reasoning