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

Technical Summary: Structured Visual Evidence Decomposition for Evidence-Grounded Multimodal Screening of OSAHS

Problem Statement
Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) is a prevalent sleep disorder with significant cardiovascular and neurocognitive consequences. While overnight polysomnography (PSG) remains the reference standard for diagnosis, it is resource-intensive and difficult to scale for preliminary risk assessment. Effective pre-PSG screening requires integrating structured clinical risk factors (e.g., BMI, neck circumference, comorbidities) with visible craniofacial and neck morphological cues. However, directly prompting general-purpose multimodal foundation models (Vision-Language Models, VLMs) to issue binary medical decisions often yields unstable, poorly calibrated outputs, such as degenerate negative predictions or excessive positive calls. Furthermore, existing systems often treat multimodal fusion as a "black-box" classification problem, lacking the interpretability required for clinical oversight. The challenge is to organize foundation models into a workflow that acquires, structures, and adjudicates evidence in a clinically controlled manner to prioritize missed-case reduction without autonomous diagnosis.

Methodology: The EviOSAHS Framework
The authors propose EviOSAHS, an evidence-grounded multimodal reasoning framework that decouples image-only anatomical evidence acquisition from final clinical adjudication. The workflow is structured into three distinct stages:

1. Structured Clinical Summary Reconstruction:
   Structured clinical variables (age, sex, BMI, neck circumference, waist-hip ratio, and comorbidities) are converted into a clean, standardized clinical summary. Crucially, this summary is withheld during the visual analysis phase to prevent clinical priors from contaminating image-derived observations. No AHI values or severity labels are included to prevent data leakage.

2. Anatomy-Specific Visual Evidence Acquisition:
   Instead of asking the VLM to diagnose OSAHS directly, the frontal facial image is decomposed into seven fixed anatomical queries targeting specific regions: neck, chin, mouth, face/neck fat, lower jaw, midface, and nose.

   • The VLM acts strictly as an anatomy-specific observer, constrained to report visual observations and visibility levels (high/medium/uncertain) for each query.
   • Direct diagnosis, treatment recommendations, or severity estimation are explicitly prohibited in the prompt.

3. Evidence-Card Construction:
   Each visual observation is parsed and converted into a structured evidence card. These cards contain:

   • Anatomical target and visual observation.
   • Visibility assessment.
   • Risk direction (support/against/uncertain).
   • Evidence strength (weak/moderate/strong).
   • Model-reported confidence (low/medium/high).
   • A concise evidence summary.
     This step transforms free-form visual descriptions into auditable, structured data.

4. Final-Only Clinical Adjudication:
   In the final stage, the seven structured evidence cards are integrated with the clean clinical summary. A Large Language Model (LLM) performs a balanced binary screening adjudication. The model is instructed to explicitly compare supporting, opposing, and uncertain evidence before issuing a final screening-positive or screening-negative label. This ensures the decision is based on an aggregated, reviewable evidence trace rather than unconstrained generation.

Key Contributions

1. Task Formulation: The study reframes OSAHS assessment as a high-sensitivity binary pre-PSG screening task, explicitly separating screening identification from autonomous diagnosis and severity grading.
2. Evidence-Grounded Framework: EviOSAHS introduces a staged workflow where VLMs are constrained to image-only anatomical observation, converting responses into structured, auditable evidence cards.
3. Temporal Separation Strategy: The framework delays the introduction of structured clinical information until the final decision stage, reducing the risk of clinical priors overwriting image-grounded observations.
4. Unified Evaluation: The method is evaluated on a 642-subject cohort against clinical-only prompting, direct multimodal prompting, naive two-stage pipelines, and various ablation variants, supported by paired statistical tests and visual-output audits.

Experimental Results
The framework was evaluated on a retrospective cohort of 642 subjects (584 screening-positive, 58 screening-negative).

• Performance: EviOSAHS achieved 88.47% accuracy, 94.86% sensitivity, 93.74% F1-score, and a 5.14% false-negative rate (FNR).
• Comparative Advantage: It significantly outperformed clinical-only prompting, direct multimodal prompting (which often collapsed to degenerate outputs), and naive two-stage pipelines. Paired McNemar tests confirmed sample-level advantages over all comparators.
• Ablation Studies: Removing the seven-question visual decomposition reduced sensitivity by 22.95 percentage points, and removing balanced final adjudication reduced sensitivity by 19.18 percentage points, confirming that structured decomposition and balanced evidence accounting are critical to the high-sensitivity operating point.
• Auditability: A question-level audit of 4,494 visual outputs showed a 100% structured parse rate and a 93.88% high-visibility rate, supporting the reliability of the image-only evidence acquisition stage.

Significance and Claims
The paper claims that structured visual evidence decomposition can organize general-purpose multimodal foundation models into an auditable, high-sensitivity workflow for binary pre-PSG OSAHS screening. By converting visual descriptions into structured evidence cards and requiring balanced adjudication, the framework enables targeted error analysis and clinician-facing interpretation.

The authors explicitly state that EviOSAHS should be interpreted as a pre-polysomnography triage assistant rather than a definitive diagnostic system. The framework prioritizes missed-case reduction but acknowledges that calibration is required when multiple weak risk cues accumulate to avoid unnecessary referrals. The study concludes that prospective validation, external testing, and calibrated operating-point control are necessary before clinical deployment. The work does not claim to replace PSG or provide autonomous diagnosis, but rather to provide reviewable evidence to support clinician oversight in high-risk triage scenarios.