Clinical Scorecard: Navigating the Explainable AI Challenge Amid Knowledge Disparities in Specialist AI for Glaucoma Referrals in Primary Care Settings
At a Glance
Category
Detail
Condition
Glaucoma, a chronic progressive optic neuropathy causing irreversible vision loss
Key Mechanisms
Damage to optic nerve head and retinal nerve fiber layer; assessed via intraocular pressure, visual field tests, and optical coherence tomography
Target Population
Patients at risk of glaucoma progression managed initially by primary eye care providers such as optometrists
Care Setting
Primary care and eye care settings with referrals to glaucoma specialists
Key Highlights
Specialist AI models trained on clinical data can identify urgent glaucoma referrals with higher accuracy than human providers alone.
Explainable AI (XAI) methods, both intrinsic and post-hoc, did not improve human-AI team performance and introduced uncertainty and biases.
Human-AI teams underperformed compared to AI alone, highlighting challenges in trust calibration and effective collaboration amid knowledge imbalance.
Guideline-Based Recommendations
Diagnosis
Combine structural (OCT) and functional (visual field) data for comprehensive glaucoma assessment.
Use AI models trained on multimodal clinical data to support referral decisions.
Management
Primary eye care providers should leverage AI decision support to improve timely and accurate referrals to glaucoma specialists.
Maintain human clinical judgment alongside AI recommendations to mitigate AI-specific errors.
Monitoring & Follow-up
Monitor patient progression through repeated intraocular pressure measurements, visual field testing, and optic nerve imaging.
Risks
AI models may exploit spurious correlations or ignore relevant clinical features, leading to errors uncommon in human reasoning.
Explainable AI can cause anchoring bias and over-reliance on incorrect AI outputs, reducing overall decision accuracy.
Patient & Prescribing Data
Patients suspected of glaucoma or at risk of progression evaluated in primary eye care
Optometrists vary in screening and referral thresholds; AI can standardize identification of urgent cases but requires careful integration to preserve human agency.
Clinical Best Practices
Integrate AI decision support tools that are transparent and tailored to primary care workflows.
Train providers to critically appraise AI recommendations and recognize AI limitations.
Avoid over-reliance on AI explanations alone; use them as adjuncts to clinical expertise.
Foster collaboration between AI developers and clinicians to design effective human-AI interfaces.
Review identifies recurring sex-based differences across specialties, including lower use of intensive and guideline-compliant treatment among female patients.
