Current progress and boundaries
Imaging assistance, risk scores, and clinical question-answering have been studied and used, but performance in a single-site test does not guarantee stability across institutions.
Why it remains difficult
Patient populations, diagnostic equipment, and care pathways all cause distribution shifts. Incorrect recommendations can have serious consequences, and accountability, oversight, and update processes must be clear.
Paths worth exploring
Conduct prospective, multi-center validation and retain data sources and model versions. Limit system authority with explicit intended-use boundaries, refusal conditions, and clinician confirmation.
How to evaluate success
Evaluate sensitivity, specificity, calibration, and net benefit across population groups; monitor missed cases, false alarms, and workload changes, and continue auditing after deployment.