Editing AI in Healthcare (section)

== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ Healthcare AI Application Maturity
! Application !! Regulatory Status !! Clinical Validation !! Deployment Maturity
|-
| Diabetic retinopathy screening || FDA cleared (IDx-DR) || Strong RCT evidence || Commercially deployed
|-
| Chest X-ray pneumonia detection || Multiple FDA clearances || Strong retrospective, some prospective || Widely deployed
|-
| Sepsis prediction || FDA-cleared (some) || Mixed prospective evidence || Deployed, contested
|-
| Drug discovery (small molecules) || N/A (research tool) || Early-stage || Research/early commercial
|-
| AlphaFold protein structure || N/A (research tool) || Strong scientific validation || Research deployed (PDB)
|-
| ECG interpretation || Multiple FDA clearances || Good validation || Deployed in wearables, hospitals
|}

'''Critical failure modes:'''
* '''Underperformance on subgroups''' — Models trained predominantly on one demographic group perform worse on others. Studies have shown certain dermatology AI systems perform worse on darker skin tones; radiology AI performs worse on images from equipment types not in training data.
* '''Feedback loops''' — Deploying an AI changes clinical behavior, which changes outcomes, which changes future training data. A sepsis alert AI might cause intervention that changes the data distribution.
* '''Alert fatigue''' — If a CDSS generates too many false positives, clinicians start ignoring its alerts — including true positives. Precision matters as much as recall.
* '''Distribution shift at deployment''' — Patient population, imaging equipment, and disease prevalence change over time. Models need continuous monitoring and revalidation.
* '''Missing data''' — EHR data has extensive missing values. Models must handle this gracefully; naive imputation can introduce systematic errors.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">