“The model catches the patients we would have lost to no-shows at the specialist. That’s where the accuracy number stops being academic,” the chief medical informatics officer at a US integrated delivery network told us earlier this year. His system deployed an FDA-cleared retinal imaging model across 46 primary care clinics. A 2024 multi-center study showed the same class of model detecting referable diabetic retinopathy with sensitivity above 90% outside the specialist office.
This is what AI in healthcare looks like in production, and it’s only one of dozens of comparable deployments now in routine use.
Diagnostics and clinical workflows
Diagnostic imaging is the most mature application of clinical AI. Deep learning models trained on millions of labeled images detect diabetic retinopathy, certain cancers, and cardiovascular abnormalities at accuracy levels that hold up under prospective validation. Radiology models triage X-rays and CTs by acuity. Digital pathology reduces inter-observer variability. Dermatology tools screen lesions from smartphone photos.
The administrative burden on physicians is one of the leading drivers of burnout. NLP tools that automate documentation give clinicians more time with patients and less time with paperwork.
NLP is changing how clinicians interact with the EHR — extracting structured data from notes, generating patient summaries, flagging drug interactions, and transcribing conversations directly into the chart. The reduction in documentation time at sites running ambient scribing tools is consistently 30-50%.
Predictive analytics and drug discovery
Hospitals are deploying real-time models that watch patient data for early signs of deterioration, sepsis, or readmission risk. Early warning systems give clinical teams hours of lead time to intervene proactively. Sepsis mortality reductions of 20% or more have been reported at sites using these tools at scale.
AI is also compressing the drug discovery pipeline. Models predict molecular behavior, screen candidate compounds, and optimize trial design. Work that used to take years of wet-lab iteration now narrows to the most promising candidates in months. The first drugs designed with substantial AI involvement are moving through clinical trials now.
Ethical considerations
As AI moves deeper into clinical decision-making, questions about bias, transparency, and consent become operational requirements rather than abstract concerns. Models trained on non-representative data underperform on the populations missing from the training set. HIPAA compliance, explainability, and audit trails are not optional. The institutions deploying these systems responsibly are the ones that built governance into the pipeline from day one.