Medical AI

Medical AI is one of the highest-stakes and most actively invested domains in applied artificial intelligence. It encompasses a diverse range of applications: computer vision systems that detect tumours in radiology scans, language models that extract structured information from clinical notes, predictive models that identify patients at risk of sepsis or readmission, drug discovery pipelines that predict molecular properties, and robotic surgical assistants that augment surgeon precision.

Radiology AI has the most mature commercial deployments. FDA-cleared algorithms can detect pneumothorax on chest X-rays, classify diabetic retinopathy in fundus photographs, detect intracranial haemorrhage on CT scans, and screen mammograms for breast cancer. These systems are designed not to replace radiologists but to triage cases (routing urgent findings for immediate review), reduce read volumes, and catch findings the radiologist might otherwise miss. Landmark studies have shown AI matching or exceeding average radiologist performance on specific tasks, though real-world clinical deployment faces significant challenges around distribution shift and generalisability.

Clinical NLP extracts structured information from unstructured clinical text - physician notes, discharge summaries, operative reports. Named entity recognition identifies mentions of conditions, medications, and procedures. Relation extraction identifies the clinical relationships between them (the patient has condition X, was prescribed medication Y for it). These structured extractions power secondary uses of clinical data for research, quality improvement, and billing.

Foundation models have entered healthcare: BioMedLM, MedPaLM 2, and similar large language models trained or fine-tuned on medical literature achieve physician-level performance on medical licensing examination questions. These models can explain diagnoses, suggest differential diagnoses, and summarise complex medical literature. However, reliability for clinical use requires careful validation - medical hallucinations carry risks that consumer hallucinations do not.

Federated learning plays a special role in medical AI: patient data is legally protected (HIPAA, GDPR) and cannot be easily centralised across hospital systems. Federated learning allows hospitals to collaboratively train models without sharing patient data, enabling larger, more representative training sets than any single institution could assemble.

Regulatory pathways are complex: AI medical devices are regulated as Software as a Medical Device (SaMD) by the FDA in the US, requiring clinical validation studies, De Novo or 510(k) clearance, and post-market surveillance. The regulatory burden is significant and slows commercial deployment, but provides safety assurance that consumer AI lacks.