Detailed · 9 events

AI in Medicine: A History

Medicine is where AI was tried earliest and judged most carefully. From diagnostic systems of the 1970s to 2018, when the US FDA authorised an AI device that gives a screening decision without a clinician reading the image, and Japan approved AI-assisted endoscopy software the same year. Protein-structure prediction reached a Nobel Prize. Meanwhile a widely deployed prediction model fell short in external validation and a flagship medical-AI business was sold off. Successes and failures, with the size and design of each study.

1980s

EVT.0011980sT1The Expert-Systems BoomA History of Artificial IntelligenceRead more →

2010s

At the end of the 2010s medical AI moved from research results to devices that regulators review before sale. In April 2018 the US FDA authorised software that reads retinal photographs for diabetic retinopathy and answers refer or rescreen with no doctor interpreting the image; that December Japan approved software that puts a number on whether a colon polyp is neoplastic, to help the endoscopist decide. One stood in for a reader, the other assisted a doctor, yet both were reviewed as locked versions that do not learn after release. Their evidence was accuracy studies; whether patients actually benefit was carried into the next decade.

2020s

In the 2020s medical AI was judged by who checked the claims, and how. A sepsis predictor built into a health-record system and running in hundreds of hospitals fell far below its vendor's figures once outsiders validated it, and IBM, which had entered medicine with Watson, sold off the data and analytics assets. A language model cleared the pass mark on licensing-exam-style questions, yet its authors wrote that it still trailed clinicians. Meanwhile a randomised trial in Swedish screening built AI support into real reading and compared interval-cancer rates, and protein-structure prediction won a share of the Nobel Prize in Chemistry in 2024.