AI in healthcare: benchmarks from diagnostics, hospitals and pharma (with sources)

In healthcare, AI already reads scans at specialist level and gives hours back to doctors. It is also the sector where a poorly validated model costs lives. Here the contrast is literal, and every number below is sourced.

Adoption is no longer a pilot. The FDA went from 6 AI-enabled devices cleared in 2015 to 295 in 2025 alone, reaching about 1,451 cumulative. Ambient clinical documentation became the largest-scale case: Abridge grew from around 100 to more than 150 health systems in one year, reaching ~63% of Epic-based hospitals.

Where AI already delivers in healthcare

Application	Result (sourced)	Study
Breast-cancer screening	+29% cancers detected and −44% reading workload, no rise in false positives (100k women)	MASAI trial (The Lancet)
Ophthalmology	Reads OCT across 50+ retinal diseases at specialist level	Moorfields/DeepMind, Nature Medicine
Clinical documentation	~30 min/day per physician saved, with lower burnout	Multicenter study, 6 health systems
Early warning	169 lives saved and −16% mortality across 21 hospitals	Kaiser Permanente AAM, NEJM
Drug discovery	Preclinical candidate in ~18 months vs 2.5–4 years; positive Phase IIa	Insilico (rentosertib), Nature Medicine
Diagnostic reasoning	86.5% on the USMLE exam; 89.4% on rare diseases	Google Med-PaLM 2 / AMIE, Nature

Notice the best cases are not promises: the MASAI trial is randomized, with more than 100,000 women; the Kaiser Permanente system was validated on over 600,000 hospitalizations before becoming routine in 21 hospitals. Validation rigor is what separates the result from the headline.

Where AI broke in healthcare

IBM Watson for Oncology at MD Anderson: about $62 million and a system internally rated "not ready for clinical use", never used on real patients.
Epic Sepsis Model: in external validation published in JAMA, sensitivity of just 33% and AUC of 0.63, missing two-thirds of sepsis cases.
Racial bias (Science, 2019): a population-health algorithm used on millions predicted cost, not illness, and under-referred Black patients. Correcting the target would raise the share of Black patients prioritized from 17.7% to 46.5%.
Google in Thailand: a diabetic-retinopathy system with over 90% lab accuracy rejected 21% of images in real clinics due to poor lighting, becoming a bottleneck instead of help.

Regulation: the clinical-validation gap

Most AI devices cleared in the US lack prospective validation. A Nature Medicine analysis found about 43% of those cleared through 2022 had no published validation data and only ~28% were tested prospectively; 96.4% enter via the 510(k) pathway, which requires no new human testing. The EU AI Act makes AI embedded in a medical device automatically high-risk, with obligations from August 2026.

In Brazil

Reference hospitals such as Albert Einstein and Hospital das Clínicas, plus the DASA and Fleury networks, already run structured radiology-AI programs, several validated on Brazilian populations. On the regulatory side, ANPD is the central data regulator and CFM Resolution 2.454/2026 sets criteria for AI use in medicine, keeping clinical decisions under the physician's responsibility.

The lesson for anyone implementing

Lab accuracy is not clinic accuracy. AI as support is one thing; AI deciding without oversight is another. The potential is huge, $200 to $360 billion a year in US savings on today's technology per NBER and McKinsey, but it only reaches those who validate in the real world and keep the professional in charge. That validation and governance rigor is exactly what Reche brings to any AI project, starting with a diagnosis of where it delivers safely.