Hospitals & Care
99 gains87 Traces101 problems
Every measured gain and problem on the record for this topic, newest first. Updated .
AI in hospitals and care
One square per record, darker for stronger evidence:Evidence: 5 high176 moderate106 emerging
Most of the evidence is chatbots tested on written questions and small single-site studies; a few randomized trials and reviews carry more weight. TRV-2026-0355TRV-2026-1613TRV-2026-2032 Chatbots often match or beat doctors on structured written cases, while a 7294-patient trial of an AI alert in live care found no drop in delirium. TRV-2026-1797TRV-2026-1664
A randomized preclinical trial of 77 surgeons watching laparoscopic video shows AI assistance cut the time to find a bleeding site from 3.23 to 2.53 seconds on average. TRV-2026-2032 In a randomized crossover study, clinical trial staff using AI spent 2.71 fewer minutes per item reviewing charts for immunotherapy side effects, though severity grading was much weaker than event detection. TRV-2026-2318 Three randomized trials with over 90,000 US patients show that messages on personal high flu risk raised vaccination 1.7-3.5 percentage points over no message. TRV-2026-1852 A randomized trial of 130 final-year nursing students found AI-adaptive simulation raised leadership scores more than standard training, under controlled simulation conditions. TRV-2026-1262
In a test of 960 responses to clinician-written cases, ChatGPT Health undertriaged 52% of true emergencies, sending diabetic ketoacidosis and impending respiratory failure to 24-48 hour evaluation. TRV-2026-0181 A randomized trial of 7294 patients at Mount Sinai found an AI-triggered alert did not reduce delirium (35.1% vs 32.7%), though only 17.2% had a documented delirium assessment. TRV-2026-1664 In 16,254 responses to 29 standardized vignettes, all 21 chatbots failed differential diagnosis at rates above 0.80 even though final-diagnosis accuracy was high, a benchmark study found. TRV-2026-0145 A test of eight speech-to-text systems found at least one clinically significant error in 30% to 68% of 800 transcripts, and 39.6% of 53 errors with potential for severe harm involved medication type or dose. TRV-2026-1687
A review of 314 FDA-cleared surgical AI devices found only 5.1% had higher-quality prospective evidence. TRV-2026-1693 A scoping review of 1181 reports on AI decision tools found 62.7% stopped at clinical validation and 14.4% evaluated implementation or impact. TRV-2026-1613 A scoping review found only seven eligible studies of agentic AI in healthcare, with just one trial involving patients. TRV-2026-0166 A review of 310 publications found no study that measured AI's effect on critical thinking or clinical reasoning in medical education. TRV-2026-0458
Studies point in different directions on whether AI scribes ease clinician burnout. TRV-2026-0478TRV-2026-0330 An observational study of 263 clinicians in six US health systems reports that self-reported burnout fell from 51.9% to 38.8% after 30 days with an AI scribe. TRV-2026-0478 A systematic review of 6 real-world studies found clinicians reported less documentation time, but burnout on standardized scales was unaffected. TRV-2026-0330 At one cancer center, 190 of 958 clinicians used the AI scribe; after-hours time fell only from 86.9 to 81.4 minutes a week. TRV-2026-1819
Most tests find chatbot answers to patient questions accurate on average, yet poor readability and occasional unsafe answers recur. TRV-2026-1856TRV-2026-2452TRV-2026-1746TRV-2026-1715
See these 79 recordsChatbots beat doctors on structured written cases, but results on triage and differential diagnosis are mixed. TRV-2026-1797TRV-2026-0181TRV-2026-0145TRV-2026-1421
See these 26 recordsAI scribes cut documentation time for many clinicians, though evidence on burnout and adoption is mixed. TRV-2026-1897TRV-2026-0478TRV-2026-0330
A Trace is one record with both a gain and a problem, so it counts on both sides: 186 records show a gain and 188 show a problem. 128 context records are in the record list but not counted, because they report no measured gain or problem.
| Evidence | Gain side | Problem side |
|---|---|---|
| HighStrong study design, backed by at least two independent sources | 3 | 3 |
| ModerateA measured study or an official source | 113 | 107 |
| EmergingEarly, small, unreviewed or reported second-hand | 70 | 78 |
By type of evidence: 95 benchmark or lab tests; 43 observational studies; 42 controlled comparisons; 33 commentaries or narrative reviews; 29 surveys; 15 interview studies; 9 systematic reviews or meta-analyses; 9 randomized trials; 4 case reports; 3 investigations by reporters; 2 news reports; 2 regulatory or legal actions. How evidence is scored
In an online French study of 204 GPs and 942 patient histories, AI recorded histories more effectively than doctors, most clearly for older patients and for allergy and family cancer history.
In an observational study of 12,468 chronic disease patients in three Chinese provinces, AI-enabled care was linked to hypertension control targets met rising from 68% to 82%, with health data accuracy above 95%.
Clinicians using AI note-taking scribes in 6 real-world studies reported spending less time on documentation, though the evidence is sparse and based on self-reports.
A literature review linked AI decision support to vulnerabilities in doctors' physical examination, differential diagnosis, clinical judgment and patient communication skills, though magnitude and reversibility remain uncertain.
In 16,254 responses to 29 standardized clinical vignettes, all 21 AI chatbots tested failed differential diagnosis at rates above 0.80, even though final-diagnosis accuracy was high.
One review in this topic printed a pooled result: one figure calculated across many studies. They are the reviews’ own numbers, checked word for word against their abstracts. They often measure different things, so they are listed side by side, not added up.
AI predicts gallbladder surgery difficulty well, but 16 of 18 studies were at high risk of bias · TRV-2026-0787
AUC 0.818 (95% CI 0.797 to 0.840)
18 studies · search to March 2, 2026
For 32 validation models, the pooled AUC was 0.818 (95% CI, 0.797-0.840).
AUC 0.848 (95% CI 0.829 to 0.868)
18 studies · search to March 2, 2026
The pooled AUC for 27 training models was 0.848 (95% CI, 0.829-0.868).
AI virtual patients and simulation helped trainees in several small studies, while reviews and interviews raise skill erosion as a risk. TRV-2026-1262TRV-2026-1044TRV-2026-0490TRV-2026-0147
See these 39 recordsSurveys find patients and clinicians wary of AI, and regulation and the NHS Palantir contract draw scrutiny. TRV-2026-1843TRV-2026-2394TRV-2026-2335TRV-2026-0359
See these 32 recordsSurgical, imaging and prediction tools mostly rest on small retrospective or preclinical tests with little prospective evidence. TRV-2026-2032TRV-2026-1623TRV-2026-0787TRV-2026-1693
See these 53 recordsEach record sits in one thread at most. A record with both a gain and a problem counts on both sides. 44 records fit none of these threads.
In a 7294-patient trial at Mount Sinai, an AI-triggered health-record alert did not reduce delirium (35.1% vs 32.7%, P = .37) or change adherence to best practices, though only 17.2% had a documented delirium assessment.
26 records · page 2 of 2
Showing the records in one thread of the summary: Diagnosis and triage. Show every record
Researchers tested ChatGPT-4.1 on 564 patients with stable three-vessel or left main coronary artery disease drawn at random from a consecutive angiography registry. The model chose among five treatment categories using patient information and a prompt of the kind real-world patients commonly use, and its answers were…
BMJ Health & Care Informatics
Researchers designed the Skyer benchmark to evaluate fifteen large language models on 55 realistic pediatric emergency department scenarios using a weighting system for over-triage and under-triage…
Researchers compared ChatGPT-5 triage and nurse triage with an expert reference decision in a retrospective review of 1564 adult emergency department patients documented as yellow or green.
Researchers tested four commercially available multimodal language models on 50 cropped clinical photographs, 25 showing burns and 25 showing nonburned skin, repeating each assessment five times for 4000 assessments.
Generative artificial intelligence in hematology is entering a new phase. The dominant question, whether large language models are accurate enough for clinical decision support, is being overtaken by a harder one, as…
This narrative review examined how AI and complementary digital health tools are being used across pre-travel, peri-travel, and post-travel care, covering machine learning risk models,…
Purpose The purpose of this study was to evaluate and compare four AI software programs-ChatGPT-5, Microsoft Copilot, Google Gemini (V 2.5), and OpenEvidence-in generating comprehensive dental…
Researchers compared large language models and clinicians on 34 real-world glaucoma cases, with glaucoma specialists scoring responses on medical accuracy, key-point recall, and logical completeness.
Researchers introduced AMIE, an LLM-based system for diagnostic dialogue, and tested it against 20 primary care physicians in 159 text-based scenarios with patient-actors from Canada, the UK and…
In a structured stress test published February 23, 2026, researchers evaluated ChatGPT Health, OpenAI's consumer health tool launched in January 2026, using 60 clinician-authored vignettes across 21…
Researchers evaluated 21 off-the-shelf large language models, including GPT-5, Claude 4.5 Opus, Gemini 3.0 and Grok 4, on 29 standardized MSD Manual clinical vignettes representing 16,254 responses…
ace