🤖AIBytes
Two clinical studies that deserve a closer look.
Can AI Improve Hospital Discharge Summaries?
This randomized trial tested whether patient-friendly discharge summaries generated with GPT-4o could improve patients confidence and ability to manage their care after leaving the hospital.
Methods
Researchers enrolled 128 adults from one internal medicine department in Germany.
Patients received either:
a standard discharge summary, or
an AI-generated, patient-oriented summary
Every AI draft was reviewed by the treating physician. The main outcome was patient activation, measured with the 13-item Patient Activation Measure.
Results
Patients receiving the AI-generated summary had significantly higher activation scores:
Median difference: 9.6 points
95% confidence interval: 4.3 to 15.1
More patients reached the highest activation level:
AI group: 34 of 64, or 53%
Control group: 12 of 64, or 19%
Patients also rated the AI summaries as more helpful, understandable, empathetic, and satisfactory.
Trust was similar between groups, and health literacy did not improve significantly.
No critical errors were identified after physician review.
Key Takeaways
AI-generated discharge summaries improved patient activation in this randomized trial.
Physician review remained essential.
The study did not measure readmissions, medication adherence, long-term outcomes, cost, or performance without clinician oversight.
Larger multicenter studies are still needed.
🔗 Rust P, Frings J, Meister S, et al. Effects of large language model-generated, patient-oriented discharge summaries on patient activation: a single-centre, single-blind, randomised controlled trial in Germany. Lancet Digit Health. 2026;8:100991. Published online May 18, 2026. doi:10.1016/j.landig.2026.100991.
Can AI and Physicians Work Better Together?
NOHARM is a benchmark that checks how safely and completely AI systems give clinical management advice. It looks at both harmful recommendations and important actions the AI leaves out.
This updated version also includes a randomized study of physician–AI teamwork.
It asks three questions: How often can AI advice lead to harm? Do clinical AI tools perform better than general-purpose models? And does AI help physicians make better management plans?
Methods
NOHARM included 1,100 tasks based on 100 physician-to-specialist consultations across 10 specialties.
Researchers tested 45 large language models and four clinical AI tools: AMBOSS LiSA, Doximity Ask, OpenEvidence, and Glass Health.
The update also included a randomized crossover study of 101 U.S. attending physicians using conventional resources, AI assistance, or any available resource.
Results
Potentially severe errors ranged from 2.9% to 24.6% of cases, depending on the system.
More than 80% of severe errors came from missing an important action, such as a test, treatment, or follow-up step, rather than recommending something harmful.
The clinical AI tools ranked highest by severity-weighted performance:
AMBOSS LiSA: 86.15
Doximity Ask: 84.51
OpenEvidence: 80.02
Glass Health: 79.74
All four performed better than the general-purpose models, although the differences among the clinical tools were not statistically significant.
AI assistance improved physicians’ management plans. However, physicians often left out useful recommendations the AI had already suggested.
The best performance came from combining the physician’s plan with all appropriate AI recommendations. Some multi-agent AI teams also improved results when a strong model reviewed the others.
Source: Wu et al. Figure 3 (Preprint).
Key Takeaways
Clinical AI tools performed better than general-purpose models on these consultation tasks.
The main safety risk was incomplete advice, not obviously harmful advice.
AI improved physician performance, but physicians did not always use its best recommendations.
The study measured potential harm in benchmark cases, not harm observed in real patients.
🔗 Wu D, Haredasht FN, Maharaj SK, et al. First, do NOHARM: a medical safety benchmark and randomized study of physician and AI teaming on clinical consultations. doi.org/10.48550/arXiv.2512.01241 Preprint.
🧬AIMedily Snaps
Fast updates clinicians should not miss.
United Nations: Preliminary Report of the Independent International Scientific Panel on AI. (Link)
Stanford healthcare AI industry report 2026. (Link)
OpenEvidence launches EvidenceGrade to help assess the strength of evidence behind their answers. (Link)
Medscape: When AI Makes the Call, Doctors May Take the Blame. (Link)
Healthcare costs:The impact of AI on U.S. healthcare spending (Link)
Forus and the American Gastroenterological Association partner to improve medication access for GI patients. (Link)
🧪Research Signals
New papers worth your time.
NEJM: Mount Sinai’s AI triage solution. (Paper)
JAMA: Speed and safety in pediatric AI. (Paper)
JAMA: AI gestational age assessment using blind-sweep ultrasound. (Paper)
Nature: Fairness in multimodal clinical AI. (Paper)
Nature: AI prediction of response to neoadjuvant therapy. (Paper)
NEJM: Algorithmic scheduling reduced dermatology wait times. (Paper)
🦾TechTools
AI medical tools
Uses AI to predict cancer outcomes from tumor pathology supporting more personalized treatment decisions.
Uses AI to estimate whether a lung nodule is cancerous to help guide follow-up and management.
📈 Productivity AI tool of the week:
Analyzes spreadsheets and datasets using natural language. Creates charts, performs statistical analyses, and explains results without coding.