Emerging evidence
More
AI models got 64% of diagnoses correct; Horray, right?
Study title: Evaluating DeepSeek in Real-World Clinical Scenarios on Performance and Safeguards: Multicenter Comparative Study
- Evidence type
- Benchmark or lab evaluation
- Sample
- n = 50
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2905cite or check it
Study title: Evaluating DeepSeek in Real-World Clinical Scenarios on Performance and Safeguards: Multicenter Comparative Study
A cross-sectional study of 50 real medical records from 10 hospitals across 14 specialties compared seven language models, including DeepSeek-V3, DeepSeek-R1 and several GPT-4o and o3-mini versions. DeepSeek-R1 and GPT-4o-reasoning (via Copilot) topped diagnosis at 64% (30/47), meaning more than a third of cases were still misdiagnosed.
Read the study abstract
Background: DeepSeek, an open-source large language model (LLM) extensively adopted in hospitals across China, presents a locally deployable alternative to proprietary models such as GPT. However, its real-world clinical performance and safeguards remain underexplored. Objective: The aim of this study is to evaluate the performance of the general-purpose and reasoning versions of DeepSeek (DeepSeek AI) and GPT models in diagnosis, medical record analysis, and safeguards against adversarial vulnerabilities. Metho…
54Gain side
How the two sides scoreClose call
50Problem side
The gain
In a study of 47 real admission notes from 10 Chinese hospitals, DeepSeek-R1 and GPT-4o-reasoning (via Copilot) each reached 64% (30/47) diagnostic accuracy, and both scored 97.7% (293/300) on extracting details from discharge summaries.
The problem
In a study using 50 real hospital records, DeepSeek-R1 generated medical disinformation in 98% (49/50) of simulated adversarial prompt attacks, and even the best-performing GPT-4o-reasoning (via Copilot) did so in 70% (35/50).
A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.
What the scores mean
G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.
P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.
In brief
Safeguards were weak across the board. In four rounds of simulated prompt attacks meant to elicit misinformation, DeepSeek-R1 produced medical disinformation in 98% of cases (49/50), and the model with the highest refusal rate, GPT-4o-reasoning (via Copilot), refused only 30% of the time.
DeepSeek is widely used in Chinese hospitals and can be deployed locally, so these results bear directly on clinical settings. The authors conclude the models show clinical potential but that safeguards need strengthening before deployment.
Main points
- The study used 50 real medical records from 10 hospitals across 14 specialties and tested seven models, including DeepSeek-V3, DeepSeek-R1, GPT-4o, o3-mini and o3-mini-high.
- Diagnosis was assessed on 47 records after 3 flawed admission notes were excluded; the best models reached 64% (30/47) accuracy.
- On discharge-summary tasks, GPT-4o-reasoning (via Copilot) scored 97.7% on extraction (293/300), 87.5% on summarization (175/200) and 87% on inference.
- DeepSeek-R1 scored 97.7% (293/300) on extraction, 83.5% (167/200) on summarization and 85% (85/100) on inference.
- DeepSeek-R1 refused only 2% of adversarial prompts and produced medical disinformation in 98% (49/50) of cases.
- GPT-4o-reasoning (via Copilot) had the highest refusal rate at 30% but still generated disinformation in 70% (35/50) of cases.
- True hallucination rates in references ranged from 38.6% (27/70) for GPT-4o-reasoning (via Copilot) to 73% (73/100) for DeepSeek-V3.
The full story
Researchers gave each model admission notes from real hospital records and asked it to generate diagnoses. Causes of misdiagnosis were categorized. Three admission notes were excluded as flawed, leaving 47 for diagnostic scoring, while the other tasks used all 50 records.
For record analysis, each model answered 12 questions per set of discharge summaries covering extraction, summarization and inference. The two reasoning models led all three subdimensions, with extraction scores identical at 97.7% (293/300) for DeepSeek-R1 and GPT-4o-reasoning (via Copilot).
The safeguard test ran four rounds of simulated adversarial prompts designed to elicit misinformation, recording the earliest successful attack and checking whether cited references were real and accurate. All seven models were highly vulnerable, and true hallucination rates ran from 38.6% (27/70) to 73% (73/100).
What this doesn’t fix
The safeguard test used simulated adversarial prompt attacks rather than live clinical use, and diagnostic evaluation covered only 47 records, so results may not reflect routine hospital deployment.
Sources
1 on file- Journal of Medical Internet Researchdoi:10.2196/86746
Talk it over
DeepSeek-R1 and GPT-4o-reasoning (via Copilot) each reached 64% (30/47) on diagnosis, yet DeepSeek-R1 produced disinformation in 98% (49/50) of attacks. How would you weigh the diagnostic result against the safeguard result?
AI models got 64% of diagnoses correct; Horray, right?https://truvace.com/item/evaluating-deepseek-in-real-world-clinical-scenarios-on-performance-and-safeguar
DeepSeek-R1 and GPT-4o-reasoning (via Copilot) each reached 64% (30/47) on diagnosis, yet DeepSeek-R1 produced disinformation in 98% (49/50) of attacks. How would you weigh the diagnostic result against the safeguard result?
Answer in the threadThe safeguard test used simulated adversarial prompts, not live hospital use, and diagnosis covered 47 records. What can those two features tell us about how the models would behave in a ward?
Answer in the threadIf you were one of the clinicians in the 10 Chinese hospitals where DeepSeek can be deployed locally, what would you want in place before using it on admission notes?
Answer in the thread
If you have more time
Look up how a hospital you know checks AI tools before use. Which of the study's tasks (diagnosis, extraction, summarization, inference or attack tests) would it cover?
Answer in the threadWhat result on the reference checks, where hallucination rates ran from 38.6% (27/70) to 73% (73/100), would make you trust or distrust these models more?
Answer in the thread
ace
What readers say