Back

Medical concept understanding in large language models is fragmented

Deng, L.; Chen, L.; Liu, M.

2026-03-05 health informatics
10.64898/2026.03.03.26347552 medRxiv
Show abstract

Large language models (LLMs) perform strongly across a wide range of medical applications, yet it remains unclear whether such success reflects genuine understanding of medical concepts. We present an ontology-grounded, concept-centered evaluation of medical concept understanding in LLMs. Using 6,252 phenotype concepts from Human Phenotype Ontology, we decompose concept understanding into three core dimensions--concept identity, concept hierarchy, and concept meaning--and design corresponding benchmarks for each dimension. Across a representative set of contemporary LLMs, best-performing models achieve high accuracy on concept identity (90.6%) and hierarchy (83.8%), but lower performance on concept meaning (72.6%). Concept-level analysis reveals substantial fragmentation in LLM understanding: only 57.7% of concepts are consistently understood across all three dimensions, while 41.3% show partial understanding and 1.1% are not captured in any dimension. These results demonstrate that strong application-level performance of LLMs can mask fundamental gaps in concept-level understanding, highlighting the necessity for ontology-grounded evaluation in medical AI.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

1
npj Digital Medicine
○ 118 papers in training set
Top 0.4%
12.7%
Similar papers in this journal
2
Scientific Reports
○ 3612 papers in training set
Top 7%
7.8%
Similar papers in this journal
3
JAMIA Open
◐ 42 papers in training set
Top 0.2%
6.6%
Similar papers in this journal
4
Journal of the American Medical Informatics Association
◐ 71 papers in training set
Top 0.5%
6.6%
Similar papers in this journal
5
BMC Medical Informatics and Decision Making
○ 43 papers in training set
Top 0.3%
6.2%
Similar papers in this journal
6
Journal of Biomedical Informatics
○ 47 papers in training set
Top 0.3%
5.4%
7
PLOS Digital Health
● 106 papers in training set
Top 2%
4.0%
8
JMIR Medical Informatics
◐ 18 papers in training set
Top 0.1%
3.5%
50% of probability mass above
9
iScience
○ 1154 papers in training set
Top 7%
3.2%
10
Artificial Intelligence in Medicine
○ 17 papers in training set
Top 0.2%
3.2%
11
Bioinformatics
◐ 1204 papers in training set
Top 5%
3.2%
12
Computers in Biology and Medicine
○ 128 papers in training set
Top 2%
2.4%
13
PLOS ONE
● 5266 papers in training set
Top 41%
2.4%
14
JCO Clinical Cancer Informatics
● 22 papers in training set
Top 0.3%
2.4%
15
Patterns
○ 78 papers in training set
Top 1%
2.1%
16
Nature Medicine
○ 125 papers in training set
Top 1%
2.1%
17
Journal of Medical Internet Research
◐ 87 papers in training set
Top 1%
1.9%
18
Nature Communications
○ 5641 papers in training set
Top 43%
1.9%
19
Frontiers in Bioinformatics
○ 49 papers in training set
Top 0.5%
1.7%
20
eBioMedicine
○ 183 papers in training set
Top 3%
1.7%
21
The Lancet Digital Health
○ 25 papers in training set
Top 0.4%
1.4%
22
GigaScience
◐ 212 papers in training set
Top 3%
1.1%
23
BMC Bioinformatics
○ 457 papers in training set
Top 5%
1.0%
24
BMJ Health & Care Informatics
● 15 papers in training set
Top 1%
0.8%
25
Biology Methods and Protocols
◐ 61 papers in training set
Top 2%
0.8%
26
PLOS Computational Biology
● 1863 papers in training set
Top 20%
0.8%
27
European Journal of Human Genetics
○ 58 papers in training set
Top 1%
0.6%
28
International Journal of Medical Informatics
○ 26 papers in training set
Top 2%
0.6%
29
BioData Mining
○ 22 papers in training set
Top 0.9%
0.6%
30
Frontiers in Digital Health
○ 24 papers in training set
Top 1%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.