[KongchangAI]
· 2 min read· 1,150 words

AI Drug Safety Knowledge Gap: LABEL Study Reveals Why Language Models Overlook Contraindications

AI Drug Safety Knowledge Gap: LABEL Study Reveals Why Language Models Overlook Contraindications

Public text mentions drug benefits about twice as often as safety limits, and language models faithfully inherit this bias.

The paper introduces LABEL, the literature asymmetry between efficacy and limits. Analyzing 574 drugs, it finds benefit statements in Wikipedia and PubMed abstracts appear about twice as often as safety limits (within-drug odds ratio 2.19), despite FDA label requirements. Qwen2.5 scaling from 0.5B to 14B improves indication accuracy by 18.3 points but leaves contraindication accuracy near random guessing. Retrieval text or routing to FDA labels improves safety accuracy, suggesting clinical AI should prioritize regulatory labels and structured databases.

Contraindications and indications on drug labels should carry equal weight, but a new study finds that public text documents the former far less often than the latter—and language models faithfully mirror this bias. The paper, titled LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models, analyzes how 574 drugs are described in Wikipedia and PubMed abstracts, revealing this overlooked safety knowledge gap.

LABEL study illustration

Key Finding: How Large Is the Documentation Gap for Safety Limits?

The research team selected 574 drugs from the PrimeKG database and compared how often each drug's own indications versus contraindications appear in free text. The results show that in Wikipedia articles and PubMed abstracts, mentions of a drug's benefits appear roughly twice as often as mentions of its safety limits (within-drug odds ratio 2.19). The authors name this phenomenon LABEL (Literature Asymmetry between Efficacy and Limits).

This asymmetry is not coincidental. Although FDA regulations require drug labels to list contraindications with equal rigor, public-facing educational texts and academic abstracts tend to emphasize a drug's therapeutic value, with comparatively weak coverage of risk information.

Within-drug odds ratio is a statistic used in paired designs to measure the relative frequency of two events: for each drug, it counts how often its indications versus contraindications are mentioned in text, computes the ratio within that drug, and then aggregates all drug-level ratios into an overall estimate. Because the comparison occurs within the same drug, confounders such as drug popularity, time on market, and total literature volume are naturally canceled out. The value of 2.19 can therefore be read directly as "for the same drug, benefit statements outnumber safety-limit statements by about twofold," without worrying that popular drugs inflate the overall frequency.

PrimeKG is a multimodal precision medicine knowledge graph that integrates nodes for drugs, diseases, genes, proteins, side effects, and structured relationships among them (such as drug–indication and drug–contraindication edges). Using it as the sampling frame ensures that all 574 drugs have officially annotated indications and contraindications in the dataset as a "gold standard," giving the text-frequency comparison a solid basis.

How Do Language Models Mirror This Bias?

The study further tested whether language models inherit this textual bias. Using the Qwen2.5 series, from 0.5B to 14B parameters, model accuracy on indication questions improved by 18.3 percentage points, but accuracy on contraindication questions barely budged from near random guessing.

This means that simply scaling up models does not automatically close the safety knowledge gap. The models do not lack the ability to reason about risk; the training corpora simply lack sufficient safety facts.

The Remedial Effect of Retrieval Augmentation and Official Labels

When researchers provided the model with retrieved text that explicitly stated the answer, the model used safety facts as efficiently as benefit facts. This shows the model is capable of processing risk information—it just needs the right knowledge source.

Another experiment found that routing drug safety questions to official FDA labels boosted model accuracy by 15.8 percentage points, as long as the label explicitly stated the relevant fact. This points to a feasible improvement direction for clinical AI systems: rather than relying on general public text, route drug safety questions to regulatory labels and structured databases.

Parameter scaling refers to improving capabilities by increasing model parameters and training data. The value of the LABEL experiment lies in showing that scaling is not a panacea: indication accuracy improves by 18.3 percentage points with scale, indicating the model can learn and amplify knowledge about efficacy from the corpus; contraindication accuracy stays near random guessing, indicating that when a certain type of fact is itself sparse in the training corpus, scaling only makes the model more fluent at reciting the half of knowledge it has "seen a lot." In other words, scaling improves the compression and retrieval efficiency of existing information, not the凭空 filling of missing facts. This also explains why models seem to "know" many drug uses but cannot answer who should not use them or under what conditions they are contraindicated—it is a mirror of the data distribution, not a failure of the reasoning module.

Retrieval-augmented generation (RAG) refers to retrieving relevant passages from external knowledge sources and injecting them into the context before the model generates an answer, so the model answers based on the given text rather than relying solely on knowledge memorized in its parameters. The experimental design in this study is precisely a controlled version of RAG: when the retrieved text explicitly states the answer, the model's utilization of safety facts is comparable to benefit facts, ruling out the explanation that "the model does not understand risk semantics" and locating the problem on the knowledge supply side.

FDA prescribing information is a structured document that has undergone regulatory review, in which the "Contraindications" and "Warnings and Precautions" sections state risk facts in a standardized format. Routing safety questions to such labels essentially replaces open corpora with authoritative structured sources—an intervention at the level of "retrieval source selection." It does not change the model, only what the model sees, so deployment costs are low, auditability is strong, and it better meets clinical requirements for traceability.

Implications for Clinical AI Systems

The core message of the LABEL study is: the root cause of model failures on drug safety lies in the scarcity of safety facts in free-text corpora, not in defects in model reasoning. Clinical AI systems should therefore prioritize integration with official regulatory labels and structured drug databases, rather than relying solely on general public records such as Wikipedia or PubMed abstracts.

This finding also warns medical writers: when describing drug efficacy, equal attention must be paid to communicating safety limits such as contraindications, otherwise AI systems will continue to amplify this knowledge bias.

The research team has released code and data for reproduction and extension by future studies.

One-line Summary

Drug benefits are mentioned about twice as often as safety limits in public text, and language models faithfully inherit this bias.

Paragraph Summary

The paper introduces the concept of LABEL—literature asymmetry between efficacy and limits. The study selected 574 drugs and compared the frequency of their indications versus contraindications in Wikipedia and PubMed abstracts, finding that benefit statements appear about twice as often as safety limits (within-drug odds ratio 2.19), even though FDA labels require contraindications to be listed with equal rigor. Language model experiments show that as Qwen2.5 scales from 0.5B to 14B, indication accuracy improves by 18.3 percentage points, while contraindication accuracy remains near random guessing, indicating that scaling cannot compensate for the scarcity of safety facts in the corpus. When provided with retrieved text that explicitly states the answer, the model can effectively use safety information; routing questions to official FDA labels boosts accuracy by 15.8 percentage points. The study concludes that clinical AI should prioritize integration with regulatory labels and structured databases rather than relying on general public text.

#AI drug safety#LABEL study#contraindications#language models#clinical AI#FDA labels#retrieval-augmented generation
Share:

Related articles