Abstract
As large language models (LLMs) are increasingly integrated into clinical workflows, ensuring their consistency across synonymous medical nomenclature is vital for patient safety. This study evaluates the "relational awareness" of three LLMs—GPT-3.5-turbo, GPT-4-turbo, and GPT-4o—regarding brand and generic oncology drug names. While the models demonstrated high proficiency in matching brand names to their generic counterparts, they exhibited significant inconsistencies in complex clinical tasks. GPT-3.5-turbo showed notable biases toward brand names in safety and effectiveness associations. Furthermore, all models performed poorly in drug-drug interaction (DDI) detection, with accuracy rates below 26%. These findings highlight a critical gap between a model's vocabulary and its clinical reasoning, necessitating more robust assessment methods for medical AI.
Introduction
In oncology, a single therapeutic agent is often referred to by both its generic name (e.g., pembrolizumab) and its brand name (e.g., Keytruda). To a human clinician, these terms are functionally identical. However, for an LLM, these names represent different tokens that may have been encountered in different contexts during pre-training. If an LLM provides different clinical recommendations based solely on which name is used, it introduces a "relational fragility" that could compromise patient care and pharmacovigilance.
This study investigates whether LLMs maintain consistent knowledge across lexical substitutions, focusing on drug name recognition, word association, sentiment analysis, and clinical reasoning tasks such as DDI detection and adverse event diagnosis.
Methods
The researchers evaluated GPT-3.5-turbo-0125, GPT-4-turbo, and GPT-4o using a variety of datasets and tasks:
- Drug Name Recognition: A baseline assessment using 367 generic-to-brand and 2,516 brand-to-generic pairs derived from the HemOnc ontology.
- Word Association: A "list preference" task where models associated drugs with antonym pairs (e.g., "safe" vs. "unsafe").
- Sentiment Analysis: Models generated descriptions of drugs, which were then analyzed for sentiment using a RoBERTa-based classifier.
- Clinical Reasoning: The models were tested on 1,000 DDI cases from PrimeKG and 2,438 immune-related adverse event (irAE) cases generated via clinical templates and verified by physicians.
Testing was conducted across multiple "temperatures" (0.0, 0.7, and 1.0) to evaluate how stochasticity influenced consistency.
Key Insights and Results
The Vocabulary-Logic Gap
The study revealed a stark contrast between a model's ability to recognize names and its ability to use them. GPT-4o achieved a high accuracy of 97.38% in matching brand names to generics. However, this "vocabulary" success did not translate to logical consistency in downstream tasks. When the same drugs were used in clinical scenarios, the models often reached different conclusions depending on whether the brand or generic name was used.
Pre-training Biases
GPT-3.5-turbo exhibited a significant bias toward brand-name drugs. It was 43% more likely to associate brand names with "effectiveness" (OR 1.43, p < 0.05) and 76% more likely to describe them as "side-effect-free" (OR 1.76, p < 0.05). This likely reflects the prevalence of pharmaceutical marketing and promotional materials in the models' training data, which often emphasize brand names in a positive light.
Clinical Reasoning Failures
The most concerning finding was the performance in safety-critical tasks:
- DDI Detection: Accuracy for identifying contraindicated drug-drug interactions was abysmal, falling under 26% for all models. There was no significant difference in accuracy between brand and generic names, suggesting a fundamental lack of DDI knowledge rather than just a nomenclature issue.
- irAE Diagnosis: Consistency in diagnosing adverse events varied. GPT-3.5-turbo tended to be overly sensitive, frequently ranking the likelihood of an adverse event as "likely" or "very likely," which could lead to clinical over-diagnosis.
Sentiment Paradox
Interestingly, while GPT-3.5 associated brands with effectiveness in list tasks, it expressed more positive sentiment toward generic names in chatbot-style descriptions (brand mean 0.6703 vs. generic mean 0.9482, p < 0.0001). This suggests that LLM biases are highly sensitive to prompt structure and the specific "mode" of interaction.
Discussion and Conclusion
The results demonstrate that LLMs do not possess a unified "concept" of a drug; instead, their knowledge is tied to specific tokens. Changing a drug's name can trigger different pathways of retrieved information, leading to inconsistent and potentially dangerous clinical outputs.
While newer models like GPT-4o show improvements in consistency and a reduction in brand-name favoritism, the low accuracy in DDI detection remains a significant barrier to clinical adoption. The researchers conclude that before LLMs can be safely integrated into EHRs or used for clinical decision support, they must undergo rigorous robustness testing. Developers must ensure that medical AI can maintain "relational awareness," providing stable and accurate recommendations regardless of the specific nomenclature used.
Summary updated