Antibiotics prescriptions: There’s a large-language model for that

Here’s another high-risk healthcare activity to which generative AI can contribute: prescribing antibiotics. Of course, this application should not be used absent the oversight of a qualified human healthcare professional. But it’s worth noting that the option is now recognized by experts in the field. 

That’s one takeaway from commentary published this month in Infectious Diseases and Therapy. 

The use of large-language AI models for antibiotic prescribing, the authors state, “offers immense potential to improve patient outcomes. These tools can provide rapid, nuanced suggestions that complement clinical expertise, enhancing efficiency and decision-making.”

The team emphasizes that this application is fraught with challenges. Not least among these are defining acceptable error margins, addressing hallucinations and mitigating variability in performance. Taken together, these potential pitfalls call for ongoing research and careful oversight, they note. 

The paper’s lead author is Daniele Roberto Giacobbe, MD, PhD, an associate professor in infectious diseases at the University of Genoa in Italy. Senior author is Matteo Bassetti, MD, PhD, director of infectious disease care at San Martino University Hospital in the same city. They and colleagues lay out six pointers for consideration by those considering AI in antibiotic prescribing: 

1. Large-language AI models have great potential to improve outcomes for patients with infectious diseases. However, LLM-based support for antibiotic prescribing is complex.

Antibiotic prescribing adds [a] layer of complexity, as clinicians must balance two objectives: selecting the most effective treatment for the patient and minimizing the risk of resistance development. 

‘Misjudging hallucinations or omissions could disproportionately affect one of these priorities, creating new challenges for clinicians navigating this dual responsibility when exploiting the aid of LLMs.’

2. There are distinctive commonalities—and crucial conceptual differences—between the use of LLMs as assistants in scientific writing and in supporting antibiotic prescribing in real-world practice.

While writing commentary articles can be challenging, it generally lacks the immediate and direct implications for patient health that come with antibiotic prescribing. 

‘Prescribing antibiotics involves critical decisions with far-reaching consequences for individual outcomes and public health, requiring clinicians to weigh patient-specific factors against broader antimicrobial stewardship principles.’

3. LLMs operate probabilistically and in a non-explainable (or only partly explainable) way, characteristics that together make the risk of error a peculiar moving target for complex tasks like antibiotic prescribing.

Assessing and mitigating the risk of error in LLMs-generated [prescriptions] is challenging due to inherent variability. LLMs operate probabilistically, predicting the most likely next “token” (word or part of a word) in a sentence based on their training data. 

‘This approach introduces variability, even for identical prompts, and makes the risk of error a moving target.’

4. Numerous other challenges exist, including the variable quality of training data and the need for properly addressing hallucinations or omissions.

Proprietary models often do not disclose their training datasets for the most recent versions of their models, making it difficult to evaluate their robustness and the overall quality and quantity of data for specific topics. 

‘Additionally, human interaction adds another layer of variability.’

5. Training current and future clinicians to optimize their interaction with LLMs is essential to achieve synergistic improvements in performance, surpassing what humans or LLMs can achieve alone. 

Antibiotic prescribing exemplifies a challenging medical decision-making process, influenced by numerous factors, including patient-specific variables, local antimicrobial resistance patterns, and clinical guidelines. 

‘AI tools, particularly large-language models, offer an opportunity to support clinicians in navigating these complexities.’

Expounding on the latter point, the authors add: “These models, through their ability to process and synthesize vast amounts of information, could complement clinical expertise by providing rapid, contextualized suggestions for treatment.”

More: 

‘While the excitement surrounding LLMs is justified, realizing their full potential will require a cautious, methodical approach. The path forward involves balancing innovation with rigorous evaluation, ensuring that these tools are integrated safely and effectively into clinical practice.’

The paper is available in full for free.