Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
Misaligned Clinical Risk Classification and Cost Asymmetry in Open-Weight Large Language Models
TL;DR
Four open-weight large language models (LLMs) were evaluated for their ability to integrate clinical risk and cost tradeoffs. Results showed that while these models can represent risk, they struggle to make cost-correct decisions consistently.
✦ Why It Matters
Incorporate cost tradeoff evaluations in clinical AI models to enhance decision-making accuracy.
Key Takeaways
- Evaluate LLMs for clinical risk and cost integration.
- Test model responses to varying cost ratios.
- Incorporate phrasing sensitivity in clinical evaluations.
- Focus on tradeoff tests alongside predictive performance.
Full Summary
Large language models (LLMs) like Qwen-2.5 and Llama-3.1 were studied to understand how they handle patient risk and clinical cost tradeoffs. Using a public diabetes dataset, researchers varied false-negative (FN) to false-positive (FP) cost ratios and examined the models' outputs.
Patient risk was recoverable with an area under the curve (AUC) of approximately 0.83, indicating good predictive performance. However, only the larger models showed consistent shifts in output based on cost direction, and responses to cost magnitude were largely indifferent.
Only 2 out of 12 model-phrasings demonstrated the expected opposing responses to FN versus FP costs. These findings suggest that while LLMs can encode risk and cost information, they do not reliably integrate these factors into their decision-making processes, highlighting the need for clinical evaluations to include tradeoff tests and sensitivity analyses.