Lightweight Language Models are Prone to Reasoning Errors for Complex Computational Phenotyping Tasks
Although computational phenotyping is a central informatics activity with resulting cohorts supporting a wide variety of applications, it is time-intensive because of manual data review. We previously assessed the ability of LLMs to perform computational phenotyping tasks using computable phenotypes for ARF respiratory support therapies. They successfully performed concept classification and classification of single-therapy phenotypes but underperformed on multi-therapy phenotypes. To better understand issues with these complex tasks, we expanded PHEONA, a generalizable framework for evaluation of LLMs, to include methods specifically for evaluating faulty reasoning. We assessed the responses of two lightweight non-reasoning LLMs (Mistral Small 24 billion and Phi-4 14 billion) and one lightweight reasoning LLM (Qwen-distilled DeepSeek-r1 32 billion) both with and without prompt modifications to identify explanation correctness errors and unfaithfulness errors during phenotyping. For experiments without prompt modifications, both errors were present in responses from all models. For experiments with prompt modifications, we measured the mean absolute change in accuracy relative to the unbiased prompt across biasing conditions. Adding specific few-shot examples aligned with an incorrect phenotype reduced accuracy by at least 5% and up to 10% depending on the model and CoT type. Since reasoning errors were ubiquitous across models, our enhancement of PHEONA to include a component for assessing faulty reasoning provides a practical framework for evaluating LLM reasoning and empirical evidence that reasoning errors occur during complex computational phenotyping.