Lu, Qiuhao

Abstract Error analysis is a critical step in evaluating and improving clinical concept extraction models, the most common clinical natural language processing (NLP) task. Unlike corpus annotation, which follows standardized protocols for creating gold-standard datasets, error analysis requires nuanced judgment grounded in both clinical expertise and NLP knowledge. This is especially important given the heterogeneity of clinical text, where variations in documentation style, note structure, and terminology can substantially influence model behavior. Despite its importance, there is currently no standardized, user-level framework to support systematic error analysis in clinical concept extraction task. In this study, we developed and validated MedError, a machine-assisted, human-in-the-loop framework designed to standardize and enhance error analysis for clinical concept extraction tasks. We collected and manually curated a corpus of 1,187 unique errors from a total of 4,237 notes across three different distinct hospitals. The error categories were defined using our previously validated error taxonomy and included 480 false negatives and 707 false positives across 25 error types and 48 clinical concept categories. We evaluated the performance of three proprietary and three open-source large language models (LLMs) in automatically classifying these errors into 26 and 15 predefined categories. We further developed a machine-assisted framework, MedError, which integrates best practices in error analysis, LLM-assisted classification and reasoning, and a user-friendly interface to enable more efficient, reproducible, and context-aware error analysis. The framework supports both single-site and federated multisite error analysis, facilitating the effective translation of clinical NLP systems into real-world settings.

Abstract Current discussion surrounding the clinical capabilities of generative language models (GLMs) predominantly center around multiple-choice question-answer (MCQA) benchmarks derived from clinical licensing examinations. While accepted for human examinees, characteristics unique to GLMs bring into question the validity of such benchmarks. Here, we validate four benchmarks using eight GLMs, ablating for parameter size and reasoning capabilities, validating via prompt permutation three key assumptions that underpin the generalizability of MCQA-based assessments: that knowledge is applied, not memorized, that semantic consistency will lead to consistent answers, and that situations with no answers can be recognized. While large models are more resilient to our perturbations compared to small models, we globally invalidate these assumptions, with implications for reasoning models. Additionally, despite retaining the knowledge, small models are prone to answer memorization. All models exhibit significant failure in null-answer scenarios. We then suggest several adaptations for more robust benchmark designs more reflective of real-world conditions.

Abstract Objective Unlocking clinical information embedded in clinical notes has been hindered to a significant degree by domain-specific and context-sensitive language. Identification of note sections and structural document elements has been shown to improve information extraction and dependent downstream clinical natural language processing (NLP) tasks and applications. This study investigates the viability of a dynamic example selection prompting method to section classification using lightweight, open-source large language models (LLMs) as a practical solution for real-world healthcare clinical NLP systems. Materials and Methods We develop a dynamic few-shot prompting approach to classifying sections where section samples are first embedded using a transformer-based model and deposited in a vector store. During inference, the embedded samples with the most similar contextual embeddings to a given input section text are retrieved from the vector store and inserted into the LLM prompt. We evaluate this technique on two datasets comprising two section schemas, including varying levels of context. We compare the performance to baseline zero-shot and randomly selected few-shot scenarios. Results The dynamic few-shot prompting experiments yielded the highest F1 scores in each of the classification tasks and datasets for all seven of the LLMs included in the evaluation, averaging a macro F1 increase of 39.3% and 21.1% in our primary section classification task over the zero-shot and static few-shot baselines, respectively. Discussion and Conclusion Our results showcase substantial performance improvements imparted by dynamically selecting examples for few-shot LLM prompting, and further improvement by including section context, demonstrating compelling potential for clinical applications.