Large Language Model-Supported Identification of Intellectual Disabilities in Clinical Free-Text Summaries: Mixed Methods Study.

Background: Free-text clinical data are unstructured and narrative in nature, providing a rich source of patient information, but extracting research-quality clinical phenotypes from these data remains a challenge. Manually reviewing and extracting clinical phenotypes from free-text patient notes is a time-consuming process and not suitable for large-scale datasets. On the other hand, automatically extracting clinical phenotypes can be challenging because medical researchers lack gold-standard annotated references and other purpose-built resources, including software. Recent large language models (LLMs) can understand natural language instructions, which help them adapt to different domains and tasks without the need for specific training data. This makes them suitable for clinical applications, though their use in this field is limited.

Objective: We aimed to develop an LLM pipeline based on the few-shot learning framework that could extract clinical information from free-text clinical summaries. We assessed the performance of this pipeline for classifying individuals with confirmed or suspected comorbid intellectual disability (ID) from clinical summaries of patients with severe mental illness and performed genetic validation of the results by testing whether individuals with LLM-defined ID carried more genetic variants known to confer risk of ID when compared with individuals without LLM-defined ID.

Methods: We developed novel approaches for performing classification, based on an intermediate information extraction (IE) step and human-in-the-loop techniques. We evaluated two models: Fine-Tuned Language Text-To-Text Transfer Transformer (Flan-T5) and Large Language Model Architecture (LLaMA). The dataset comprised 1144 free-text clinical summaries, of which 314 were manually annotated and used as a gold standard for evaluating automated methods. We also used published genetic data from 547 individuals to perform a genetic validation of the classification results; Firth's penalized logistic regression framework was used to test whether individuals with LLM-defined ID carry significantly more de novo variants in known developmental disorder risk genes than individuals without LLM-defined ID.

Results: The results demonstrate that a 2-stage approach, combining IE with manual validation, can effectively identify individuals with suspected IDs from free-text patient records, requiring only a single training example per classification label. The best-performing method based on the Flan-T5 model and incorporating the IE step achieved an F1-score of 0.867. Individuals classified as having ID by the best performing model were significantly enriched for de novo variants in known developmental disorder risk genes (odds ratio 29.1, 95% CI 7.36-107; P=2.1×10-5).

Conclusions: LLMs and in-context learning techniques combined with human-in-the-loop approaches can be highly beneficial for extraction and categorization of information from free-text clinical data. In this proof-of-concept study, we show that LLMs can be used to identify individuals with a severe mental illness who also have suspected ID, which is a biologically and clinically meaningful subgroup of patients.