LLM in Medical Data Analysis Methodology (March 2024)

English summary

The presentation describes practical methods for using large language models as analytical tools in biomedical, clinical, and scientific research. The main applications are data cleaning, unstructured clinical text processing, hypothesis generation, natural-language analysis, and code generation with execution.

Key points

• LLMs can support manuscript drafting and editing, generate Python or R code, and assist with syntax and runtime error correction.
• Data-cleaning tasks include standardizing dates and units, correcting spelling and clinical terminology, detecting inconsistencies between clinical descriptions and measurements, and imputing missing values.
• Unstructured-data workflows include extracting drug names, doses, routes, frequencies, and durations; assigning ICD codes; categorizing literature and conference materials; conducting sentiment analysis; and summarizing histories for clinical-trial prescreening. In the presentation, the reported ChatGPT 3.5 ICD-coding result was approximately 70 percent when fully correct and partially correct cases were combined, consisting of 59 percent fully correct and 11 percent partially correct. A cited study reported that combining LLM-based initial annotation with expert refinement produced F1 scores comparable to expert-only annotation while reducing annotation time by more than 50 percent. These figures are results presented from cited studies, not independently verified here.

Audience

Medical researchers, clinical data scientists, bioinformaticians, and academic researchers developing computational data-analysis workflows

Scope and limitations

This summary is based only on the supplied evidence notes from the presentation dated March 27, 2024. The full study designs, data characteristics, comparison conditions, and statistical uncertainty of the cited studies were not supplied, so the reported performance figures cannot be independently verified from this material. LLM outputs require quantitative agreement assessment with human experts, bias evaluation, and separate validation. Important limitations include regulatory and security risks when identifiable or sensitive clinical data are sent to commercial cloud APIs, performance differences between local open-source and proprietary models, and reduced reproducibility caused by prompting, stochasticity, and changes to proprietary model versions.

Research context

The topic is thematically connected to medical artificial intelligence, clinical text processing, research data analysis, and reproducible computational workflows.