Automated extraction of key entities from non-english thorax CT reports using machine learning by large context, many-shot Generative AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial Intelligence (AI)-powered Named Entity Recognition (NER), a subfield of Natural Language Processing (NLP), can automatically analyze and extract relevant information from unstructured medical texts. However, available models are English-focused, with limited options for other languages. This study aims to identify an effective NER strategy for non-English medical reports by comparing a traditional transformer-based model against Large Language Models (LLMs) using different prompting techniques. This study performed a comparative analysis of four models for extracting key entities from anonymized Turkish Thorax Computed Tomography (CT) reports. We evaluated a spaCy-transformer model, trained with 100 reports as a baseline. This was compared against three LLM-based approaches using Google’s Gemini models: a many-shot ( n = 518 examples) prompt with Gemini 1.5 Pro, and both many-shot and five-shot prompts with Gemini 2.5 Pro. The many-shot prompts utilized a 64,000-token context. Performance was evaluated on a test set of 100 reports, focusing on five entities: anatomy (ANAT), impression (IMP), observation presence (OBS-P), absence (OBS-A), and uncertainty (OBS-U). The spaCy-transformer model achieved the highest performance in exact-match evaluation with a macro-averaged F1-score of 0.75. For relaxed-match evaluation, both the Gemini 2.5 Pro many-shot model and the spaCy-transformer achieved top-tier performance with an overall accuracy of 0.97 and a Cohen’s Kappa of 0.96. Critically, the many-shot approach with Gemini 2.5 Pro (macro F1: 0.92) significantly outperformed its five-shot counterpart (macro F1: 0.89), demonstrating the benefit of providing more examples. This study reveals that while trained transformers excel at precise boundary detection (exact match), LLMs guided by a many-shot strategy demonstrate excellent performance for relaxed-match recognition, which often carries more clinical relevance. Our results provide strong evidence that a many-shot learning approach is superior to a few-shot strategy for this task. While validated on Turkish reports, this methodology presents a promising and adaptable framework for developing high-accuracy NER tools in other languages where dedicated NLP resources are scarce.