Abstract B018: Practical benchmarking of large language models for structuring synthetic prostate cancer histopathology statements in English and Finnish
Bibliographic record
Abstract
Abstract Large Language Models (LLMs) are increasingly applied to clinical oncology tasks, such as structuring free-text format histopathology statements. Given the rapid pace of LLM development, guidance is needed for optimal model, parameter, and prompt design choices, especially in minority languages. A total of 100 novel synthetic histopathological prostate cancer (PCa) statements were manually generated to mimic real-life data of Helsinki University Hospital. First, 25 statements were generated in Finnish and manually translated to English. Manually censored versions were then generated, with key information replaced by filler text to test hallucinations. The following models were evaluated: Claude (3 variants of 3.5 Haiku/Sonnet and 3.7 Sonnet-20250219), DeepSeek (R1), Gemini (1.5 Pro with 2 snapshots, 2.0 flash-001 and flash-lite-001), GPT (4o with 3 snapshots and 4.1 normal, mini and nano), Grok (3-beta and 2-1212), Llama 4 (Maverick and Scout), and Mistral (large-2411). Each model was tested with zero-shot (no examples), one-shot (one example in prompt), and few-shot prompts (one positive and one negative example), and queried for three structured elements in JSON format: sample type (TURP/RP/Biopsy), malignancy status (Yes/No), and Gleason score (major+minor=sum). Missing values were requested as NA. All queries were replicated thrice with temperature-parameter 0. F-score (F) was used to assess model performance. Python v3.12 was used to construct experimental data via APIs followed by statistical analysis in R v4.4. All data and code is made available via GitHub. All models benefited from the few-shot prompt design, thus results are here summarized for the few-shot prompt. Median (IQR) F was 0.76 (0.70-0.86) in English and 0.75 (0.69-0.80) in Finnish across all models. Top proprietary model was GPT-4o, with all snapshots within F 0.96-0.98 in English and F 0.92-0.94 in Finnish. Top open weight model was Mistral with F 0.87 / 0.75 in English / Finnish, followed by DeepSeek’s R1 with F 0.80 / 0.75. Most notable underperformer was GPT-4.1 nano with F 0.47 / 0.37 in English / Finnish. Gemini 1.5 Pro snapshot 001 vs 002 had notable variance with F 0.71 / 0.69 vs F 0.62 / 0.64 similarly as Claude 3.5 Sonnet 20240620 vs. 20241022 with F 0.66 / 0.69 vs F 0.77 / 0.74 in English / Finnish. Our results from PCa histopathology highlight that current state-of-the-art LLMs are agnostic to extracting data from a minority language (Finnish) in comparison with the de facto largest corpus training language (English). Few-shot prompts systematically performed best, highlighting the importance of prompt engineering. While top-performer emerged from the GPT-family, open weight models such as Mistral and DeepSeek-R1 hold potential, especially since they can be locally deployed to sensitive data settings without internet access. Further, researchers ought to carefully choose their model, as light-weight models may result in poor performance (GPT-4.1 nano) and performance can vary notably between time-dependent snapshots (Gemini 1.5 Pro and Claude 3.5 Sonnet). Citation Format: Tolou Shadbahr, Antti S. Rannikko, Tuomas Mirtti, Teemu D. Laajala. Practical benchmarking of large language models for structuring synthetic prostate cancer histopathology statements in English and Finnish [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B018.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.004 | 0.018 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.002 | 0.001 |
| Insufficient payload (model declined to judge) | 0.006 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".