Optimising large language models for clinical information extraction: a benchmarking study in the context of ulcerative colitis research
Bibliographic record
Abstract
Objective: Closed-source large language models (LLMs) like generative pre-trained transformer 4o (GPT-4o) have shown promise for clinical information extraction but are potentially limited by cost, data security concerns and inflexibility. Open-source models are an attractive alternative with various adaptation strategies with no consensus on best practices. This study aims to rigorously identify optimal adaptation strategies for open-source models and evaluate their performance relative to closed-source alternatives. Methods and Analysis: We studied three LLM adaptation strategies: chain-of-thought prompting, few-shot prompting and fine-tuning. Our target for information extraction was the Mayo Endoscopic Subscore (MES). We applied those strategies in all combinations to six open-source models (8-70 billion parameters) using an annotated set of colonoscopy procedure reports from the University of California, San Francisco (N=608) and San Francisco General Hospital (N=217). We analysed the relationship of these strategies to several performance metrics with a mixed-effects model, accounting for the variability between centres and LLMs. GPT-4o served as a closed-source oracle and provided in-depth commentary on the cost-effectiveness of these options. Results: < 0.001) the performance of open-source LLMs by 9.1-15.7 percentage points across accuracy, precision recall and annotation eligibility accuracy. However, GPT-4o with prompt engineering outperforms the best open-source model by 4.9%-11.2%. A simple cost-effectiveness analysis suggests that GPT-4o is more affordable compared with open-source alternatives. Conclusion: GPT-4o is currently the most efficient LLM for MES extraction. If unavailable, QLoRA-optimised open-source models are a competitive alternative. However, results also suggest that current instruction-following LLMs including GPT-4o do not fully follow user-provided instructions, leaving room for improvement. More work is needed to achieve consistent, near-perfect performance in clinical information extraction by LLMs.
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 distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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 teacher head, 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".