Limitations of GPT-3.5 and GPT-4 in Applying Fleischner Society Guidelines to Incidental Lung Nodules
Bibliographic record
Abstract
Purpose: To evaluate the accuracy of GPT-3.5, GPT-4, and a fine-tuned GPT-3.5 model in applying Fleischner Society recommendations to lung nodules. Methods: We generated 10 lung nodule descriptions for each of the 12 nodule categories from the Fleischner Society guidelines, incorporating them into a single fictitious report (n = 120). GPT-3.5 and GPT-4 were prompted to make follow-up recommendations based on the reports. We then incorporated the full guidelines into the prompts and re-submitted them. Finally, we re-submitted the prompts to a fine-tuned GPT-3.5 model. Results were analyzed using binary accuracy analysis in R. Results: GPT-3.5 accuracy in applying Fleischner Society guidelines was 0.058 (95% CI: 0.02, 0.12). GPT-4 accuracy was improved at 0.15 (95% CI: 0.09, 0.23; P = .02 for accuracy comparison). In recommending PET-CT and/or biopsy, both GPT-3.5 and GPT-4 had an F-score of 0.00. After explicitly including the Fleischner Society guidelines in the prompt, GPT-3.5 and GPT-4 significantly improved their accuracy to 0.42 (95% CI: 0.33, 0.51; P < .001) and to 0.66 (95% CI: 0.57, 0.74; P < .001), respectively. GPT-4 remained significantly better than GPT-3.5 ( P < .001). The fine-tuned GPT-3.5 model accuracy was 0.46 (95% CI: 0.37, 0.55), not different from the GPT-3.5 model with guidelines included ( P = .53). Conclusion: GPT-3.5 and GPT-4 performed poorly in applying widely known guidelines and never correctly recommended biopsy. Flawed knowledge and reasoning both contributed to their poor performance. While GPT-4 was more accurate than GPT-3.5, its inaccuracy rate was unacceptable for clinical practice. These results underscore the limitations of large language models for knowledge and reasoning-based tasks.
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 | 0.051 | 0.257 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.003 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.003 | 0.003 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.002 |
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".