Integrating GPT-4o Into Data Mining in Neurosurgery: Feasibility and Proof-of-Concept Study
Bibliographic record
Abstract
Background: Large language models offer new possibilities for transforming unstructured clinical text into structured datasets. However, their performance in specialized and complex documentation environments, such as neurosurgery, remains insufficiently characterized. GPT-4o is a large language model with enhanced natural language capabilities, but its accuracy in extracting structured data from neurosurgical reports has not been systematically assessed. Objective: This proof-of-concept study evaluated the feasibility and accuracy of GPT-4o for extracting predefined structured variables from unstructured neurosurgical reports of patients with vestibular schwannoma. Specific aims were to measure accuracy across variable types, assess the impact of prompt refinement, and explore the model's potential utility for research-oriented data mining. Methods: In this retrospective single-center study, 10 consecutive patients with histologically confirmed vestibular schwannoma who underwent surgery between August and December 2023 were included. Four anonymized German-language documents per patient (discharge, surgical, histopathology, and 3-month follow-up reports) were processed using GPT-4o. Seventeen variables were extracted using a standardized zero-shot prompt. Targeted prompt refinements were subsequently applied for variables with low baseline accuracy. Two board-certified neurosurgeons independently validated all outputs, with discrepancies resolved by a senior neurosurgeon. Accuracy metrics, 95% CIs (Wilson method), and descriptive comparisons between variable types were calculated. Results: GPT-4o achieved 100% accuracy for structured variables requiring minimal interpretation, including patient ID, date of birth, date of surgery, histopathological diagnosis, and World Health Organization grade. Several interpretative variables, such as symptoms at presentation, symptom type, symptom duration, extent of resection, and permanence of postoperative deficits, were also extracted with 100% accuracy. In contrast, intraoperative complications and new postoperative deficits were correctly identified in only 50% (5/10) of cases using the zero-shot prompt. After targeted prompt refinement, accuracy for these variables improved substantially, reaching 90% to 100% in most cases. The mean accuracy was highest for structured categorical variables (97.5%, SD 4.6%), intermediate for binary variables (80%, SD 27.4%), and lowest for conditional text variables (66.7%, SD 28.9%), without statistically significant differences (P=.25). Conclusions: GPT-4o demonstrated strong feasibility for structured data extraction from standardized neurosurgical reports, particularly for variables with limited semantic complexity. However, the high accuracy observed reflects a narrow and highly controlled context and should not be interpreted as evidence of general reliability across diverse clinical settings. Larger, multi-institutional, and multilingual studies are needed to determine broader applicability and potential clinical integration.
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 | 0.012 | 0.023 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
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".