Information Extraction of Doctoral Theses Using Two Different Large Language Models vs Health Services Researchers: Development and Usability Study
Bibliographic record
Abstract
Background: The Archive of German-Language General Practice (ADAM) stores about 500 paper-based doctoral theses published from 1965 to today. Although they have been grouped in different categories, no deeper systematic process of information extraction (IE) has been performed yet. Recently developed large language models (LLMs) like ChatGPT have been attributed the potential to help in the IE of medical documents. However, there are concerns about LLM hallucinations. Furthermore, there have not been reports regarding their usage in nonrecent doctoral theses yet. Objective: The aim of this study is to analyze if LLMs can help to extract information from doctoral theses by using GPT-4o and Gemini-1.5-Flash for paper-based doctoral theses in ADAM. Methods: We randomly selected 10 doctoral theses published between 1965 and 2022. After preprocessing, we used two different LLM pipelines, using models by OpenAI and Google. Pipelines were used to extract dissertation characteristics and generate uniform abstracts. Furthermore, one pooled human-generated abstract was written for comparison. Furthermore, blinded raters were asked to evaluate LLM-generated abstracts in comparison to the human-generated ones. Bidirectional encoder representations from transformers scores were calculated as the evaluation metric. Results: Relevant dissertation characteristics and keywords could be extracted for all theses (n=10): institute name and location, thesis title, author name(s), and publication year. For all except one doctoral thesis, an abstract could be generated using GPT-4o, while Gemini-1.5-Flash provided abstracts in all cases (n=10). The modality of abstract generation showed no influence in raters' evaluation using the nonparametric Kruskal-Wallis test for independent groups (P=.44). The creation of LLM-generated abstracts was estimated to be 24-36 times faster than creation by humans. Evaluation metrics showed moderate-to-high semantic similarity (mean bidirectional encoder representations from transformers F1-score, GPT-4o: 0.72 and Gemini: 0.71). Translation from German into English did not result in a loss of information (n=10). Conclusions: An accumulating body of unpublished doctoral theses makes it difficult to extract relevant evidence. Recent advances in LLMs like ChatGPT have raised expectations in text mining, but they have not yet been used in the IE of "historic" medical documents. This feasibility study suggests that both models (GPT-4o and Gemini-1.5-Flash) helped to accurately simplify and condense doctoral theses into relevant information, while LLM-generated abstracts were perceived as similar to human-generated ones, were semanticly similar, and took about 30 times less time to create. This pilot study demonstrates the feasibility of a regular office-scanning workflow and use of general-purpose LLMs to extract relevant information and produce accurate abstracts from ADAM doctoral theses. Taken together, this information could help researchers to better search the family medicine scientific literature over the last 60 years, helping to develop current research questions.
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 | 0.025 | 0.091 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.004 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.003 | 0.004 |
| Open science | 0.001 | 0.003 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.003 | 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".