Weakly-supervised natural language processing with BERT-Clinical for automated lesion information extraction from free-text MRI reports in multiple sclerosis patients
Bibliographic record
Abstract
Purpose: To investigate how bidirectional encoder representations from transformers (BERT)-based models help extract treatment response information from free-text radiology reports. Materials and methods: This study involved 400 brain MRI reports from 115 participants with multiple sclerosis. New MRI lesion activity including new or enlarging T2 (newT2) and enhancing T1 (enhanceT1) lesions for assessing treatment responsiveness was identified using the named entity recognition technique along with BERT. Likewise, 2 other associated entities were also identified: the remaining brain MRI lesions (regT2), and lesion location. Report sentences containing any of the 4 entities were labeled for model development, totally 2568. Four recognized BERT models were investigated, each with conditional random field integrated for lesion versus location classification, trained using variable sample sizes (500–2000 sentences). Regularity was then applied for lesion subtyping. Model evaluation utilized a flexible F1 score, among others. Results: The Clinical-BERT performed the best. It achieved the best testing flexible F1 score of 0.721 in lesion and location classification, 0.741 in lesion only classification, and 0.771 in regT2 subtyping. With growing sample sizes, only Clinical-BERT performed increasingly better, which also had the best area under the curve of 0.741 in lesion classification at training using 2000 sentences. The PubMed-BERT achieved the best testing flexible F1 score of 0.857 in location only classification, and 0.846 and 0.657 in subtyping newT2 and enhanceT1, respectively. Conclusion: Based on a small sample size, our methods demonstrate the potential for extracting critical treatment-related information from free-text radiology reports, especially Clinical-BERT.
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 | 0.003 | 0.009 |
| 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 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| 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".