Clinical Trial Design Approach to Auditing Language Models in Health Care Setting
Bibliographic record
Abstract
PURPOSE: Rapid advancements in natural language processing have led to the development of sophisticated language models. Inspired by their success, these models are now used in health care for tasks such as clinical documentation and medical record classification. However, language models are prone to errors, which can have serious consequences in critical domains such as health care, ensuring that their reliability is essential to maintain patient safety and data integrity. METHODS: To address this, we propose an innovative auditing process based on principles from clinical trial design. Our approach involves subject matter experts (SMEs) manually reviewing pathology reports without previous knowledge of the model's classification. This single-blind setup minimizes bias and allows us to apply statistical rigor to assess model performance. RESULTS: Deployed at the British Columbia Cancer Registry, our audit process effectively identified the core issues in the operational models. Early interventions addressed these issues, maintaining data integrity and patient care standards. CONCLUSION: The audit provides real-world performance metrics and underscores the importance of human-in-the-loop machine learning. Even advanced models require SME oversight to ensure accuracy and reliability. To our knowledge, we have developed the first continuous audit process for language models in health care, modeled after clinical trial principles. This methodology ensures that audits are statistically sound and operationally feasible, setting a new standard for evaluating language models in critical applications.
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 | 0.455 | 0.557 |
| Meta-epidemiology (narrow) | 0.002 | 0.002 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.004 | 0.003 |
| Science and technology studies | 0.002 | 0.005 |
| Scholarly communication | 0.006 | 0.005 |
| Open science | 0.005 | 0.006 |
| Research integrity | 0.004 | 0.006 |
| Insufficient payload (model declined to judge) | 0.010 | 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; the direct Gemma label and the distilled Codex classifier agree on what is shown here.
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".