Diagnosing and remediating harmful data shifts for the responsible deployment of clinical AI models
Bibliographic record
Abstract
Abstract Harmful data shifts occur when the distribution of data used to train a clinical AI system differs significantly from the distribution of data encountered during deployment, leading to erroneous predictions and potential harm to patients. We evaluated the impact of data shifts on an early warning system for in-hospital mortality that uses electronic health record data from patients admitted to a general internal medicine service, across 7 large hospitals in Toronto, Canada. We found model performance to differ across subgroups of clinical diagnoses, sex and age. To explore the robustness of the model, we evaluated potentially harmful data shifts across demographics, hospital types, seasons, time of hospital admission, and whether the patient was admitted from an acute care institution or nursing home, without relying on model performance. Interestingly, many of these harmful data shifts were unidirectional. We found models trained on community hospitals experience harmful data shifts when evaluated on academic hospitals, whereas models trained on academic hospitals transfer well to the community hospitals. To improve model performance across hospital sites we employed transfer learning, a strategy that stores knowledge gained from learning one domain and applies it to a different but related domain. We found hospital type-specific models that leverage transfer learning, perform better than models that use all available hospitals. Furthermore, we monitored data shifts over time and identified model deterioration during the COVID-19 pandemic. Typically, machine learning models remain locked after deployment, however, this can lead to model deterioration due to harmful data shifts that occur over time. We used continual learning, the process of learning from a continual stream of data in a sequential manner, to mitigate data shifts over time and improve model performance. Overall, our study is a crucial step towards the deployment of clinical AI models, by providing strategies and workflows to ensure the safety and efficacy of these models in real-world settings.
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 distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.010 | 0.009 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.004 | 0.009 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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 teacher head, 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".