Improving Detection of Hospital Adverse Events Using Machine Learning on Real-World Narrative EMR Data
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
ObjectiveAdministrative data often underrepresents hospital adverse events (AEs) due to limitations in International Classification of Diseases (ICD) coding. By leveraging electronic medical records (EMRs), we aim to mitigate these discrepancies and enhance the precision of healthcare surveillance and performance evaluations. To this end, we have developed a machine learning (ML)-based approach that utilizes EMR text data to detect common AEs. MethodsWe sampled adult admissions from four Calgary hospitals (2017 - 2022). Registered nurses assessed charts for 17 AEs, and the results were used as reference standard. We compared two AE detection methods: the standard ICD-based method following Canadian guidelines, and our ML algorithm applied to EMR narratives. Sensitivity, positive predictive value (PPV), negative predictive value (NPV), and specificity for both methods were calculated and compared against the reference standard. ResultsWe analyzed 9,566 patients, of whom 1,506 were identified with AEs. Of the 17 AEs, the sensitivity in ICD-coded data ranged from 0-37%, and in EMRs, it was between 75-100%. Both showed low PPV (0-50% ICD vs.1-34% EMR). ICD data had high specificity ranging from 99-100% and NPV (99%-100%), while EMRs had specificity between 68-94% and an NPV of 100%. Conclusion and ImplicationsML significantly enhances sensitivity for AE detection compared to ICD-10-CA coding, despite both methods experiencing low PPV due to imbalances in EMR data. This marked improvement in sensitivity highlights ML's potential to transform AE surveillance and reporting, promising significant advancements in patient safety and healthcare quality by enabling more accurate and comprehensive identification of AEs.
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.
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 | 0.005 | 0.005 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.005 |
| Open science | 0.004 | 0.002 |
| Research integrity | 0.000 | 0.000 |
| 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 it