Determining a minimum set of variables for machine learning cardiovascular event prediction: results from REFINE SPECT registry
Problem
Manual collection of clinical variables for machine learning prediction of major adverse cardiac events from SPECT myocardial perfusion imaging is time-consuming and error-prone.
Why it matters
Faster, simpler risk stratification could improve care for patients undergoing cardiac imaging.
What the AI did
ML models were trained on 20,414 registry patients and incrementally pruned by variable importance to find the fewest manually input and imaging variables preserving prognostic accuracy, with external testing on 2,984 patients.
Results
not stated in source
Impact
Multicentre data with an external test cohort described; exact accuracy metrics not in abstract.
Record
| Venue | Cardiovascular Research |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Richard Ríos, Robert J.H. Miller, Lien-Hsin Hu, Yuka Otaki, Ananya Singh, Márcio A. Diniz, Tali Sharir, Andrew Jeffrey Einstein et al. |
| Citations | 42 (OpenAlex, at import time) |
| References | 23 |
| Model or method | not stated in source |
| Organization | REFINE SPECT registry; University of Calgary (external test) |
| Primary source | https://doi.org/10.1093/cvr/cvab236 |
| Open access | green |
| Keywords | cardiovascular risk, machine learning, SPECT, MACE, feature selection |
Importance rating
significant — thousands to millions affected
Multicentre study on 20,414 registry patients plus 2,984 external tests; simpler MACE risk stratification could benefit many cardiac imaging patients.