Author + information
- Julian Andres Betancura,b,
- Yuka Otakia,b,
- Matthews Fisha,b,
- Mark Lemleya,b,
- Damini Deya,b,
- Balaji Tamarappooa,b,
- Guido Germanoa,b,
- Daniel Bermana,b and
- Piotr Slomkaa,b
Background: We aimed to compare the prediction of major adverse cardiac events (MACE) by machine learning (ML) using clinical information and automatically derived image variables from stress and rest scans.
Methods: 2619 consecutive patients (48% male, 62±13 years), undergoing high speed SPECT, were monitored. A cohort of 2348 patients without past myocardial infarction (MI) or coronary artery bypass grafting (CABG) was also analyzed. MACE comprised all-cause mortality, non-fatal MI, unstable angina, or late (>90 days) coronary revascularization. ML involved automated feature selection by information gain ranking, Logitboost ensemble model building and 10-fold cross validation. Prediction of MACE by stress-only ML, stress/rest ML, expert summed stress/difference scores (SSS, SDS), and automatic stress/ischemic total perfusion deficits (TPD) was assessed by area under receiver operating characteristics curve (AUC).
Results: During follow-up (3.0±0.8 years), 320 patients (12%) had MACE. 235 patients (10%) had MACE in the cohort without past-MI/CABG. AUCs for prediction of MACE in the two populations by stress-only ML, stress TPD and SSS were similar or greater than for stress/rest ML, ischemic TPD and SDS, respectively (Figure). Stress-only ML prediction was significantly higher than for stress-rest visual score or TPD (p<0.001).
Conclusions: Adding rest scan information did not improve ML prediction of 3-year MACE when combined with clinical variables and quantitative stress analysis.
Poster Hall, Hall C
Saturday, March 18, 2017, 3:45 p.m.-4:30 p.m.
Session Title: Nuclear Cardiology: Quality
Abstract Category: 30. Non Invasive Imaging: Nuclear
Presentation Number: 1246-225
- 2017 American College of Cardiology Foundation