Mitigating bias in deep learning for diagnosis of coronary artery disease from myocardial perfusion SPECT images.

PURPOSE: Artificial intelligence (AI) has high diagnostic accuracy for coronary artery disease (CAD) from myocardial perfusion imaging (MPI). However, when trained using high-risk populations (such as patients with correlating invasive testing), the disease probability can be overestimated due to selection bias. We evaluated different strategies for training AI models to improve the calibration (accurate estimate of disease probability), using external testing. METHODS: Deep learning was trained using 828 patients from 3 sites, with MPI and invasive angiography within 6 months. Perfusion was assessed using upright (U-TPD) and supine total perfusion deficit (S-TPD). AI training without data augmentation (model 1) was compared to training with augmentation (increased sampling) of patients without obstructive CAD (model 2), and patients without CAD and TPD < 2% (model 3). All models were tested in an external population of patients with invasive angiography within 6 months (n = 332) or low likelihood of CAD (n = 179). RESULTS: Model 3 achieved the best calibration (Brier score 0.104 vs 0.121, p < 0.01). Improvement in calibration was particularly evident in women (Brier score 0.084 vs 0.124, p < 0.01). In external testing (n = 511), the area under the receiver operating characteristic curve (AUC) was higher for model 3 (0.930), compared to U-TPD (AUC 0.897) and S-TPD (AUC 0.900, p < 0.01 for both). CONCLUSION: Training AI models with augmentation of low-risk patients can improve calibration of AI models developed to identify patients with CAD, allowing more accurate assignment of disease probability. This is particularly important in lower-risk populations and in women, where overestimation of disease probability could significantly influence down-stream patient management.

Unsupervised learning to characterize patients with known coronary artery disease undergoing myocardial perfusion imaging.

PURPOSE: Patients with known coronary artery disease (CAD) comprise a heterogenous population with varied clinical and imaging characteristics. Unsupervised machine learning can identify new risk phenotypes in an unbiased fashion. We use cluster analysis to risk-stratify patients with known CAD undergoing single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI). METHODS: From 37,298 patients in the REFINE SPECT registry, we identified 9221 patients with known coronary artery disease. Unsupervised machine learning was performed using clinical (23), acquisition (17), and image analysis (24) parameters from 4774 patients (internal cohort) and validated with 4447 patients (external cohort). Risk stratification for all-cause mortality was compared to stress total perfusion deficit (< 5%, 5-10%, ≥10%). RESULTS: Three clusters were identified, with patients in Cluster 3 having a higher body mass index, more diabetes mellitus and hypertension, and less likely to be male, have dyslipidemia, or undergo exercise stress imaging (p < 0.001 for all). In the external cohort, during median follow-up of 2.6 [0.14, 3.3] years, all-cause mortality occurred in 312 patients (7%). Cluster analysis provided better risk stratification for all-cause mortality (Cluster 3: hazard ratio (HR) 5.9, 95% confidence interval (CI) 4.0, 8.6, p < 0.001; Cluster 2: HR 3.3, 95% CI 2.5, 4.5, p < 0.001; Cluster 1, reference) compared to stress total perfusion deficit (≥10%: HR 1.9, 95% CI 1.5, 2.5 p < 0.001; < 5%: reference). CONCLUSIONS: Our unsupervised cluster analysis in patients with known CAD undergoing SPECT MPI identified three distinct phenotypic clusters and predicted all-cause mortality better than ischemia alone.

Prognostic value of early left ventricular ejection fraction reserve during regadenoson stress solid-state SPECT-MPI.

BACKGROUND: We hypothesized early post-stress left ventricular ejection fraction reserve (EFR) on solid-state-SPECT is associated with major cardiac adverse events (MACE). METHODS: 151 patients (70 ± 12 years, male 50%) undergoing same-day rest/regadenoson stress 99mTc-sestamibi solid-state SPECT were followed for MACE. Rest imaging was performed in the upright and supine positions. Early stress imaging was started 2 minutes after the regadenoson injection in the supine position and followed by late stress acquisition in the upright position. Total perfusion deficit (TPD) and functional parameters were quantified automatically. EFR, ∆end-diastolic volume (EDV), and end-systolic volume (ESV) were calculated as the difference between stress and rest values in the same position. EFR < 0%, ∆EDV ≥ 5 ml, or ∆ESV ≥ 5 ml was defined as abnormal. RESULTS: During the follow-up (mean 3.2 years), 28 MACE occurred (19%). In Kaplan-Meier analysis, there was a significantly decreased event-free survival in patients with early EFR < 0% (P = 0.004). Similarly, there was a decreased event-free survival in patients with ∆ESV ≥ 5 ml at early stress (P = 0.003). However, EFR, ∆EDV, and ∆ESV at late stress were not associated with MACE-free survival. Cox proportional hazards model adjusting for clinical information and stress TPD demonstrated that EFR, ∆EDV, and ∆ESV at early stress were significantly associated with MACE (P < 0.05 for all). CONCLUSIONS: Reduced early post-stress EFR on vasodilator stress solid-state SPECT is associated with MACE.

Machine learning to predict abnormal myocardial perfusion from pre-test features.

BACKGROUND: Accurately predicting which patients will have abnormal perfusion on MPI based on pre-test clinical information may help physicians make test selection decisions. We developed and validated a machine learning (ML) model for predicting abnormal perfusion using pre-test features. METHODS: We included consecutive patients who underwent SPECT MPI, with 20,418 patients from a multi-center (5 sites) international registry in the training population and 9019 patients (from 2 separate sites) in the external testing population. The ML (extreme gradient boosting) model utilized 30 pre-test features to predict the presence of abnormal myocardial perfusion by expert visual interpretation. RESULTS: In external testing, the ML model had higher prediction performance for abnormal perfusion (area under receiver-operating characteristic curve [AUC] 0.762, 95% CI 0.750-0.774) compared to the clinical CAD consortium (AUC 0.689) basic CAD consortium (AUC 0.657), and updated Diamond-Forrester models (AUC 0.658, p < 0.001 for all). Calibration (validation of the continuous risk prediction) was superior for the ML model (Brier score 0.149) compared to the other models (Brier score 0.165 to 0.198, all p < 0.001). CONCLUSION: ML can predict abnormal myocardial perfusion using readily available pre-test information. This model could be used to help guide physician decisions regarding non-invasive test selection.

Prevalence and predictors of automatically quantified myocardial ischemia within a multicenter international registry.

BACKGROUND: The utility of cardiac stress testing depends on the prevalence of myocardial ischemia within candidate populations. However, a comprehensive assessment of the factors influencing frequency of myocardial ischemia within contemporary populations referred for stress testing has not been performed. METHODS: We assessed 19,690 patients undergoing nuclear stress testing from a multicenter registry. The chi-square test was used to assess the relative importance of features for predicting myocardial ischemia. RESULTS: In the overall cohort, LVEF, male gender, and rest total perfusion deficit (TPD) were the top three predictors of ischemia, followed by CAD status, age, typical angina, and CAD risk factors. Myocardial ischemia was observed in 13.6 % of patients with LVEF > 55 %, in 26.2 % of patients with LVEF 45 %-54 %, and in 48.3% among patients with LVEF < 45 % (P < 0.001). A similar pattern was noted for rest TPD (P < 0.001). Men had a threefold higher frequency of ischemia versus women (25.8 % vs. 8.4%, P < 0.001). Although the relative ranking of ischemia predictors varied among centers, LVEF and/or rest TPD were among the two most potent predictors of myocardial ischemia within each center. CONCLUSION: The prevalence of myocardial ischemia varied markedly according to clinical and imaging characteristics. LVEF and rest TPD are robust predictors of myocardial ischemia.

Automated quantitative analysis of CZT SPECT stratifies cardiovascular risk in the obese population: Analysis of the REFINE SPECT registry.

BACKGROUND: Obese patients constitute a substantial proportion of patients referred for SPECT myocardial perfusion imaging (MPI), presenting a challenge of increased soft tissue attenuation. We investigated whether automated quantitative perfusion analysis can stratify risk among different obesity categories and whether two-view acquisition adds to prognostic assessment. METHODS: Participants were categorized according to body mass index (BMI). SPECT MPI was assessed visually and quantified automatically; combined total perfusion deficit (TPD) was evaluated. Kaplan-Meier and Cox proportional hazard analyses were used to assess major adverse cardiac event (MACE) risk. Prognostic accuracy for MACE was also compared. RESULTS: Patients were classified according to BMI: BMI < 30, 30 ≤ BMI < 35, BMI ≥ 35. In adjusted analysis, each category of increasing stress TPD was associated with increased MACE risk, except for 1% ≤ TPD < 5% and 5% ≤ TPD < 10% in patients with BMI ≥ 35. Compared to visual analysis, single-position stress TPD had higher prognostic accuracy in patients with BMI < 30 (AUC .652 vs .631, P < .001) and 30 ≤ BMI < 35 (AUC .660 vs .636, P = .027). Combined TPD had better discrimination than visual analysis in patients with BMI ≥ 35 (AUC .662 vs .615, P = .003). CONCLUSIONS: Automated quantitative methods for SPECT MPI interpretation provide robust risk stratification in the obese population. Combined stress TPD provides additional prognostic accuracy in patients with more significant obesity.

Diagnostic safety of a machine learning-based automatic patient selection algorithm for stress-only myocardial perfusion SPECT.

BACKGROUND: Stress-only myocardial perfusion imaging (MPI) markedly reduces radiation dose, scanning time, and cost. We developed an automated clinical algorithm to safely cancel unnecessary rest imaging with high sensitivity for obstructive coronary artery disease (CAD). METHODS AND RESULTS: Patients without known CAD undergoing both MPI and invasive coronary angiography from REFINE SPECT were studied. A machine learning score (MLS) for prediction of obstructive CAD was generated using stress-only MPI and pre-test clinical variables. An MLS threshold with a pre-defined sensitivity of 95% was applied to the automated patient selection algorithm. Obstructive CAD was present in 1309/2079 (63%) patients. MLS had higher area under the receiver operator characteristic curve (AUC) for prediction of CAD than reader diagnosis and TPD (0.84 vs 0.70 vs 0.78, P < .01). An MLS threshold of 0.29 had superior sensitivity than reader diagnosis and TPD for obstructive CAD (95% vs 87% vs 87%, P < .01) and high-risk CAD, defined as stenosis of the left main, proximal left anterior descending, or triple-vessel CAD (sensitivity 96% vs 89% vs 90%, P < .01). CONCLUSIONS: The MLS is highly sensitive for prediction of both obstructive and high-risk CAD from stress-only MPI and can be applied to a stress-first protocol for automatic cancellation of unnecessary rest imaging.

Comparison of diabetes to other prognostic predictors among patients referred for cardiac stress testing: A contemporary analysis from the REFINE SPECT Registry.

BACKGROUND: Diabetes mellitus (DM) is increasingly prevalent among contemporary populations referred for cardiac stress testing, but its potency as a predictor for major adverse cardiovascular events (MACE) vs other clinical variables is not well delineated. METHODS AND RESULTS: From 19,658 patients who underwent SPECT-MPI, we identified 3122 patients with DM without known coronary artery disease (CAD) (DM+/CAD-) and 3564 without DM with known CAD (DM-/CAD+). Propensity score matching was used to control for the differences in characteristics between DM+/CAD- and DM-/CAD+ groups. There was comparable MACE in the matched DM+/CAD- and DM-/CAD+ groups (HR 1.15, 95% CI 0.97-1.37). By Chi-square analysis, type of stress (exercise or pharmacologic), total perfusion deficit (TPD), and left ventricular function were the most potent predictors of MACE, followed by CAD and DM status. The combined consideration of mode of stress, TPD, and DM provided synergistic stratification, an 8.87-fold (HR 8.87, 95% CI 7.27-10.82) increase in MACE among pharmacologically stressed patients with DM and TPD > 10% (vs non-ischemic, exercised stressed patients without DM). CONCLUSIONS: Propensity-matched patients with DM and no known CAD have similar MACE risk compared to patients with known CAD and no DM. DM is synergistic with mode of stress testing and TPD in predicting the risk of cardiac stress test patients.

Rationale and design of the REgistry of Fast Myocardial Perfusion Imaging with NExt generation SPECT (REFINE SPECT).

BACKGROUND: We aim to establish a multicenter registry collecting clinical, imaging, and follow-up data for patients who undergo myocardial perfusion imaging (MPI) with the latest generation SPECT scanners. METHODS: REFINE SPECT (REgistry of Fast Myocardial Perfusion Imaging with NExt generation SPECT) uses a collaborative design with multicenter contribution of clinical data and images into a comprehensive clinical-imaging database. All images are processed by quantitative software. Over 290 individual imaging variables are automatically extracted from each image dataset and merged with clinical variables. In the prognostic cohort, patient follow-up is performed for major adverse cardiac events. In the diagnostic cohort (patients with correlating invasive angiography), angiography and revascularization results within 6 months are obtained. RESULTS: To date, collected prognostic data include scans from 20,418 patients in 5 centers (57% male, 64.0 ± 12.1 years) who underwent exercise (48%) or pharmacologic stress (52%). Diagnostic data include 2079 patients in 9 centers (67% male, 64.7 ± 11.2 years) who underwent exercise (39%) or pharmacologic stress (61%). CONCLUSION: The REFINE SPECT registry will provide a resource for collaborative projects related to the latest generation SPECT-MPI. It will aid in the development of new artificial intelligence tools for automated diagnosis and prediction of prognostic outcomes.

Upper reference limits of transient ischemic dilation ratio for different protocols on new-generation cadmium zinc telluride cameras: A report from REFINE SPECT registry.

BACKGROUND: Upper reference limits for transient ischemic dilation (TID) have not been rigorously established for cadmium-zinc-telluride (CZT) camera systems. We aimed to derive TID limits for common myocardial perfusion imaging protocols utilizing a large, multicenter registry (REFINE SPECT). METHODS: One thousand six hundred and seventy-two patients with low likelihood of coronary artery disease with normal perfusion findings were identified. Images were processed with Quantitative Perfusion SPECT software (Cedars-Sinai Medical Center, Los Angeles, CA). Non-attenuation-corrected, camera-, radiotracer-, and stress protocol-specific TID limits in supine position were derived from 97.5th percentile and mean + 2 standard deviations (SD). Reference limits were compared for different solid-state cameras (D-SPECT vs. Discovery), radiotracers (technetium-99m-sestamibi vs. tetrofosmin), different types of stress (exercise vs. four different vasodilator-based protocols), and different vasodilator-based protocols. RESULTS: TID measurements did not follow Gaussian distribution in six out of eight subgroups. TID limits ranged from 1.18 to 1.52 (97.5th percentile) and 1.18 to 1.39 (mean + 2SD). No difference was noted between D-SPECT and Discovery cameras (P = 0.71) while differences between exercise and vasodilator-based protocols (adenosine, regadenoson, or regadenoson-walk) were noted (all P < 0.05). CONCLUSIONS: We used a multicenter registry to establish camera-, radiotracer-, and protocol-specific upper reference limits of TID for supine position on CZT camera systems. Reference limits did not differ between D-SPECT and Discovery camera.

"Same-Patient Processing" for multiple cardiac SPECT studies. 1. Improving LV segmentation accuracy.

OBJECTIVES: This paper describes a novel approach (same-patient processing, or SPP) aimed at improving left ventricular segmentation accuracy in patients with multiple SPECT studies, and evaluates its performance compared to conventional processing in a large population of 962 patients undergoing rest and stress electrocardiography-gated SPECT MPI, for a total of 5,772 image datasets (6 per patient). METHODS: Each dataset was independently processed using a standard algorithm, and a shape quality control score (SQC) was produced for every segmentation. Datasets with a SQC score higher than a specific threshold, suggesting algorithmic failure, were automatically reprocessed with the SPP-modified algorithm, which incorporates knowledge of the segmentation mask location in the other datasets belonging to the same patient. Experienced operators blinded as to whether datasets had been processed based on the standard or SPP approach assessed segmentation success/failure for each dataset. RESULTS: The SPP approach reduced segmentation failures from 219/5772 (3.8%) to 42/5772 (0.7%) overall, with particular improvements in attenuation corrected (AC) datasets with high extra-cardiac activity (from 100/962 (10.4%) to 12/962 (1.4%) for rest AC, and from 41/962 (4.3%) to 9/962 (0.9%) for stress AC). The number of patients who had at least one of their 6 datasets affected by segmentation failure decreased from 141/962 (14.7%) to 14/962 (1.7%) using the SPP approach. CONCLUSION: Whenever multiple image datasets for the same patient exist and need to be processed, it is possible to deal with the images as a group rather than individually. The same-patient processing approach can be implemented automatically, and may substantially reduce the need for manual reprocessing due to cardiac segmentation failure.

Clinical phenotypes among patients with normal cardiac perfusion using unsupervised learning: a retrospective observational study.

BACKGROUND: Myocardial perfusion imaging (MPI) is one of the most common cardiac scans and is used for diagnosis of coronary artery disease and assessment of cardiovascular risk. However, the large majority of MPI patients have normal results. We evaluated whether unsupervised machine learning could identify unique phenotypes among patients with normal scans and whether those phenotypes were associated with risk of death or myocardial infarction. METHODS: Patients from a large international multicenter MPI registry (10 sites) with normal perfusion by expert visual interpretation were included in this cohort analysis. The training population included 9849 patients, and external testing population 12,528 patients. Unsupervised cluster analysis was performed, with separate training and external testing cohorts, to identify clusters, with four distinct phenotypes. We evaluated the clinical and imaging features of clusters and their associations with death or myocardial infarction. FINDINGS: Patients in Clusters 1 and 2 almost exclusively underwent exercise stress, while patients in Clusters 3 and 4 mostly required pharmacologic stress. In external testing, the risk for Cluster 4 patients (20.2% of population, unadjusted hazard ratio [HR] 6.17, 95% confidence interval [CI] 4.64-8.20) was higher than the risk associated with pharmacologic stress (HR 3.03, 95% CI 2.53-3.63), or previous myocardial infarction (HR 1.82, 95% CI 1.40-2.36). INTERPRETATION: Unsupervised learning identified four distinct phenotypes of patients with normal perfusion scans, with a significant proportion of patients at very high risk of myocardial infarction or death. Our results suggest a potential role for patient phenotyping to improve risk stratification of patients with normal imaging results. FUNDING: This work was supported by the National Heart, Lung, and Blood Institute at the National Institutes of Health [R35HL161195 to PS]. The REFINE SPECT database was supported by the National Heart, Lung, and Blood Institute at the National Institutes of Health [R01HL089765 to PS]. MCW was supported by the British Heart Foundation [FS/ICRF/20/26002].

Direct Risk Assessment From Myocardial Perfusion Imaging Using Explainable Deep Learning.

BACKGROUND: Myocardial perfusion imaging (MPI) is frequently used to provide risk stratification, but methods to improve the accuracy of these predictions are needed. OBJECTIVES: The authors developed an explainable deep learning (DL) model (HARD MACE [major adverse cardiac events]-DL) for the prediction of death or nonfatal myocardial infarction (MI) and validated its performance in large internal and external testing groups. METHODS: Patients undergoing single-photon emission computed tomography MPI were included, with 20,401 patients in the training and internal testing group (5 sites) and 9,019 in the external testing group (2 different sites). HARD MACE-DL uses myocardial perfusion, motion, thickening, and phase polar maps combined with age, sex, and cardiac volumes. The primary outcome was all-cause mortality or nonfatal MI. Prognostic accuracy was evaluated using area under the receiver-operating characteristic curve (AUC). RESULTS: During internal testing, patients with normal perfusion and elevated HARD MACE-DL risk were at higher risk than patients with abnormal perfusion and low HARD MACE-DL risk (annualized event rate, 2.9% vs 1.2%; P < 0.001). Patients in the highest quartile of HARD MACE-DL score had an annual rate of death or MI (4.8%) 10-fold higher than patients in the lowest quartile (0.48% per year). In external testing, the AUC for HARD MACE-DL (0.73; 95% CI: 0.71-0.75) was higher than a logistic regression model (AUC: 0.70), stress total perfusion deficit (TPD) (AUC: 0.65), and ischemic TPD (AUC: 0.63; all P < 0.01). Calibration, a measure of how well predicted risk matches actual risk, was excellent in both groups (Brier score, 0.079 for internal and 0.070 for external). CONCLUSIONS: The DL model predicts death or MI directly from MPI, by estimating patient-level risk with good calibration and improved accuracy compared with traditional quantitative approaches. The model incorporates mechanisms to explain to the physician which image regions contribute to the adverse event prediction.

Time and event-specific deep learning for personalized risk assessment after cardiac perfusion imaging.

Standard clinical interpretation of myocardial perfusion imaging (MPI) has proven prognostic value for predicting major adverse cardiovascular events (MACE). However, personalizing predictions to a specific event type and time interval is more challenging. We demonstrate an explainable deep learning model that predicts the time-specific risk separately for all-cause death, acute coronary syndrome (ACS), and revascularization directly from MPI and 15 clinical features. We train and test the model internally using 10-fold hold-out cross-validation (n = 20,418) and externally validate it in three separate sites (n = 13,988) with MACE follow-ups for a median of 3.1 years (interquartile range [IQR]: 1.6, 3.6). We evaluate the model using the cumulative dynamic area under receiver operating curve (cAUC). The best model performance in the external cohort is observed for short-term prediction - in the first six months after the scan, mean cAUC for ACS and all-cause death reaches 0.76 (95% confidence interval [CI]: 0.75, 0.77) and 0.78 (95% CI: 0.78, 0.79), respectively. The model outperforms conventional perfusion abnormality measures at all time points for the prediction of death in both internal and external validations, with improvement increasing gradually over time. Individualized patient explanations are visualized using waterfall plots, which highlight the contribution degree and direction for each feature. This approach allows the derivation of individual event probability as a function of time as well as patient- and event-specific risk explanations that may help draw attention to modifiable risk factors. Such a method could help present post-scan risk assessments to the patient and foster shared decision-making.

Explainable Deep Learning Improves Physician Interpretation of Myocardial Perfusion Imaging.

Artificial intelligence may improve accuracy of myocardial perfusion imaging (MPI) but will likely be implemented as an aid to physician interpretation rather than an autonomous tool. Deep learning (DL) has high standalone diagnostic accuracy for obstructive coronary artery disease (CAD), but its influence on physician interpretation is unknown. We assessed whether access to explainable DL predictions improves physician interpretation of MPI. Methods: We selected a representative cohort of patients who underwent MPI with reference invasive coronary angiography. Obstructive CAD, defined as stenosis ≥50% in the left main artery or ≥70% in other coronary segments, was present in half of the patients. We used an explainable DL model (CAD-DL), which was previously developed in a separate population from different sites. Three physicians interpreted studies first with clinical history, stress, and quantitative perfusion, then with all the data plus the DL results. Diagnostic accuracy was assessed using area under the receiver-operating-characteristic curve (AUC). Results: In total, 240 patients with a median age of 65 y (interquartile range 58-73) were included. The diagnostic accuracy of physician interpretation with CAD-DL (AUC 0.779) was significantly higher than that of physician interpretation without CAD-DL (AUC 0.747, P = 0.003) and stress total perfusion deficit (AUC 0.718, P < 0.001). With matched specificity, CAD-DL had higher sensitivity when operating autonomously compared with readers without DL results (P < 0.001), but not compared with readers interpreting with DL results (P = 0.122). All readers had numerically higher accuracy with CAD-DL, with AUC improvement 0.02-0.05, and interpretation with DL resulted in overall net reclassification improvement of 17.2% (95% CI 9.2%-24.4%, P < 0.001). Conclusion: Explainable DL predictions lead to meaningful improvements in physician interpretation; however, the improvement varied across the readers, reflecting the acceptance of this new technology. This technique could be implemented as an aid to physician diagnosis, improving the diagnostic accuracy of MPI.

Differences in Prognostic Value of Myocardial Perfusion Single-Photon Emission Computed Tomography Using High-Efficiency Solid-State Detector Between Men and Women in a Large International Multicenter Study.

BACKGROUND: Semiquantitative assessment of stress myocardial perfusion defect has been shown to have greater prognostic value for prediction of major adverse cardiac events (MACE) in women compared with men in single-center studies with conventional single-photon emission computed tomography (SPECT) cameras. We evaluated sex-specific difference in the prognostic value of automated quantification of ischemic total perfusion defect (ITPD) and the interaction between sex and ITPD using high-efficiency SPECT cameras with solid-state detectors in an international multicenter imaging registry (REFINE SPECT [Registry of Fast Myocardial Perfusion Imaging With Next-Generation SPECT]). METHODS: Rest and exercise or pharmacological stress SPECT myocardial perfusion imaging were performed in 17 833 patients from 5 centers. MACE was defined as the first occurrence of death or myocardial infarction. Total perfusion defect (TPD) at rest, stress, and ejection fraction were quantified automatically by software. ITPD was given by stressTPD-restTPD. Cox proportional hazards model was used to evaluate the association between ITPD versus MACE-free survival and expressed as a hazard ratio. RESULTS: In 10614 men and 7219 women, with a median follow-up of 4.75 years (interquartile range, 3.7-6.1), there were 1709 MACE. In a multivariable Cox model, after adjusting for revascularization and other confounding variables, ITPD was associated with MACE (hazard ratio, 1.08 [95% CI, 1.05-1.1]; P<0.001). There was an interaction between ITPD and sex (P<0.001); predicted survival for ITPD<5% was worse among men compared to women, whereas survival among women was worse than men for ITPD≥5%, P<0.001. CONCLUSIONS: In the international, multicenter REFINE SPECT registry, moderate and severe ischemia as quantified by ITPD from high-efficiency SPECT is associated with a worse prognosis in women compared with men.

Handling missing values in machine learning to predict patient-specific risk of adverse cardiac events: Insights from REFINE SPECT registry.

BACKGROUND: Machine learning (ML) models can improve prediction of major adverse cardiovascular events (MACE), but in clinical practice some values may be missing. We evaluated the influence of missing values in ML models for patient-specific prediction of MACE risk. METHODS: We included 20,179 patients from the multicenter REFINE SPECT registry with MACE follow-up data. We evaluated seven methods for handling missing values: 1) removal of variables with missing values (ML-Remove), 2) imputation with median and unique category for continuous and categorical variables, respectively (ML-Traditional), 3) unique category for missing variables (ML-Unique), 4) cluster-based imputation (ML-Cluster), 5) regression-based imputation (ML-Regression), 6) missRanger imputation (ML-MR), and 7) multiple imputation (ML-MICE). We trained ML models with full data and simulated missing values in testing patients. Prediction performance was evaluated using area under the receiver-operating characteristic curve (AUC) and compared with a model without missing values (ML-All), expert visual diagnosis and total perfusion deficit (TPD). RESULTS: During mean follow-up of 4.7 ± 1.5 years, 3,541 patients experienced at least one MACE (3.7% annualized risk). ML-All (reference model-no missing values) had AUC 0.799 for MACE risk prediction. All seven models with missing values had lower AUC (ML-Remove: 0.778, ML-MICE: 0.774, ML-Cluster: 0.771, ML-Traditional: 0.771, ML-Regression: 0.770, ML-MR: 0.766, and ML-Unique: 0.766; p < 0.01 for ML-Remove vs remaining methods). Stress TPD (AUC 0.698) and visual diagnosis (0.681) had the lowest AUCs. CONCLUSION: Missing values reduce the accuracy of ML models when predicting MACE risk. Removing variables with missing values and retraining the model may yield superior patient-level prediction performance.

Determining a minimum set of variables for machine learning cardiovascular event prediction: results from REFINE SPECT registry.

AIMS: Optimal risk stratification with machine learning (ML) from myocardial perfusion imaging (MPI) includes both clinical and imaging data. While most imaging variables can be derived automatically, clinical variables require manual collection, which is time-consuming and prone to error. We determined the fewest manually input and imaging variables required to maintain the prognostic accuracy for major adverse cardiac events (MACE) in patients undergoing a single-photon emission computed tomography (SPECT) MPI. METHODS AND RESULTS: This study included 20 414 patients from the multicentre REFINE SPECT registry and 2984 from the University of Calgary for training and external testing of the ML models, respectively. ML models were trained using all variables (ML-All) and all image-derived variables (including age and sex, ML-Image). Next, ML models were sequentially trained by incrementally adding manually input and imaging variables to baseline ML models based on their importance ranking. The fewest variables were determined as the ML models (ML-Reduced, ML-Minimum, and ML-Image-Reduced) that achieved comparable prognostic performance to ML-All and ML-Image. Prognostic accuracy of the ML models was compared with visual diagnosis, stress total perfusion deficit (TPD), and traditional multivariable models using area under the receiver-operating characteristic curve (AUC). ML-Minimum (AUC 0.798) obtained comparable prognostic accuracy to ML-All (AUC 0.799, P = 0.19) by including 12 of 40 manually input variables and 11 of 58 imaging variables. ML-Reduced achieved comparable accuracy (AUC 0.796) with a reduced set of manually input variables and all imaging variables. In external validation, the ML models also obtained comparable or higher prognostic accuracy than traditional multivariable models. CONCLUSION: Reduced ML models, including a minimum set of manually collected or imaging variables, achieved slightly lower accuracy compared to a full ML model but outperformed standard interpretation methods and risk models. ML models with fewer collected variables may be more practical for clinical implementation.

Clinical Deployment of Explainable Artificial Intelligence of SPECT for Diagnosis of Coronary Artery Disease.

BACKGROUND: Explainable artificial intelligence (AI) can be integrated within standard clinical software to facilitate the acceptance of the diagnostic findings during clinical interpretation. OBJECTIVES: This study sought to develop and evaluate a novel, general purpose, explainable deep learning model (coronary artery disease-deep learning [CAD-DL]) for the detection of obstructive CAD following single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI). METHODS: A total of 3,578 patients with suspected CAD undergoing SPECT MPI and invasive coronary angiography within a 6-month interval from 9 centers were studied. CAD-DL computes the probability of obstructive CAD from stress myocardial perfusion, wall motion, and wall thickening maps, as well as left ventricular volumes, age, and sex. Myocardial regions contributing to the CAD-DL prediction are highlighted to explain the findings to the physician. A clinical prototype was integrated using a standard clinical workstation. Diagnostic performance by CAD-DL was compared to automated quantitative total perfusion deficit (TPD) and reader diagnosis. RESULTS: In total, 2,247 patients (63%) had obstructive CAD. In 10-fold repeated testing, the area under the receiver-operating characteristic curve (AUC) (95% CI) was higher according to CAD-DL (AUC: 0.83 [95% CI: 0.82-0.85]) than stress TPD (AUC: 0.78 [95% CI: 0.77-0.80]) or reader diagnosis (AUC: 0.71 [95% CI: 0.69-0.72]; P < 0.0001 for both). In external testing, the AUC in 555 patients was higher according to CAD-DL (AUC: 0.80 [95% CI: 0.76-0.84]) than stress TPD (AUC: 0.73 [95% CI: 0.69-0.77]) or reader diagnosis (AUC: 0.65 [95% CI: 0.61-0.69]; P < 0.001 for all). The present model can be integrated within standard clinical software and generates results rapidly (<12 seconds on a standard clinical workstation) and therefore could readily be incorporated into a typical clinical workflow. CONCLUSIONS: The deep-learning model significantly surpasses the diagnostic accuracy of standard quantitative analysis and clinical visual reading for MPI. Explainable artificial intelligence can be integrated within standard clinical software to facilitate acceptance of artificial intelligence diagnosis of CAD following MPI.

Prognostically safe stress-only single-photon emission computed tomography myocardial perfusion imaging guided by machine learning: report from REFINE SPECT.

AIMS: Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) stress-only protocols reduce radiation exposure and cost but require clinicians to make immediate decisions regarding rest scan cancellation. We developed a machine learning (ML) approach for automatic rest scan cancellation and evaluated its prognostic safety. METHODS AND RESULTS: In total, 20 414 patients from a solid-state SPECT MPI international multicentre registry with clinical data and follow-up for major adverse cardiac events (MACE) were used to train ML for MACE prediction as a continuous probability (ML score), using 10-fold repeated hold-out testing to separate test from training data. Three ML score thresholds (ML1, ML2, and ML3) were derived by matching the cancellation rates achieved by physician interpretation and two clinical selection rules. Annual MACE rates were compared in patients selected for rest scan cancellation between approaches. Patients selected for rest scan cancellation with ML had lower annualized MACE rates than those selected by physician interpretation or clinical selection rules (ML1 vs. physician interpretation: 1.4 ± 0.1% vs. 2.1 ± 0.1%; ML2 vs. clinical selection: 1.5 ± 0.1% vs. 2.0 ± 0.1%; ML3 vs. stringent clinical selection: 0.6 ± 0.1% vs. 1.7 ± 0.1%, all P < 0.0001) at matched cancellation rates (60 ± 0.7, 64 ± 0.7, and 30 ± 0.6%). Annualized all-cause mortality rates in populations recommended for rest cancellation by physician interpretation, clinical selection approaches were higher (1.3%, 1.2%, and 1.0%, respectively) compared with corresponding ML thresholds (0.6%, 0.6%, and 0.2%). CONCLUSION: ML, using clinical and stress imaging data, can be used to automatically recommend cancellation of rest SPECT MPI scans, while ensuring higher prognostic safety than current clinical approaches.