Integration of coronary artery calcium scoring from CT attenuation scans by machine learning improves prediction of adverse cardiovascular events in patients undergoing SPECT/CT myocardial perfusion imaging

Abstract

Background.

Machine learning (ML) has been previously applied for prognostication in patients undergoing SPECT myocardial perfusion imaging (MPI). We evaluated whether including attenuation CT coronary artery calcification (CAC) scoring improves ML prediction of major adverse cardiovascular events (MACE) in patients undergoing SPECT/CT MPI.

Methods.

From the REFINE SPECT Registry 4770 patients with SPECT/CT performed at a single center were included (age: 64 ± 12 years, 45% female). ML algorithm (XGBoost) inputs were clinical risk factors, stress variables, SPECT imaging parameters, and expert-observer CAC scoring using CT attenuation correction scans performed to obtain CT attenuation maps. The ML model was trained and validated using tenfold hold-out validation. Receiver Operator Characteristics (ROC) curves were analyzed for prediction of MACE. MACE-free survival was evaluated with standard survival analyses.

Results.

During a median follow-up of 24.1 months, 475 patients (10%) experienced MACE. Higher area under the ROC curve for MACE was observed with ML when CAC scoring was included (CAC-ML score, 0.77, 95% confidence interval [CI] 0.75–0.79) compared to ML without CAC (ML score, 0.75, 95% CI 0.73–0.77, P = .005) and when compared to CAC score alone (0.71, 95% CI 0.68–0.73, P < .001). Among clinical, imaging, and stress parameters, CAC score had highest variable importance for ML. On survival analysis patients with high CAC-ML score (> 0.091) had higher event rate when compared to patients with low CAC-ML score (hazard ratio 5.3, 95% CI 4.3–6.5, P < .001).

Conclusion.

Integration of attenuation CT CAC scoring improves the predictive value of ML risk score for MACE prediction in patients undergoing SPECT MPI.

INTRODUCTION

Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) is a widely used technique for the diagnosis of coronary artery disease (CAD). It is estimated that 20 million SPECT studies are performed yearly worldwide with approximately 6.5 million studies done annually in the USA alone.¹ SPECT MPI has undergone major technological advances recently with the implementation of high-efficiency scanners incorporating cadmium zinc telluride (CZT) solid-state detectors, specialized collimators, and software-based resolution recovery. These advances resulted in improvement in count sensitivity and superior image quality when compared to conventional SPECT cameras. The Registry of Fast Myocardial Perfusion Imaging With Next-Generation SPECT registry (REFINE SPECT)² is an international multicenter observational cohort study of patients with suspected or known CAD who underwent SPECT MPI using CZT systems. This registry enables the analysis of high-quality, state-of-the-art SPECT images at a centralized core laboratory with detailed analysis of myocardial perfusion, function, and coronary calcifications.

Coronary artery calcium (CAC) score derived from ECG-gated low-dose non-contrast computed tomography (CT) scans can be used to quantify coronary calcium burden and can serve as a strong and independent predictor of cardiovascular events³ and can provide complemental diagnostic and prognostic data in addition to MPI.⁴ In addition, SPECT has been increasingly acquired with hybrid SPECT/CT systems for attenuation purposes. Previous studies suggest that CT performed for attenuation correction (CTAC) can be used to quantify CAC burden with good correlation between Agatston score derived from CTAC and from ECG-gated dedicated CAC scoring CT examinations.^5–7

Previously our group has provided evidence about the prediction of coronary artery disease (CAD) using machine learning (ML) models with MPI.^8–10 In addition, we have demonstrated that ML models outperformed standard interpretation methods and risk models in predicting major adverse cardiovascular events (MACE) and could be employed in defining the fewest clinical and imaging variables required to maintain the prognostic accuracy for MACE.¹¹ However, to date it is unknown whether incorporation of CAC scoring from CTAC improves the performance of ML for predicting MACE in patients undergoing SPECT MPI. Therefore, we aimed to evaluate whether including CTAC coronary artery calcification (CAC) scoring improves ML prediction of major adverse cardiovascular events (MACE) in patients undergoing SPECT/CT MPI.

METHODS

Study population

The REFINE SPECT registry² is an international multicenter observational cohort study of patients with suspected or known CAD who underwent SPECT MPI using CZT solid-state camera systems. From the REFINE SPECT Registry, patients with CZT SPECT/CT performed at Yale New Haven Hospital were included after exclusion of those patients who did not undergo attenuation correction. We also excluded patients who had a low-resolution CT scan acquired by a 4-slice CT which could be used for attenuation correction, but not for diagnosing coronary artery calcifications.

Clinical data

Demographic data were collected about the participants’ age, gender, body mass index, family history of CAD, smoking status and about the presence of hypertension, dyslipidemia, diabetes, peripheral artery disease, history of previous myocardial infarction (MI), prior percutaneous coronary intervention (PCI), and prior coronary artery bypass graft (CABG) surgery. Resting blood pressure and heart rate were recorded prior to SPECT examination. Resting ECG was assessed for presence of left ventricular hypertrophy or conduction disease.

Image acquisition and protocol

All patients underwent stress perfusion and gated SPECT imaging using ^99mTc-tetrofosmin with a Discovery NM 530c or Discovery 570c scanner (GE, Healthcare, Haifa, Israel). Patients underwent either symptom-limited exercise treadmill stress testing or pharmacological stress with regadenoson. Static and gated images were acquired. Static images were reconstructed with and without attenuation correction, whereas gated images were reconstructed without attenuation correction. Patients were selected for exercise or pharmacological stress using clinical parameters. Stress results including exercise duration, symptoms and both resting and stress heart rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), and electrocardiogram (ECG) findings were recorded and verified by experienced nuclear cardiologists during clinical interpretation.

CT attenuation map acquisition

CTAC scans were performed free breathing without ECG gating in helical mode acquired with Discovery 570c for nuclear images acquired for both NM530c and Discovery 570c. The acquisition parameters were adjusted by the technologists based on patients’ body mass index (BMI). For patients with BMI < 40 kg/m², the following parameters were used: tube current: 60 mA, tube voltage: 120 kV, rotation time: 0.4 seconds, pitch: 0.98, number of slices: 89, helical slice thickness: 2.5 mm, and slice spacing: 2.5 mm. For patient with BMI ≥ 40 kg/m², similar parameters were used with adjusting the tube current to 150 mA. Images were reconstructed with 2.5 mm thickness using a full-angle reconstruction.

SPECT image analysis

All image data were deidentified and transferred to the core laboratory at Cedars-Sinai. Quality control was performed by experienced technologists at the core laboratory for perfusion and gating images.² The myocardial perfusion and function variables, including total perfusion deficit (TPD), left ventricular end-diastolic volume, left ventricular end-diastolic wall volume, left ventricular end-systolic volume, and phase variables, were derived automatically using Quantitative Perfusion SPECT/Quantitative Gated SPECT software (Cedars-Sinai Medical Center, Los Angeles, CA).

CTAC calcium scoring

CAC scoring of CTAC scans was performed by an experienced observer using standard clinical tool (Cardiac Suite Cedars-Sinai Medical Center, Los Angeles, CA) using a standard 130-HU threshold.^5,12 For each patient, the Agatston score was computed.

Machine learning model

The ML model inputs included clinical variables (traditional risk factors for CAD), stress testing parameters related to exercise or pharmacological stress response, SPECT imaging parameters, and CTAC CAC scoring. Models were built using XgBoost classification model, a state-of-the-art ensemble boosting ML algorithm which has demonstrated high performance in cardiac CT-based risk stratification.^13–16 The machine learning development and reporting were performed according to RELAINCE guidelines.¹⁷

Training validation and testing

To avoid reporting biased results and limit overfitting, we validated our ML algorithm using tenfold hold-out testing as previously described.¹³ Briefly, this involved stratifying and dividing the data into tenfolds of equal size: eight folds (80%) were used for training, one fold (10%) was used for tuning model parameters, and one fold (10%) was used only for testing. Stratification was used to ensure a similar distribution of events across the tenfolds. This process was repeated 10 times, always using a different fold for model training, tuning, and testing. The ML risk scores were concatenated from all 10 testing data folds to allow assessment of model performance over the entire dataset. This way it was possible to maintain a large sample size for both training and testing while reducing the variance in prediction error and preventing the results from being dependent on an arbitrary split of data.

Comparison with logistic regression

To allow comparison with more basic machine learning methods we additionally trained Bayesian logistic regression (LR) models using the same input data and training-validation regimen as for Xgboost model. Performance of the LR model was evaluated using ROC curves.

Outcomes

The primary end point was MACE which included all-cause mortality, non-fatal MI, or late coronary revascularization (percutaneous coronary intervention or coronary artery bypass surgery > 90 days after SPECT imaging). MACE was determined by review of electronic medical records.² The first event that occurred among the above components of MACE was considered the primary end point.

Explainable individualized ML risk prediction

To demonstrate the clinical applicability of our ML model and provide explanation about how the model provides accurate predictions, we provided examples of detailed description of individualized risk predictions made by the ML algorithm. The model allowed identification of important patient-specific variables and the role of the variable in the predicted score. We analyzed the specific path a subject takes in the model; in each decision stump (or split) of the model, the individual landed in one of two leaves in a 1-level decision tree.¹⁸ Each leaf was associated with a weight: one leaf decreasing the risk of the event occurring and the other one increasing the risk. These weights were associated with the variables used to generate the corresponding split. By cumulating all the weights used to refine per-patient prediction for each variable in the model, we could determine whether a parameter had a protective influence, depending on the weight sign. By considering the absolute values of cumulated weights, we could also obtain the global contribution of each parameter.

Statistical analysis

Categorical variables were compared by the χ² test, and continuous variables were compared by the Student’s t test or Mann–Whitney U test, as appropriate. Using receiver-operating characteristic (ROC) analysis and pairwise comparisons according to Delong et al¹⁹the predictive performance for MACE was compared between ML without CAC, ML with CAC, and CAC score alone (all scores were continuous variables). The highest Youden’s J index (J = sensitivity + specificity—1) was used to identify an optimal cutoff for the ML score and stratify subjects into ‘high’ or ‘low’ ML risk. Additionally, we used net reclassification index (NRI) to investigate the improvement of ML model with CAC over ML without CAC.²⁰ Kaplan– Meier analysis was performed according to this threshold, and survival curves were compared with the log-rank test. All statistical analyses were performed with SPSS 27.0.0 (Microsoft Inc., College Station, TX) or R version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria). Training and validation of the models were performed in R Studio using xgboost (version 1.5.0.2) and caret (version 6.0.90) packages. NRI analysis was performed using R and nricens package (version 1.6). A two-sided P < .05 was considered statistically significant.

RESULTS

Patient characteristics

The final study population comprised 4770 patients after exclusion of 230 studies without CTAC or with non-diagnostic CTAC from the total of 5000 Yale New Haven Hospital studies included in the REFINE SPECT registry. Out of the 4770 included studies 4133 were performed on the Discovery 570c with the remaining performed on NM 530c. Table 1 summarizes the baseline characteristics of subjects. During the median follow-up of 22 months (interquartile range [IQR] 13, 33 months) 475 patients (10%) experienced MACE, including 207 deaths, 153 MIs, and 115 late revascularizations (100 percutaneous coronary interventions and 15 coronary artery bypass surgeries). The median time until MACE was 14 months (IQR 6, 22). Subjects who experienced events were older, more likely to be male, had lower body mass index, higher systolic and lower diastolic blood pressure readings, higher rate of left ventricular hypertrophy on ECG, higher rate of smoking, and had higher rate of cardiovascular comorbidities, including hypertension, dyslipidemia, diabetes, peripheral artery disease, prior MI, percutaneous, and surgical revascularization. In addition, subjects who experienced events were more likely to undergo pharmacological testing and had lower stress heart rates and stress systolic and diastolic blood pressures (Table 2). The stress end-diastolic wall volume and stress end-diastolic volumes were lower, whereas stress end-systolic volumes, stress non-corrected total perfusion deficit (TPD), and attenuation corrected TPD were higher in patients who experienced MACE. In addition, calcium score was higher in patients who experienced MACE.

Table 1.

Baseline characteristics

N	Overall N= 4770	MACE N = 475	No MACE N = 4295	P value
Age, years	64 (56–73)	69 (60–78)	64 (56–72)	< .001
Female (%)	2122 (45%)	148 (31%)	1974 (46%)	< .001
BMI, kg/m2	25.5 (29.3–33.7)	28.3 (24.3–32.6)	29.4 (25.6–33.9)	< .001
Family history of CAD	681 (14%)	43 (9%)	638 (15%)	< .001
Smoking	951 (20%)	113 (24%)	838 (20%)	.03
Hypertension	3060 (64%)	331 (70%)	2729 (64%)	.008
Dyslipidemia	2549 (53%)	276 (58%)	2273 (53%)	.03
Diabetes	1240 (26%)	169 (36%)	1071 (25%)	< .001
PAD	1248 (26%)	225 (47%)	1023 (24%)	< .001
History of MI	393 (8%)	72 (15%)	321 (8%)	< .001
History of PCI	532 (11%)	104 (22%)	428 (10%)	< .001
History of CABG	265 (6%)	67 (14%)	198 (5%)	< .001
Resting SBP, mmHg	138 (125–153)	140 (126–158)	138 (124–152)	.002
Resting DBP, mmHg	80 (73–86)	77 (70–85)	80 (73–86)	< .001
Resting HR, beats/min	71 (63–80)	70 (62–79)	71 (63–80)	.53
LVH on resting ECG	308 (7%)	43 (9%)	265 (6%)	.02

Categorical variables are shown as numbers (%), and continuous variables are shown as median values (interquartile range) MACE, major adverse cardiac events; BMI, body mass index; CAD, coronary artery disease; PAD, peripheral artery disease; MI, myocardial infarction; PCI, percutaneous coronary intervention; CABG, coronary artery bypass grafting; SBP, systolic blood pressure; DBP, diastolic blood pressure; HR, heart rate; LVH, left ventricular hypertrophy

Table 2.

Stress test and imaging characteristics

N	Overall N = 4770	MACE N= 475	No MACE N = 4295	P value
Stress type
Exercise	1724 (36%)	65 (14%)	1659 (39%)	< .001
Regadenoson	3046 (64%)	410 (86%)	2636 (61%)
Stress HR, beats/min	107 [89–144]	91 (80–110)	110 (90–146)	< .001
Stress SBP, mmHg	153 (131–175)	142 (120–164)	154 (132–176)	< .001
Stress DBP, mmHg	80 (71–88)	73 (64–82)	80 (72–88)	< .001
Exercise duration
≤ 6 min	435 (25% exercise)	21 (32% exercise)	414 (25% exercise)	.21
7–9 min	828 (48% exercise)	32 (49% exercise)	796 (48% exercise)
≥ 10 min	461 (27% exercise)	12 (19% exercise)	449 (27% exercise)
ECG response to test
Negative	3340 (70%)	298 (63%)	3042 (71%)	< .001
Positive	473 (10%)	40 (8%)	433 (10%)
Equivocal	279 (6%)	20 (4%)	259 (6%)
Non-diagnostic	666 (14%)	115 (24%)	551 (13%)
ECG ST slope
Upsloping	188 (4%)	10 (2%)	178 (4%)	.16
Down-sloping	134 (3%)	12 (3%)	122 (3%)
Horizontal	425 (9%)	40 (8%)	385 (9%)
Stress end-diastolic volume, mL	88 (68–115)	83 (59–113)	87 (68–113)	< .001
Stress end-diastolic wall volume, mL	128 (110–151)	102 (77–137)	127 (109–149)	< .001
Stress end-systolic volume, mL	31 (20–49)	42 (26–70)	30 (19–47)	< .001
Stress TPD (after AC), %	2.34 (0.78–5.13)	3.89 (1.61–7.94)	2.22 (0.71–4.80)	< .001
Stress TPD (no AC), %	2.21 (0.76–5.02)	3.55 (1.16–8.04)	2.12 (0.71–4.75)	< .001
Stress quality control	1.62 (1.27–2.03)	1.76 (1.35–2.25)	1.61 (1.26–2.00)	< .001
Calcium score	88 (0–602)	564 (130–1364)	62 (0–521)	< .001

Categorical variables are shown as numbers (%), and continuous variables are shown as median values (interquartile range) MACE, major adverse cardiac events; HR, heart rate; SBP, systolic blood pressure; DBP, diastolic blood pressure; TPD, total perfusion deficit; AC, attenuation correction

Feature importance

Among clinical, imaging, and stress parameters, stress heart rate, stress diastolic blood pressure, and stress end-diastolic wall volume had highest variable importance in ML score (Figure 1). CAC score showed the highest variable importance in the CAC-ML model (Figure 2). Peripheral vascular disease, stress end-systolic volume, age, body mass index, resting systolic blood pressure, and resting heart rate were among higher ranked clinical variables in both ML score and CAC-ML score. In addition, stress TPD both with and without attenuation correction and stress end-diastolic and end-systolic volumes were highly ranked imaging variables in both ML score and CAC-ML score.

Figure 1.

Variable importance for the classification of cardiac events with machine learning (ML) score without attenuation CT coronary artery calcium score. The top 26 variables are displayed: clinical risk factors in gray, quantitative imaging measures in red, and stress parameters in yellow. The “gain” denotes how much a variable contributes to the prediction made by the XGBoost algorithm. BP, blood pressure; TPD, total perfusion deficit; PCI, percutaneous coronary intervention; CAD, coronary artery disease.

Figure 2.

Variable importance for the classification of cardiac events in machine learning (ML) including attenuation CT coronary artery calcification (CAC) score. The top 26 variables are displayed: CAC score in blue, clinical risk factors in gray, quantitative imaging measures in red, and stress parameters in yellow. The “gain” denotes how much a variable contributes to the prediction made by the XGBoost algorithm. CAC, coronary artery calcium; CTAC, CT attenuation correction; other abbreviations as in Figure 1.

Performance of the ML model

The CAC-ML model integrating CAC score with clinical, imaging, and stress parameters had a significantly higher AUC (CAC-ML score, 0.77, 95% confidence interval [CI] 0.75–0.79) for prediction of MACE when compared to CAC score alone (0.71, 95% CI 0.73–0.77, P < .001) or when compared to ML score without the addition of CAC (ML score, 0.75, 95% CI 0.68–0.73, P = .005) (Figure 3). These results were also confirmed in NRI analysis where the model with CAC had overall NRI of 0.09 (95% CI 0.02, 0.17). The positive and negative NRI were 0.14 (95% CI 0.08, 0.2) and - 0.05 (95% CI - 0.07, - 0.03), respectively. Patients with high ML score (> 0.090) had higher event rate when compared to patients with low ML score (< 0.090, hazard ratio [HR] 4.38, 95% confidence interval [CI] 3.57 to 5.37, P < .001) (Figure 4A). Patients with high CAC-ML score (> 0.091) had higher event rate when compared to patients with low CAC-ML score (< 0.091, HR 5.25, 95% CI 4.25 to 6.49, P < .001) (Figure 4B). Performance of LR models, while inferior to XgBoost, confirmed the added value of CAC score to input variables (AUC 0.753, 95% CI 0.730, 0.775, and 0.747, 95% CI 0.724, 0.77 for LR model with and without CAC, respectively, P = .002).

Figure 3.

Receiver-operating characteristic (ROC) curves for the prediction of major adverse cardiac events (MACE). The machine learning (ML) model with attenuation CT coronary artery calcium (CAC) score performed significantly better compared to ML without CAC (P = .005) and compared to attenuation CT CAC score alone (P < .001).

Figure 4.

Kaplan–Meier curves of cardiac events with a high versus low machine learning (ML) risk score (Panel A) and high versus low coronary artery calcification (CAC)-ML risk score. CI, confidence interval.

Explainable individualized ML risk prediction

Figure 5 demonstrates a case example of individualized ML risk score prediction for a 70-year-old male who suffered myocardial infarction 78 days after SPECT MPI (Figure 5A). Myocardial perfusion imaging demonstrated a small to medium-sized, mild-intensity reversible perfusion defect in the basal to apical inferior wall consistent with ischemia (Figure 5B). Attenuation CT showed severe calcifications in the left anterior descending and right coronary arteries with a calculated calcium score of 1314 (Figure 5C). The ML score without CAC stratified the patient in the low-risk category. After the addition of CAC assessment to the model the patient was re-classified to the high-risk category. Figure 6 demonstrates another case example with individualized ML risk prediction in an 84-year-old female who survived 3.3 years without MACE (Figure 6A). Myocardial perfusion was normal (Figure 6B) and attenuation CT showed no CAC (Figure 6C). The ML score without CAC stratified the patient in the high-risk category. After adding the information about the lack of CAC to the model the patient was re-classified to the low-risk category.

Figure 5.

Individualized machine learning (ML) risk prediction for a 70-year-old male after SPECT MPI (Panel A) with representative images of SPECT myocardial imaging (Panel B) and attenuation CT (Panel C) showing left anterior descending artery coronary calcifications (yellow arrow). HR, heart rate; DBP, diastolic blood pressure; SBP, systolic blood pressure; PAD, peripheral artery disease; AC, attenuation correction; NC, non-corrected; EDWV, end-diastolic wall volume; BMI, body mass index; PCI, percutaneous coronary intervention; CAC, coronary artery calcification.

Figure 6.

Individualized machine learning (ML) risk prediction for an 84-year-old female after SPECT MPI (Panel A) with representative images of SPECT myocardial imaging (Panel B) and attenuation CT (Panel C). Abbreviations as in Figure 5.

DISCUSSION

To our knowledge this is the first study to show that integration of attenuation CT CAC scoring in a ML-based risk assessment algorithm improves the value of risk prediction in patients undergoing SPECT myocardial perfusion imaging. Our data suggest that quantifying CAC on attenuation CT scans can help to better risk stratify patients with the help of explainable ML methods in patients undergoing SPECT MPI using CZT systems. Our findings further support the expanding use of hybrid myocardial perfusion imaging systems, especially the ones equipped with CZT technology.

CAC scoring by non-contrast CT provides accurate estimation of CAC which serves as imaging evidence of coronary artery disease burden within the epicardial coronary arteries.²¹ CAC score is a strong predictor of future cardiovascular risk with a prognostic value higher than traditional risk assessment tools, such as the Framingham risk score and the Atherosclerotic Cardiovascular Disease risk score.^21,22 Recently our group⁴ and other investigators^23,24 demonstrated that CAC scoring on dedicated CAC scoring CT scans increases the diagnostic accuracy for obstructive CAD when performed in conjunction with MPI. In addition, limited studies suggest that CAC score and myocardial perfusion may be independent predictors for MACE in symptomatic patients undergoing evaluation for suspected CAD.^25,26 Recently Patel et al. found modest increase in predicting MACE with adding dedicated calcium scoring results to a model including clinical risk factors and myocardial perfusion imaging results in a cohort of patients undergoing PET myocardial perfusion imaging.²⁶ This is in line with our current findings as the CAC-ML prediction score which incorporated both perfusion metrics and CAC score derived from SPEC/CT attenuation maps outperformed the ML score, which included clinical and imaging variables without CAC data.

In hybrid SPECT/CT systems and PET MPI a low-dose CT attenuation correction scan is performed to correct for attenuation. This attenuation CT can be used for qualitative and quantitative estimation of CAC. The main argument for omitting the acquisition of a dedicated CAC scoring CT without losing additional prognostic data is for decreasing scan time, expense, and radiation. Small studies have reported good agreement between visual estimation of CAC from low-dose CT attenuation correction scans in hybrid PET/CT and SPECT/CT with standard Agatston score.^6,7 The presence of CAC on attenuation CT by visual estimation has been shown to increase diagnostic accuracy for obstructive CAD²⁷ and to carry significant prognostic information beyond perfusion imaging.^27–29 We have to mention that these studies, similar to our study, employed high-spatial and temporal resolution CT scanners, whereas CT scanners on the most frequently used hybrid SPECT-CT systems have limited contrast and temporal resolution, which limits the generalizability of our findings. In addition to the qualitative assessment, CAC quantification has been demonstrated to be feasible on attenuation CT scans with excellent correlation with standard Agatston scores.^7,30–32 As a recognition of the important prognostic value of CAC on non-gated CT scans, in the recently published joint guidelines the Society of Cardiovascular Computed Tomography and the Society of Thoracic Radiology, in addition to the mandatory visual estimation for presence of CAC, encourage computation of a non-gated Agatston score for all non-contrast CT examinations.³³ Our current paper provides further evidence about the prognostic value of CAC score from CTAC scans performed to obtain CT attenuation maps. Optimization of CTAC acquisition protocols could potentially further improve the incremental value of CTAC calcium scoring.

By objectively integrating clinical data, SPECT imaging measures, and CAC score, our ML score showed superior performance for MACE prediction compared to CAC score alone. We have used tenfold cross-validation to maintain a robust estimation of prediction accuracy with minimal bias, which can serve as powerful alternative when separate validation cohorts are not available.³⁴ We have previously used similar ensemble boosting ML algorithm (XGBoost) to analyze the prospective EISNER trial showing that ML integration of novel circulating biomarkers and noninvasive imaging measures including CAC score provided superior long-term risk prediction for cardiac events compared to current risk assessment tools.¹³ We have also used similar boosting algorithm successfully to identify the minimal set of imaging and clinical variables to maintain good prognostic accuracy for MACE prediction in the REFINE SPECT cohort.¹¹ A major advantage of Xgboost model is that it provides a description about the influence of each variable for the individualized risk prediction as demonstrated by case examples in the current study. This description may help physicians to identify specific targets for the individual patients for risk reduction strategies.

Our results extend the findings of several other recent studies which report on the use of ML for prediction of MACE with myocardial perfusion imaging.^35–37 We have recently reported about the potential use of ML for the safe cancelation of rest imaging in SPECT after ML-based stress image evaluation.³⁵ The ML approach showed a better risk stratification to qualify for stress-only imaging with an annual MACE risk of 1.4%, compared to the visual risk assessment that resulted in 2.1% MACE risk while selecting a similar proportion of patients (~ 60%). Related to this, we have found ML-based approach to be highly sensitive for the prediction of obstructive CAD from stress-only MPI.³⁸ We have also previously applied ML-based risk assessment without integration of CAC scoring in a different cohort undergoing SPECT MPI.¹¹ Importantly, this study used a more inclusive composite end point by including early revascularization in addition to cardiac death, myocardial infarction, and late revascularization which could potentially explain the higher AUC. Altogether, these studies suggest that ML techniques can be applied to improve risk stratification in patients undergoing MPI. Our findings complement these previous studies by showing that incorporation of attenuation CT CAC score could potentially improve the performance of ML models.

Limitations

Despite the relatively large patient number, this is a retrospective study with all of its inherent limitations. In the current study we have included patients only from a single academic center. However, the images were analyzed in a core laboratory with blinded image analysis. The composite outcome included late revascularization which is not considered to be a ‘hard’ cardiac event. However, it is important to mention that the primary outcome was driven by non-fatal MI and all-cause death. All images were acquired with SPECT systems employing CZT technology, therefore the results may not be generalizable to all hybrid MPI systems. XGBoost, being a gradient boosting decision tree algorithm, requires longer training times and is prone to overfitting. Our ML model will require further internal or external validation in an independent cohort.

CONCLUSION

In this study, ML integration of clinical, stress, and myocardial perfusion imaging parameters with attenuation CT CAC scoring provided superior risk prediction for cardiac events compared to ML score without CAC scoring or CAC score alone. Future studies are warranted to confirm these findings in an external cohort and to test the value of ML-based risk prediction in a prospective setting.

NEW KNOWLEDGE GAINED

Our study demonstrated that integration of attenuation CT CAC scoring in a ML-based risk assessment algorithm improves the value of risk prediction in patients undergoing SPECT myocardial perfusion imaging. Our findings further support the expanding use of hybrid myocardial perfusion imaging systems and suggests that quantifying CAC on attenuation CT scans can help to better risk stratify patients with the help of explainable ML methods.

Supplementary Material

Supplementary Information The online version contains supplementary material available at https://doi.org/10.1007/s12350-022-03099-x.