Machine learning for prognostic prediction in coronary artery disease with SPECT data: a systematic review and meta-analysis

Abstract

Background

Single-photon emission computed tomography (SPECT) analysis relies on qualitative visual assessment or semi-quantitative measures like total perfusion deficit that play a critical role in the non-invasive diagnosis of coronary artery disease by assessing regional blood flow abnormalities. Recently, machine learning (ML) -based analysis of SPECT images for coronary artery disease diagnosis has shown promise, with its utility in predicting long-term patient outcomes (prognosis) remaining an active area of investigation. In this review, we comprehensively examine the current landscape of ML-based analysis of SPECT imaging with an emphasis on prognostication of coronary artery disease.

Main body

Our systematic search yielded twelve retrospective studies, investigating SPECT-based ML models for prognostic prediction in coronary artery disease patients, with a total sample size of 73,023 individuals. Several of these studies demonstrate the superior prognostic capabilities of ML models over traditional logistic regression (LR) models and total perfusion deficit, especially when incorporating demographic data alongside SPECT imaging. Meta-analysis of 6 studies revealed promising performance of the included ML models, with sensitivity and specificity exceeding 65% for major adverse cardiovascular events and all-cause mortality. Notably, the integration of demographic information with SPECT imaging in ML frameworks shows statistically significant improvements in prognostic performance.

Conclusion

Our review suggests that ML models either independently or in combination with demographic data enhance prognostic prediction in coronary artery disease.

Background

Chronic coronary syndromes (CCS), encompassing both macrovascular epicardial coronary artery disease (CAD) and microvascular dysfunction, represent a significant global burden, ranking as the leading cause of mortality and morbidity [1]. Accurate diagnosis and risk stratification are crucial for optimizing patient management in CCS. Myocardial perfusion single-photon emission computed tomography (SPECT) has emerged as a valuable non-invasive tool for the assessment of CAD, primarily through subjective visual interpretation [2]. While quantitative perfusion analysis offers advantages in cardiovascular risk stratification and guiding treatment decisions, it often overlooks other significant SPECT findings that can contribute to long-term risk prediction [2].

SPECT myocardial perfusion imaging (MPI) offers valuable insights into myocardial blood flow abnormalities, providing prognostic information for patients with coronary artery disease (CAD) [3]. However, integrating this information with a broader range of clinical data poses challenges for clinicians due to the complexity of manual data fusion. This limitation can potentially adversely affect the accuracy of prognostic assessment and selection of optimal treatment strategies. A data-driven approach that leverages both SPECT-MPI data and comprehensive clinical information has the potential to overcome these hurdles. By quantitatively integrating these multimodal data sources, one could provide clinicians with a more robust and objective assessment of a patient’s risk for adverse cardiovascular outcomes, such as cardiac mortality.

An increasing number of AI applications in cardiovascular medicine has the potential to improve cost-effectiveness and guide decision-making. Simultaneously, over the last decade, there has been a tremendous progress in technology related to non-invasive imaging techniques, which are crucial in the diagnosis and treatment of cardiovascular disease [3]. In terms of terminology, AI is the application of computer systems to carry out operations that often call for human intelligence. These computer systems may automatically learn to do a task and get better with experience through exposure to vast volumes of data thanks to) ML, a subdiscipline of AI neural networks used in deep learning (DL). Convolutional neural networks (CNN) are the most often utilized DL networks in medical image analysis. Explainable AI is a term used to describe ML techniques that allow humans to understand the learning outcomes, with an aspired goal of enabling more transparency in interpreting the inner workings of ML models [4].

The field of medicine, particularly cardiovascular medicine, is witnessing a transformative convergence of artificial intelligence (AI) and advancements in non-invasive imaging techniques. Machine learning (ML) as a subdiscipline of AI, has the potential to revolutionize cardiovascular care by enabling automated analysis of vast datasets and extracting hidden patterns for improved decision-making. This aligns perfectly with the tremendous progress witnessed in non-invasive imaging modalities over the past decade. These advancements have provided crucial insights for diagnosis, risk stratification, and treatment guidance in cardiovascular disease (CVD) [3].

ML has been always a hot topic in medicine; a seminal paper titled ‘’ImageNet’’ (ref), has given way to a “deep learning (DL) tsunami” in almost all fields, replacing conventional ML methods, largely owing to human-level and occasinally superhuman level performances. DL based approaches have been shown to predict health outcomes across various specializations. In cardiovascular medicine, DL has shown promising results in analyzing SPECT-MPI data [5]. By identifying novel relationships within medical imaging data and integrating it with clinical information, both ML and DL techniques have been used to build predictive models for a range of outcomes [5] (Fig. 1). In this context, DL and ML methods were mostly explored for in the context of assessing the diagnostic utility of SPECT-MPI for CAD [6]; its role in prognostic assessment remains an active area of investigation. This review, thus, comprehensively examines all published studies on the use of DL and ML with SPECT data specifically for prognostication in CAD patients. By gathering the current knowledge base, this review aims to not only highlight the existing research landscape but also emphasize the need for further exploration in this rapidly evolving field.

Fig. 1.

PRISMA flowchart for the study screening and selection

Overview of machine learning and deep learning

The field of ML emerged in the 1950s with the goal of enabling computers to learn and adapt. This discipline empowers computers to acquire new skills and potentially uncover patterns possibly beyond human comprehension. Artificial neural networks, a core component of the ML revolution in medical imaging, are further analyzed within the subfield of DL [7]. DL involves stacked layers of artificial neurons that iteratively extract increasingly complex features from the input data. By feeding the output of each layer into the next, the system progressively refines its understanding of the incoming information. This hierarchical architecture enables DL models to learn from massive datasets, achieving state-of-the-art performance in various tasks such as image and audio recognition, natural language processing, and more [8, 9].

The success of DL hinges on three key factors: the development of more efficient learning algorithms (e.g., backpropagation), the availability of large-scale datasets, and the emergence of powerful computing resources like graphics processing units (GPUs) [8, 9]. These advancements have propelled DL to demonstrate remarkable results in several applications, ranging from facial recognition and medical diagnostics to drug discovery and autonomous vehicles [8, 9].

While the number of layers directly correlates with a model’s complexity, this does not necessarily confer better efficacy. As the learning process becomes more challenging, modeling said data and extracting meaningful or more subtle discriminative features become much more difficult. Some of the most prevalent DL architectures include convolutional neural networks (CNNs), recurrent neural networks, variational autoencoders, and generative adversarial networks [10]. Despite significant progress, DL is not without limitations. Building highly accurate and generalizable models often requires vast amounts of training data. Often times, models can be quite fragile, provide black-box (lack of interpretability) output, and appear over-confident in their decisions (lack of calibration). Additionally, the quality and quantity of input data, along with the degree of human intervention in model development, can significantly impact the final outcome [11].

Brief history of machine learning in nuclear myocardial perfusion imaging

Early research on applying ML models to MPI, beginning in the 1990s, primarily focused on multi-layer perceptron for perfusion map classification [12]. These models faced limitations in handling the high dimensionality of imaging data due to their fully connected layers, unlike CNN based methods where contemporary designs include convolutional layers. The rise of the 2000s saw continued exploration of ML for MPI, with studies comparing MLPs to other conventional algorithms like support vector machines and K-nearest neighbors [13, 14].

Over the past few years, there has been a significant shift towards DL, particularly CNNs, for MPI analysis. Unlike MLPs, CNNs utilize convolutional layers with weight sharing, significantly reducing the number of parameters needed to learn image features. This hierarchical architecture allows CNNs to progressively extract low-level features, edges, and higher-order patterns from raw pixel data, ultimately leading to improved prediction accuracy [15, 16]. A review by Alskaf et al. even suggests that CNNs outperform human experts in identifying perfusion defects on SPECT images [6].

Beyond diagnosis, DL applications in MPI are demonstrating promise in predicting adverse outcomes like death, myocardial infarction, and the need for revascularization [6]. Additionally, DL offers the potential for generating high-quality full-dose MPI images from low-dose or shortened scans, thereby reducing radiation exposure for patients.

Main text

Design

This systematic review and meta-analysis is registered in the International Prospective Register of Systematic Reviews (PROSPERO, CRD42024528286) and reported accordingly to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy (PRISMA-DTA) guidelines [17]. Searching activities were carefully conducted by two independent researchers (VC, AE and MB) and all decisions were made on mutual agreement by all authors. The research question for this review was created using the PICO approach as follows:

• Population: Patients with coronary artery disease (CAD).

• Intervention: SPECT imaging-based machine learning (ML) models.

• Comparison: Conventional CAD imaging.

• Outcome: Prediction of patient outcomes based on SPECT and CAD.

Selection criteria

To comprehensively identify relevant studies, we conducted a systematic search across PubMed, Embase, and Cochrane databases. The search strategy employed a combination of Medical Subject Headings (MeSH) terms and relevant keywords encompassing “single-photon emission computed tomography (SPECT),” “myocardial perfusion imaging (MPI),” “machine learning (ML),” “artificial intelligence (AI),” “deep learning (DL),” “prognosis,” and “coronary artery disease (CAD).” Two independent reviewers (VC and YT) screened the retrieved articles based on pre-defined inclusion and exclusion criteria. Disagreements were resolved through discussion by all authors to ensure consistent application of selection criteria. This rigorous two-step screening process minimized bias and ensured the inclusion of high-quality studies relevant to the current review’s focus on AI-based prognostication using SPECT-MPI in CAD. In total, 859 relevant studies were reached.

Data extraction

Two independent reviewers (VC and YT) meticulously evaluated the full texts of all eligible studies using a standardized data extraction form. This form captured key study characteristics, including first author, publication year, study design (prospective/ retrospective), patient population size, specific ML/DL model employed, follow-up duration, and the area under the receiver operating characteristic curve (AUC) for the ML/DL model compared to other models (if applicable). This structured approach ensured consistent data extraction and facilitated a comprehensive analysis of both clinical and artificial intelligence aspects within the studies.

Statistical analysis

The narrative synthesis was used to show the performance of the applied ML/DL model on prediction of outcomes and a meta-analysis was performed to measure the prognostic accuracy of models and presented as forest plots. In the meta-analysis, specificity and sensitivity were used to measure the prognostic accuracy. In the absence of reported values for true positive (TP), false positive (FP), false negative (FN), and true negative (TN) in the included studies, a confusion matrix was constructed for each study within the meta-analysis. This involved utilizing the reported sample size (S) along with the sensitivity and specificity data provided in the studies. First, FN was calculated by multiplying (1 - sensitivity) by the sample size (S). Next, TN was derived by subtracting FN from the sample size (S). Similarly, FP was computed by multiplying (1 - specificity) by the sample size (S), and TP was obtained by subtracting FP from the sample size (S). This approach allowed for the estimation of TP, FP, FN, and TN values necessary for the meta-analysis. All statistical analysis was performed using RevMan 5.4.1 (The Nordic Cochrane Center, Copenhagen, Denmark). A p-value of less than 0.05 was considered statistical significance.

Results

The search procedure identified 859 records. Following duplicate removal, 397 studies underwent title and abstract screening. An additional 14 studies were identified through reference review, resulting in a total of 411 studies for initial screening. After meticulous full-text evaluation, twelve studies met the inclusion criteria for data extraction and qualitative synthesis within this review. Six of these twelve studies were further eligible for meta-analysis. Figure 1 shows the review process and outlines the reasons for exclusion when applicable.

Characteristics of studies

This review included 12 studies with a total sample size of 73,023 individuals. The first of these studies in predicting the prognosis of SPECT-based ML/DL models was published by Betancur et al. [18] in 2018, as seen in Table 1, and the last study was published by Ananya Singh et al. [19] in 2023. Among these studies, four of them [19–22] were conducted through a multi-center approach with the remained being single center-based. The studies included in this review investigated a range of patient outcomes following CAD diagnosis using SPECT-MPI and ML/DL models. These outcomes encompassed major adverse cardiovascular events (MACE), the need for early revascularization procedures, all-cause mortality including myocardial infarction (MI), and specifically cardiac death. Table 1 summarizes demographic characteristics, population size, follow-up periods and long-term outcomes.

Table 1.

It contains a summary of studies in which a prognostic prediction model was developed with ML using SPECT images

First author	Year	Outcome age gender (male) center	Participants \outcome	Ml model used	Follow up time	Stress tpd vs ml-spect model (AUC)	Stress tpd vs ml-spect + ehr model (AUC)	Lr model vs ml-spect model (AUC)	Lr model vs ml-spect + ehr model (AUC)
Julian Betancur et all [18]	2018	MACE 62 ± 13 YEARS 48% SINGLE CENTER	2619\239	WEKA	3 YEARS	0.73 Vs 0.78 P < 0.05	0.73 Vs 0.81 P < 0.05	-	-
David Haro Alonso et all [19]	2019	CARDIAC DEATH 71 ± 12 YEARS 52% SINGLE CENTER	8321\551	AdaBoost	3.15 ± 1.99 YEARS	-	-	0.76 Vs 0/83 P < 0.001	-
Lien-Hsin Hu et all. [20]	2020	EARLY REVASCULARIZATION 65 ± 11 YEARS 66% MULTICENTER	1970\958	WEKA	6 MONTHS	-	0.71 Vs 0.79 P < 0.001	-	-
Luis Eduardo Juarez-Orozca et all. [21]	2020	MACE 68 ± 9 YEARS - SINGLE CENTER	1085\159	CNN	385 days	0.84 Vs 0.90 P < 0.05		0.78 Vs 0.90 P < 0.05
Valeria Cantoni et all. [22]	2021	MACE 64 ± 10 YEARS 73% SINGLE CENTER	453\41	SVM	2.5 ± 0.5 YEARS	P = 0.043 Vs P < 0.001	-	-	-
Richard Rios et all [23]	2021	MACE 65 ± 12 YEARS 57% MULTICENTER	20,414\3541	XGBoost	4.7 ± 1.5 YEARS	0.69 Vs 0.75 P < 0.05	0.69 Vs 0.79 P < 0.001	-	-
Ananya Singh et all. [24]	2022	ALL CAUSE MORTALITY 71 YEARS 60% SINGLE CENTER	4735\877	CNN	6 YEARS	0.60 Vs 0.82 P < 0.001	-	0.75 Vs 0.82 P < 0.05	-
Eero Lehtonen et all [25]	2023	MACE 62 ± 9 YEARS 42% SINGLE CENTER	2411\210	XGBoost	4 YEARS	-	-	-	P < 0.05**
Luis Eduardo Juarez-Orozca et all. [26]	2023	ALL CAUSE MORTALITY-MI 61 YEARS 43% SINGLE CENTER	739\46	GBM	6.1 YEARS	-	-	-	p = 0.002**
Fares Alahdad et all. [27]	2023	MACE 61.1 ± 14.2 YEARS 54% MULTICENTER	956\102	Auto Sklearn	31 months	-	-	-	- P < 0.001**
Ananya Singh et all.-internal * [28]	2023	ALL CAUSE MORTALITY-MI 64 YEARS 57% MULTICENTER	20,201\1913	HARD MACE-DL	4.6 YEARS	0.63 VS 0.76 P < 0.001	-	0.72 Vs 0.76 P < 0.05	-
Ananya Singh et all.-external * [28]	2023	ALL CAUSE MORTALITY-MI 68 YEARS 54% SINGLE CENTER	9019\719	HARD MACE-DL	3.5 YEARS	0.65 Vs 0.73 P < 0.001	-	0.70 Vs 0.73 P < 0.05	-

*In the relevant study, 2 different models were developed from internal and external data. Therefore, the results of the study are explained in 2 separate columns

**The model was created with DL-SPECT + EHR but it wasn`t compared any model

DL-SPECT: only spect imaging based DL model

DL-SPECT + EHR MODEL: spect imaging + Electornic healt record datas based DL model

LR MODEL: Logistic Regresyon Model

Stress TPD: Total Perfusion Defect Based model

Demographic characteristics

The studies included retrospective cohorts with a mean age of participants up to 60 years (Table 1). Gender distribution data was not available in all studies; however, studies reporting gender breakdown indicated a male predominance exceeding 50%. Patient population sizes ranged from 453 to 20,414 [21, 23].

Outcomes and comparison of models

The reviewed studies investigated a variety of patient outcomes using SPECT-MPI and ML/DL models, including early revascularization, all-cause mortality, major adverse cardiovascular events (MACE), cardiac death, and myocardial infarction (Table 1). Notably, several studies reported statistically significant superiority of ML/DL models derived solely from SPECT imaging compared to stress TPD-based models for predicting said outcomes [18, 24, 25, 19]. Furthermore, five studies demonstrated that SPECT imaging-based ML/DL models significantly outperformed conventional logistic regression (LR) models built using combined features of demographics, laboratory values, and echocardiographic parameters [19, 24–26]. This recapitulates the potential of SPECT MPI data in capturing prognostic information beyond traditional clinical features. Additionally, six studies suggested that ML/DL models incorporating both SPECT imaging and demographic data (SPECT-EHR) achieved statistically significant improvements compared to conventional statistics-based methods, such as LR and stress TPD [18, 20, 21]. Last, but not least, two studies revealed that including demographic information in SPECT-based ML/DL models enhanced their predictive capabilities [18, 21]. These findings collectively highlight the promise of SPECT MPI-based ML for improved prognostic assessment in CAD patients.

Meta-analysis of ML models

We conducted a meta-analysis to evaluate the sensitivity and specificity of ML models in predicting MACE and all-cause mortality in patients with CAD. Six studies investigated MACE prediction using ML/DL models, while four studies focused on all-cause mortality (Table 1). Among these studies, three provided data for sensitivity and specificity on both outcomes, allowing us to perform a meta-analysis. The forest plots (Fig. 2) and SROC curves (Fig. 3) depict the sensitivity and specificity estimates for both MACE and all-cause mortality. The pooled sensitivity for predicting MACE across the included studies was 83.1% (95% CI: 78.6–87.6%), and the pooled specificity was 89.4% (95% CI: 88.2–90.6%). For all-cause mortality, the pooled sensitivity was 69.6% (95% CI: 67.9–71.3%) and the pooled specificity was 65.7% (95% CI: 62.7–66.6%). These results reveal a promising performance of the included ML models, with sensitivity and specificity exceeding 65% for both outcomes.

Fig. 2.

Forest plot and SROC plot of sensitivity and specificity on MACE outcomes with SPECT imaging based ML models

Fig. 3.

Forest plot and SROC plot of sensitivity and specificity on all cause mortality outcomes with SPECT imaging based ML models

Discussion

ML/DL algorithms have demonstrated remarkable capabilities in medical imaging analysis. In the context of coronary artery disease (CAD), ML/DL models have achieved accuracy comparable to human experts in classifying normal and abnormal myocardium in SPECT images [29]. Furthermore, studies have shown success in utilizing ML/DL to identify specific regions of abnormal myocardium, potentially surpassing conventional scoring methods [30]. Building upon these advancements, our current research explores the application of SPECT-based ML models for CAD prognosis.

Our analysis of the reviewed studies (Table 1) revealed that all SPECT-ML/DL models achieved statistically significant prognostic prediction for CAD outcomes. For instance, Nakajima and colleagues discovered that a neural network trained with expert interpretations of SPECT images was superior at detecting abnormalities compared to conventional scoring [31]. This finding aligns with prior research suggesting that integrating clinical data with quantitative imaging features has the potential to enhance the accuracy of DL algorithms [32]. Indeed, three studies within our review demonstrated improved model performance when incorporating demographic data alongside SPECT imaging data.

In SPECT-MPI, quantitative perfusion defect measures like TPD play an important role in diagnosing CAD. TPD offers accuracy comparable to visual interpretation for detecting obstructive stenosis [29]. Several studies have explored TPD-based ML models for CAD diagnosis, demonstrating superior performance to stress TPD alone [30, 33]. These ML/DL models achieved rapid image analysis speeds of less than 0.5 s [33]. Beyond diagnosis, TPD can also be used for prognosis, as evidenced by its application in eight studies included within this review (Table 1). Here, we directly compared the prognostic performance of TPD-based models with SPECT-ML/DL models. Seven studies addressed this clinical inquiry [18, 20–25, 19], with three studies additionally focusing on SPECT-ML/DL models incorporating electronic health records (EHR) [18, 20, 21]. All studies reported statistically significant superiority of AI-based models (either SPECT-ML/DL or SPECT-ML/DL + EHR) compared to TPD models for predicting patient outcomes.

Logistic regression (LR) is a widely used statistical method in medicine for both diagnosis and prognosis of various diseases [34, 35]. LR models can analyze the relationship between a binary outcome (e.g., presence/absence of disease) and multiple influencing factors. Several studies have successfully employed LR for predicting long-term patient outcomes in conditions like CAD [36, 37]. Our review identified studies that compared the prognostic performance of SPECT-ML/DL models with LR models too (Table 1). In all eight studies evaluating this comparison [26, 24], SPECT-ML/DL models demonstrated statistically significant superiority in predicting patient outcomes following CAD diagnosis.

Limitations

Our systematic review identified several limitations. First, the inherent heterogeneity of ML/DL applications and techniques across the included studies limited the ability to perform in-depth comparisons of individual studies. Secondly, while the overall sample size encompassing all studies was substantial, the number of studies appraising particular outcomes or utilizing specific ML/DL techniques remained relatively small. This limited our ability to draw definitive conclusions about the generalizability of findings. Additionally, a number of studies employed TPD scores derived from both rest and stress SPECT images, potentially introducing variability in TPD calculation methods. Finally, reviewed studies evaluated the performance of their proposed ML/DL model against the “best performing” model from prior studies, rather than employing a standardized benchmark. This approach makes it challenging to directly compare the effectiveness of different ML models across studies.

Conclusion

This systematic review demonstrates the promising potential of machine learning (ML) and deep learning (DL) models for prognostic assessment in patients with coronary artery disease (CAD). Our analysis of SPECT MPI-based ML/DL models revealed statistically significant superiority in predicting patient outcomes compared to stress TPD and logistic regression (LR) approaches. Furthermore, several studies demonstrated the added value of incorporating demographic data into SPECT-based ML/DL models for improved prognostic performance. These findings suggest that SPECT-MPI data can capture prognostic information beyond traditional clinical features, potentially leading to more robust risk stratification in CAD patients.

While the field of SPECT MPI-based ML/DL for CAD prognosis is budding, the current evidence highlights substantial promise. Future research should address the current limitations by promoting standardized protocols for ML/DL development and validation in the context of SPECT-MPI. Larger, multicenter studies are warranted to evaluate the generalizability and clinical utility of these models. Additionally, investigating the integration of other imaging modalities and clinical data sets with SPECT-MPI holds similar promise for further enhancing prognostic accuracy.

Abbreviations

AI

Artificial Inteliigence

CAD

Coronary Artery Disease

CCS

Chronic coronary syndromes

CNN

Convolutional neural networks

DL

Deep learning

LR

Logistic regression

MACE

Major adverse cardiovascular events

ML

Machine Learning

MPI

Myocardial Perfusion Imaging

SPECT

Single-photon emission computed tomography

TPD

Total perfusion deficit

Author contributions

All authors contributed to the study conception and design. Material preparation, data collection were performed by, M,B; A,E; Y,T,; M, L,M.; Systatistical analysis were performed by E, H,K, Cand T, C, The first draft of the manuscript was written by V, C; H, S,U, B and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work is partially supported by the NIH grants: R01-CA246704, R01-CA240639, U01-DK127384-02S1, and U01-CA268808.

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declatarations

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.