Artificial Intelligence in Cardiology: What's New?
Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. In this episode, we will discuss the groundbreaking impact that machine learning and deep learning technologies have had on the field of cardiology. From the update of mortality risk scores in cardiovascular disease to the skyrocketing growth it has meant for Nuclear Cardiology, and how artificial intelligence promises to revolutionize cardiovascular health care.
Did you know that non-ST-segment elevation acute coronary syndromes account for approximately three-quarters of acute coronary syndrome cases in women?1
The Global Registry of Coronary Events score, known as GRACE score, is currently used in both male and female patients. This score is used to estimate the risk of mortality from electrocardiographic and biochemical variables in patients with acute coronary syndromes, or ACS. However, several studies have questioned the adequacy of this score in women with non-ST-segment elevation acute coronary syndromes, since it was initially developed and tested in a predominantly male population.1
We know that women with non-ST-segment elevation-ACS have a different set of plaque characteristics and that they have a higher prevalence of plaque erosion as the primary causal mechanism. In addition, they tend to present with the disease at an older age, have a higher burden of comorbidities, have longer delays in going to the hospital, and have a higher risk of unadjusted mortality risk than men. However, coronary angiography and early invasive treatment tend to be less frequently performed in women.1
Let's examine how artificial intelligence, or AI, was recently applied to update the GRACE score. The research team led by Drs. Wenzl and Kraler developed a machine learning-based risk score for in-hospital mortality in non-ST-segment elevation-ACS patients, called GRACE 3.0, that considers sex differences in disease characteristics.1
The researchers found that female patients had significantly reduced renal function, as indicated by a lower estimated glomerular filtration rate, or eGFR. Creatinine is a proxy for the renal function used by the GRACE score, without considering the different physiological ranges between women and men. However, creatinine concentrations were lower in female patients. In addition, female patients were more likely to have signs of congestion, had higher levels of N-terminal prohormone of brain natriuretic peptide, and did not have higher rates of cardiogenic shock.1
Examining the performance of the GRACE 2.0 score, which is the currently used version, in 420,781 patients with non-ST-segment elevation acute coronary syndromes from four European countries revealed limited discriminatory performance, suboptimal calibration, and underestimation of mortality risk in female patients with different baseline risk profiles; thereby promoting a systematic sex bias in early risk stratification and guideline-directed care.1
Although the study was limited to populations in the United Kingdom and Switzerland, the sex-based performance difference was evident across geographic boundaries. This finding suggests that knowledge of sex differences in disease biology and patient risk profile at presentation is critical to improving outcomes for all patients with non-ST-segment elevation-ACS.1
So, how does this newly proposed GRACE 3.0 score work?
The models for predicting in-hospital mortality were based on the GRACE variables and developed with sex-disaggregated data. The 8 variables included were: age, heart rate, systolic blood pressure, Killip class, creatinine concentration, cardiac arrest, presence of ST-segment deviation, and troponin elevation. A supervised machine-learning approach was applied, called ensemble learning, to capture potential non-linear relationships between patient characteristics and mortality.1
The importance of clinical features in the GRACE 2.0 in-hospital mortality prediction model differed in regression-based analyses in female and male patients with non-ST-segment elevation acute coronary syndromes, thus suggesting that sex-specific weighting of GRACE components improves overall model performance.1
The GRACE 3.0 score yielded an area under the curve of 0.89 in male patients and 0.86 in female patients in the training cohort, while in the internal validation cohort the area under the curve was 0.88 in males and 0.84 in females. When applied to the external validation cohort, the area under the curve was 0.91 in male patients and 0.87 in female patients. The power of discrimination of in-hospital death was better with the GRACE 3.0 score than with the 2.0 version in both validation cohorts, irrespective of sex.1
By applying a machine learning algorithm to these features in sex-disaggregated cohorts, the investigators developed and validated the GRACE 3.0 score, which separately predicts in-hospital mortality in women and men with non-ST-segment elevation-ACS. Gender-specific risk estimates from this new score led to the reclassification of women into the high-risk group and men into the low- to intermediate-risk group. With the implementation of this model, personalized treatment could be optimized and structural inequalities in managing non-ST-segment elevation acute coronary syndrome patients could be overcome, opening the doors to future management implications.1
Cápsula
In the last 50 years, nuclear cardiology imaging has evolved and developed enormously. Thanks to these advances, single photon emission computed tomography myocardial perfusion imaging, known as SPECT MPI, is the most widely performed noninvasive technique for detecting coronary artery disease, which is the most frequent cause of death in the developed world.2
A key aspect for the widespread clinical usage of nuclear imaging modalities has been the continuous development of automated processing, quantification, and visualization algorithms and techniques. Thus, highly efficient AI tools such as machine learning, deep learning, and, in particular, convolutional neural networks, have demonstrated great effectiveness in imaging applications. These tools have already been applied to various tasks, such as image quality enhancement, image segmentation, diagnosis and, especially risk stratification and outcome prediction.2
Cierre de cápsula
Welcome back. In the previous segment, we explored how artificial intelligence can help improve existing risk-prediction tools by identifying biases that miscalculate the risk in an essential subset of patients. Now, we will examine how this technology is revolutionizing a field with an essential diagnostic and prognostic value: nuclear cardiology imaging.2
Shiri and collaborators proposed a convolutional neural network method to reduce by half either the acquisition time or the number of projections for SPECT MPI, consequently reducing the radiation dose. The convolutional network achieved a better image quality in half-time projections, this result was not observed when the number of projections was halved.2
Another study in 2020, by Ramon and collaborators, sought to eliminate the high image noise typical of low-dose acquisitions in 1,052 SPECT MPI datasets by testing two common reconstruction techniques. The team of researchers achieved significantly improved noise reduction while increasing diagnostic accuracy through a deep learning method, even at lowered dose levels.2
Furthermore, artificial intelligence algorithms have been applied to improve test safety. Stress-only protocols have been shown to reduce effective radiation exposure by up to 60 percent; however, they remain largely underused.2
In 2020, a machine learning-based algorithm was submitted by the team led by Lien-Hsin Hu for risk prediction using only stress myocardial perfusion imaging and clinical data. The algorithm was designed to select low-risk patients who could benefit from the cancellation of the rest-test. The results showed that the machine learning algorithm's cancellation recommendations achieved lower rates of major adverse cardiovascular events, or MACE, than those obtained when the recommendation was made based on physician’s interpretation or clinical guidelines.2
AI can also improve attenuation correction with low-dose computed tomography scanning. There is ample evidence that the use of attenuation correction can increase accuracy and reduce ambiguity, yet it is underutilized due to its cost and low availability.2
In 2021, Luyao Shi and researchers successfully developed a feasible method for generating reliable attenuation correction maps directly from SPECT images, achieving synthetic maps that are qualitatively and quantitatively consistent with the true maps.2
Another area where AI is making a breakthrough is image segmentation and interpretation. For example, many left ventricular segmentation algorithms have been proposed previously for a reliable assessment of myocardial perfusion and function; however, their robustness has notably increased thanks to more recent technological advances.2
Another example is a team that used machine learning techniques in the automated identification of the valve plane from SPECT images, which included expert knowledge and high-level image features related to valve plane location, shape, and intensity, as well as anatomical variations.2
This fully automated method exhibited a diagnostic performance equivalent to that obtained after correction by experienced readers. Similarly, another team developed a model for delineating left ventricular epicardial and endocardial profiles from gated SPECT images, the model was trained with left ventricular contours delineated by clinicians. Preliminary results demonstrated excellent accuracy in myocardial volume assessment.2
Thus, image-based diagnosis and outcome prediction are possibly the areas of nuclear cardiology that have seen the highest progress thanks to artificial intelligence techniques.2
Let us now discuss the results achieved by Arsanjani and team, who combined several quantitative variables to detect coronary artery disease from a cohort of 957 SPECT studies, comparing the diagnostic accuracy of a support vector machine model with that of visual interpretation performed by two expert readers. The algorithm showed better diagnostic accuracy than the readers.2
In a follow-up investigation, the authors obtained additional improvements when quantitative perfusion measures and clinical variables were combined. This could become a fundamental aid to a diagnostician in the future, especially in less experienced centers, as interpretation is highly dependent on the readers’ experience. Also, the integration of numerous heterogeneous variables can be challenging.2
One more example of this promising technology is the findings of Hu and collaborators, who used a machine learning algorithm to predict per-vessel early revascularization within 90 days from SPECT MPI.2
A good number of patients scheduled for invasive angiography ultimately do not have obstructive disease. Thus, predicting the need for revascularization is clinically important in the evaluation of coronary artery disease. The investigators used 18 clinical variables and 28 imaging features to feed the model. The algorithm outperformed current quantitative methods on a per-patient and per-vessel basis, as well as expert interpretation.2
Moving on, we will now explore the influential role that AI, and particularly machine learning, plays in the field of medicine today, given that it is very likely that, soon nearly every physician will be using machine learning technology, either consciously or by its routine integration into medical software. Hence, both physicians and scientists will need to have a basic understanding of this technology to enable a safe and conscious use in daily practice.3
Machine learning is primarily focused on providing optimal predictions and often sacrifices interpretability on several fronts that can become impenetrable in neural networks, which can be difficult to accept, as we humans are used to understanding linear associations between variables.3
Machine learning models can be broadly divided into unsupervised and supervised approaches. While unsupervised ones focus on discovering connections between variables, supervised models aim to achieve optimal prediction from labeled samples.3
Supervised algorithms are seeing a rapid rise in medical science. Yet, they are not the perfect solution to every unsolved clinical problem and, like any other statistical model, are limited by the quality and magnitude of the dataset from which they are trained. Still, they are an optimal choice when it comes to providing probabilities.3
Machine learning-based predictive algorithms now make it possible to better model the intricate workings of physiology and pathophysiology alike, presenting very promising opportunities in the cardiovascular field.3
Take cardiovascular outcome prediction, for example, which, despite numerous established clinical risk scores, remains challenging at the individual level. Machine learning algorithms have not only shown improved long-term predictive capabilities for specific biomarkers but have also shown superiority in short-term prediction of mortality among patients with cardiogenic shock, which can lead to treatment intensification in high-risk patients.3
Beyond optimizing the treatment of cardiovascular diseases, these algorithms can also help to understand their pathogenesis and underlying mechanisms. Thus, machine learning has been commonly used in genome-wide association studies, abbreviated as GWAS. Several investigations have successfully applied them to predict the incidence of hypertension using polygenic risk factors, to predict advanced coronary calcium, inherited heart disease, and to predict type II diabetes in a multiethnic cohort.3
Another interesting approach with potential future applications is the so-called “liquid biopsy”, a minimally invasive technology that has been established in oncology to detect and characterize circulating molecular biomarkers of tumor origin without the need for invasive biopsy. Just like in oncology, this technology could one day help to characterize cardiovascular diseases such as heart failure, acute myocarditis, or coronary artery disease, without further invasive testing.3
Similarly, machine learning can also be employed to identify previously unnoticed patterns by conventional approaches. For example, a recent deep learning framework has been shown to be able to detect structural variations in human genome data, capturing a set of fusion genes that whole-genome sequencing had missed. This approach has been successfully applied to single-cell and mass genomics to study the heterogeneity of structural variation in primary tumors from Hi-C maps.3
These advances may allow us to broaden our understanding of biological processes. For example, the work of Clerx and collaborators in ion channel research, where they investigated the effects of mutations in the SCN5A gene on clinical phenotypes with a two-step machine learning model. The researchers achieved improvements in predicting the functional properties of sodium channels and outperforming traditional approaches.3
Although the application of machine learning and deep learning tools in cardiovascular medicine can facilitate precision medicine, clinicians are often reluctant to use AI methods mainly due to a poor understanding of the underlying predictive mechanisms. However, interpretability can mean different things to clinicians, scientists, and even patients, and interpretability itself can also be measured at different levels. For example, Van der Schaar's three proposed types of interpretability for machine-learning models:3
● The first is to show the explanatory characteristics of the patients, in other words, to highlight which characteristics are taken into account in the model and how they are weighted.
● The second type suggests grouping patients according to similarities regarding predicted outcome. This allows to identifiy specific high-risk or low-risk phenotypes.
● The third type recommends identifying the rules and laws that cause the model to significantly alter a decision. This can help narrow the model down to the most important decisions.
Although the first and second types are already a practiced reality and are used in many investigations, they do not provide visibility into the causal relationship between the outcome and the variables of interest. However, currently, there is no valid method to establish the third type, which may be the most sought-after approach in clinical practice.3
It is indispensable to keep in mind that AI-based approaches are not always the optimal solution, either in the clinical or in laboratory settings. However, as development continues, it will be essential that clinicians apply structured criteria to be able to interpret the algorithm recommendations, and meticulously validate those in prospective clinical trials.3
In this regard, Meskó and colleagues have proposed a list of questions that can help assess the applicability of artificial intelligence models in clinical practice, including:3
● Questioning the relevance of the model. For instance, the incremental value the model may offer over other non-AI approaches, as well as the actionable clinical consequences it may have.
● Questioning the validation methods and sample sizes on which the model was developed and trained.
● Exposing potential selection biases, including the degree of representativeness of the investigated cohort for the patient population in which the method will be applied and its validity in real-world situations.
● Examining the labeling methods followed by the developers, including definitions, standardization, datasets, and updating with gold standards.
● Inquiring about adherence to the most current guidelines.
● Questioning about transparency, including the public availability status, potential conflicts of interest, and derived decision processes of the model.
All these considerations deserve special attention when introducing these technologies into clinical practice and even trials. After all, we must be able to trust the answers the algorithm provides, even if we cannot understand the individual calculations it performs.3
Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode.
Discover more medical news on Viatris Connect!
Referências:
- 1. Wenzl FA, Kraler S, Ambler G, Weston C, Herzog SA, Räber L, et al. Sex-specific evaluation and redevelopment of the GRACE score in non-ST-segment elevation acute coronary syndromes in populations from the UK and Switzerland: a multinational analysis with external cohort validation. Lancet. 2022; 400(10354):744-756. Available at: https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(22)01483-0/fulltext
- 2. Garcia EV, Piccinelli M. Preparing for the Artificial Intelligence Revolution in Nuclear Cardiology. Nucl Med Mol Imaging. 2023;57(2):51-60. Available at: Preparing for the Artificial Intelligence Revolution in Nuclear Cardiology
- 3. Kresoja KP, Unterhuber M, Wachter R, Thiele H, Lurz P. A cardiologist's guide to machine learning in cardiovascular disease prognosis prediction. Basic Res Cardiol. 2023;118(1):10. Available at: A cardiologist's guide to machine learning in cardiovascular disease prognosis prediction
Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.
NON-2023-9468
NON-2023-2512