Are neural networks the ultimate risk prediction models in patients at high risk of acute myocardial infarction?

Identifying and phenotyping patients at risk of developing major cardiovascular events is an ongoing research priority. With an abundance of variables and confounders, the current prediction tools based on linear multivariate regression models are becoming outpowered by the emerging machine learning algorithms. These algorithms have the potential to identify, with a higher prediction power healthy patients at risk, prompting preventive interventions, such as early screening imaging tools or medical therapy (e.g. antiplatelet therapy).1 There is evidence that the development of acute myocardial infarction (AMI) is associated with increased mortality, acute kidney injury, hospital stay and increased costs.2 These adverse outcomes are further augmented by the association of cardiogenic shock.3 Machine learning aims to address important clinical questions, inherently difficult to tackle through the classic approach. One of the most common algorithms used is pattern recognition. An eloquent example is the diagnosis of obstructive coronary disease on computed tomography angiography.4 Another type of machine learning is reinforcement learning, which is reward-based learning, while some algorithms are focused on discriminative learning rather than generative learning. Machine learning algorithms frequently employed in practice include linear and logistic regression, artificial neural networks, support vector machines and tree-based methods. When training neural networks, an optimal fit is usually desired, but under-fitting and over-fitting may be present due to the selection of data for inclusion in the analysis. Datasets used in machine learning projects are typically partitioned into: the training set, which encompasses the bulk of all available data, is used for the primary development of the model; the validation set is used to estimate overall model performance or to fine tune its ‘hyperparameters’. The performance benchmarks and metrics are problem dependent and are tested against existing dataset benchmarks. The implementation of the machine learning model within a clinical setting could help automate the process of selecting appropriate candidates for further diagnostic evaluation, while circumventing more cumbersome routine clinical steps. The current evidence suggests that neural networks can find hidden features in input patterns that are not visible by conventional statistical methods. The current European guidelines on cardiovascular disease prevention in clinical practice5 recommend (class I, level C) the systematic risk assessment for cardiovascular disease in individuals at high risk, and the repeat of this assessment every 5 years (class I, level C). Patients who have undergone previous CABG are considered at very high risk.5 Furthermore, the use of validated scores and strategies for the prevention of cardiovascular disease is recommended (class I, level C). Existing scores (e.g. GRACE) have a limited discriminatory power. The role of neural networks may be to address the need for a powerful discriminatory risk stratification tool, which is an essential requisite of management in these patients. These have been shown to have a superior performance when compared to established risk prediction algorithms (e.g. American Heart Association/American College of Cardiology).6 This may be explained through the inclusion of emerging novel biomarkers in the machine learning algorithms, which are not routinely included in existing risk scores. There is emerging evidence that neural network models are useful in predicting mortality after coronary artery bypass grafting (CABG),7 the need for revascularisation,8 or the detection of arrhythmias.9 Machine learning has recently been shown to be superior to predict cardiovascular or all-cause mortality by combining clinical or imaging modalities when compared with clinical and imaging metrics alone.6,10 This has an added importance in the context of emerging standard electronic health records, which may be used following consent to predict adverse outcomes. One of the inherent limitations of machine learning models is the high number of variables and interactions, which may preclude the identification of specific therapeutic targets. Another limitation of neural networks may be the lack of standardisation across datasets, which hinders the selection of optimal models for clinical integration. This may be mitigated by regular and systematic reviews of these algorithms by humans to avoid traditional biases. One of the ongoing ethical considerations is the accountability of the output generated through machine learning. In conclusion, neural networks have shown promising utility in improving risk stratification and guide management, but there is still a lack of understanding on how and the extent to which these tools can be integrated in clinical practice. The scientific community is excited to learn of further developments and the implementation of these ultimate risk stratification tools.