Keywords: Cardiac; Vascular; Cardiovascular; Imaging; Cardiac Computed Tomography; Coronary Computed Tomography; Coronary Artery Disease; Cardiac Magnetic Resonance; CMR; Atherosclerosis; CAD; CCT; CTA; diagnosis; prediction; prognosis; Machine Learning; Artificial Intelligence; Deep Learning; Neural Network
Calcium Score, as performed by Coronary/Cardiac Computed Tomography (CT) is a quantitative method to measure the burden of coronary atherosclerosis which focuses on the calcific component (which is not the only component). It is validated and approved in guidelines for the assessment and improved risk stratification in asymptomatic individuals. In fact, it relies on a wide and established literature. For symptomatic patients it has always been not as effective.
CCT is also an established diagnostic method for the assessment of coronary artery disease (CAD) and stenosis; it has also been extensively validated and quite recently (at the ESC 2019) it has been introduced in the guidelines for the diagnosis and management of chronic coronary syndromes as a Class I tool (1).
The recent burst of research and marketing of Artificial Intelligence (AI) and derived methods has brought to everybody’s attention the potential increase in the speed of automation applied to image analysis and probability estimation. Cardiac imaging is surely one of the fields in which this push is strong. However, some considerations can be discussed.
Following the social media, the online flow of information coming from congresses, scientific societies and scientific journals’ websites, one can find several ideas and opinions on AI.
I like, for instance, the following statement: "When we raise money it’s AI, when we hire it's machine learning, and when we do the work it's logistic regression." Attributed to Juan Miguel Lavista. This citation is particularly interesting in my view because it reconciles the fact that AI is not something that appeared overnight and it is based skillwise on something we have been doing for decades. It means that AI is cool and sexy to market a startup or a fundraise but ultimately the job is similar to what we did with advanced statistical methods in the not so recent past.
Hence, it appears that there is a thin line between AI and Machine Learning (ML) as compared to Statistical Methods/Models (SM).
Frank Harrell, Professor of Biostatistics at Vanderbilt University, in his online blog, developed some definitions that I find quite useful here. He manages to provide simple clues about the differences between ML and Statistical Models (SM) using the following considerations:
- Uncertainty: SM explicitly take uncertainty into account by specifying a probabilistic model for the data.
- Structural: SM typically start by assuming additivity of predictor effects when specifying the model.
- Empirical: ML is more empirical including allowance for high-order interactions that are not pre-specified, whereas SM have identified parameters of special interest.
Prof. Harrel also summarizes what are the difference between SM and ML and, in my understanding, these can be simplified as follows: SM are suitable when there is low noise ratio in the data, when the incremental value of a variable (e.g. treatment) has to be tested, when we work with a limited number of variables, when sample size is small, and ultimately when interpretability of data is key. Instead, ML is more suitable when noise ratio is large, when samples are massive, when we do not expect one single variable to be predominant, when repeated measurements/testing can be performed, and ultimately when the fact that we obtain a black box is not an issue.
ML algorithms can become as good as the number of parameters they are fed with and the more repeated measurements and iterations, the more the model can improve; but, as we were used to “force” relationship between parameters into a linear regression fit, ML can create totally different models with multiple clustering of variables.
We all know from experience and basic science that “correlation is not causation”, even though on the news media tend to mix things up on this particular aspect; in the new era of AI we can update the concept as “association is not prediction”. This is where ML can actually help.
Al'Aref SJ et al. published on the European Heart Journal the following study: Machine learning of clinical variables and coronary artery calcium scoring for the prediction of obstructive coronary artery disease on coronary computed tomography angiography: analysis from the CONFIRM registry. Data are derived from the large CONFIRM prognostic registry (2).
This study shows that an ML model incorporating clinical features in addition to CACS can better estimate the pretest likelihood of obstructive CAD on CCT. In the case of the study from Al'Aref SJ et al. some of the aforementioned criteria are fulfilled (i.e. superlarge sample, multiple variables, no need to test the incremental value of a specific variable,…).
The Authors conclude that in clinical practice, the utilization of such an approach could improve risk stratification and help guide downstream management. The improvement is sharp from the data and it is comparable to the one reported by the CAD-CONSORTIUM previously (3,4). This is good news. However, some may also say that in everyday practice diagnosis is more important and that prediction is mostly an accessory. I agree but I must say that prediction is an accessory when it is not strongly connected to a specific diagnosis; otherwise it can be quite important. When we deal with prediction the point is the degree of improvement. Predictions are nice but we make them every single day and we honestly do not really use them so much. Why? I think it’s because we know they are not very accurate especially in extreme cases. Just think about the weather; given current technologies and resources you would guess that weather predictions are always right. In fact, they are mostly right when the weather is going to be very good or very bad; the prediction task is less reliable in between, as for many diagnostic and prognostic tests. Usual prediction methods never had the needed accuracy; therefore, we all hope that AI will be able to improve prediction performance to the level where they can actually be of clinical relevance and allow clinicians to make strong decisions.