Study results in context
A criticism of ML models in the past has been their lack of
interpretability as the predictors that drive model output are unknown.
Kachroo et. al. have addressed this limitation by providing
analysis of the important model predictors, thus providing novel
mechanistic and clinical insights. In this study, the authors’ feature
importance analysis reinforces the notion that each predictor on its own
may not contain much predictive power, but, the combination of
predictors, and the use of a powerful tool, like ML, can generate
predictive power. When each predictor in this study is placed in order
of importance, up to 20 predictors have up to 20% relative importance
to the most important predictor and include features like hypertension
and diabetes that when evaluated alone contain minimal predictive
power.
Despite the lack of echocardiographic measurements of cardiomyopathy in
the STS dataset, there were some novel findings. In this paper, of the
five most important predictors of recurrent MR, three of them (bypass to
the ramus, obtuse marginal II and diagonal II) represent
revascularization targets. Perhaps these represent the geometric changes
that take place due to ischemia and ventricular remodelling, that result
in leaflet tethering and mitral annular dilation and circularization
[1]. These measures are consistent with previously identified
markers of recurrent MR such as basal inferior aneurysms or dyskinesia.
This suggests that certain patterns of myocardial ischemia may be more
likely to suffer from recurrent MR post repair. Additionally, these
insights can be used to deliver surgical strategies that reduce risk of
recurrent MR. For example, as a bypass to the ramus is an important
predictor of recurrent MR, surgeons would be more vigilant about
revascularization of the anterior wall. The use of beta blockers also
was one of the five most important predictors and is likely collinear
with NYHA III/IV symptoms and represents symptomatic reduced ejection
fraction heart failure. In this setting, it is likely that patients that
were not in beta blockers were at higher risk for recurrent MR. This
shines light on the role of ventricular reverse remodelling and the role
of ongoing optimal guideline directed medical therapy in order for valve
repair to be durable.
In summary, the great benefit of the STS database is that these
predictors are easy to obtain and do not rely on patient reported
outcomes or subjective and expensive echocardiographic assessment of the
heart. In this way, Kachroo et. al . have succeeded in using
universally accessible, objective measures that are routinely collected
at a national level to identify patients that receive durable mitral
valve repairs. This tool can be used by all surgeons for all patients
using only clinical demographics and a revascularization plan. As the
amount of data collected in medicine continues to grow and become more
digitized and multidimensional, ML will be the tool that can amalgamate,
analyze and interpret this high dimensionality data. ML and its
counterpart, Artificial Intelligence, will allow for the integration of
datasets, like the STS, with raw imaging data, such as echocardiograms
and cardiac MRI, with waveform data, such as electrocardiograms, with
genomic data for highly personalized care (Figure). This study by
Kachroo et. al. is the first building block in this pursuit,
showing that high prediction accuracy can be achieved from routinely
collected STS variables, and can only be improved as datasets continue
to grow in dimensionality. We can look forward to a rich future in data
and data analytics and using ML with the STS database can prove to be
beneficial for clinicians, scientists and patients.