On using machine learning algorithms to define clinically meaningful patient subgroups

Improved taxonomy will drive our efforts to personalise medicine over time. Ideally improved taxonomy is fueled by our detailed insight in pathogenesis leading to subgrouping syndrome’s into more homogeneous diseases. An alternative is to cluster subgroups of patients based on similar manifestations and prognosis. So, the recent publication of Spielman et al 1 as well as the correspondence on that study written by Pinal-Fernandez and Mammen2 is very timely and interesting. Spielman et al 1 identified three clinical clusters in patient with anti-Ku-positive myositis by applying hierarchical clustering analysis on both clinical and biological features. Pinal-Fernandez and Mammen suggest that the results of Spielman’s work might be flawed as they disagree with the method of number of cluster selection. Indirectly they also challenge the idea of using (hierarchical) clustering techniques to identify clinically meaningful patient populations. Of course, we agree that improper use of analytical methods can lead to incorrect conclusions. Therefore, we applaud the ongoing discussion on how to reliably use ‘big-data techniques’ partly fueled by EULAR’s point to consider for the use of …