Clinical subtyping using community detection: Limited utility?

Abstract

Objectives: To discover psychiatric subtypes, researchers are adopting a method called community detection. This method was not subjected to the same scrutiny in the psychiatric literature as traditional clustering methods. Furthermore, many community detection algorithms have been developed without psychiatric sample sizes and variable numbers in mind. We aim to provide clarity to researchers on the utility of this method.

Methods: We provide an introduction to community detection algorithms, specifically describing the crucial differences between correlation-based and distance-based community detection. We compare community detection results to results of traditional methods in a simulation study representing typical psychiatry settings, using three conceptualizations of how subtypes might differ.

Results: We discovered that the number of recovered subgroups was often incorrect with several community detection algorithms. Correlation-based community detection fared better than distance-based community detection, and performed relatively well with smaller sample sizes. Latent profile analysis was more consistent in recovering subtypes. Whether methods were successful depended on how differences were introduced.

Conclusions: Traditional methods like latent profile analysis remain reasonable choices. Furthermore, results depend on assumptions and theoretical choices underlying subtyping analyses, which researchers need to consider before drawing conclusions on subtypes. Employing multiple subtyping methods to establish method dependency is recommended.

Keywords: community detection; hierarchical clustering; k-means; latent profile analysis; psychiatric subtypes.

FIGURE 1

Illustration of how different definitions of similarity can lead to different communities with the same data. In panel (a), example profiles are displayed for three participants (Ali, Bobby, and Charu) who have obtained scores on 10 continuous variables. In panel (b), the distances between participants' scores from panel (a) are displayed for each pair of participants. Note that in practice, multivariate distances would be used, but here all 10 univariate distances are displayed side‐by‐side for illustrative purposes. In panel (c), the correlations between participants' scores from panel (a) are displayed for each pair of participants. In panels (b and c), communities are enclosed in red. Positive similarities are displayed in green, negative similarities in dark blue.

FIGURE 2

Illustration of the three types of simulated data structure. In panels (a and b), the points refer to different variables; in panel (c) the points refer to participants. In panel (a), the lines between points are illustrative, to show that points are from the same subtype. In panel (b), the lines denote correlations between variables. In panel (c), the lines denote correlations between participants. Positive correlations are displayed in green, negative correlations in dark blue.

FIGURE 3

Simulation results. For each true number of subtypes on the y‐axis, the distribution over the recovered number of subtypes on the x‐axis is plotted. The size of the dots indicates the percentage of the 1000 simulations where that number of subtypes was recovered. If recovery is perfect, all five dots are on the diagonal on the left of the plot, as in the latent profile analysis/covariance of participants panel.

FIGURE 4

Recommended methods as derived from the simulated datasets. C, correlation‐based; D, distance‐based; LPA, latent profile analysis