Clinical subtyping using community detection: Limited utility?

Area of Science:

• Psychiatric research
• Computational psychiatry
• Statistical modeling

Background:

• Researchers are increasingly using community detection for psychiatric subtyping.
• This method lacks extensive validation in psychiatric literature compared to traditional clustering.
• Existing community detection algorithms may not be optimized for typical psychiatric data characteristics.

Purpose of the Study:

• To clarify the utility of community detection for psychiatric subtype discovery.
• To compare community detection algorithms with traditional methods in simulated psychiatric settings.
• To guide researchers in selecting appropriate subtyping methodologies.

Main Methods:

• Introduction to community detection algorithms, distinguishing correlation-based from distance-based approaches.
• Simulation study using three distinct subtype difference conceptualizations relevant to psychiatry.
• Comparison of community detection results against traditional methods like latent profile analysis.

Main Results:

• Several community detection algorithms inaccurately recovered the number of subgroups.
• Correlation-based community detection outperformed distance-based methods, particularly with smaller sample sizes.
• Latent profile analysis demonstrated greater consistency in identifying subtypes.
• Method success was contingent on the nature of simulated subtype differences.

Conclusions:

• Traditional methods, such as latent profile analysis, remain reliable choices for psychiatric subtyping.
• Subtyping conclusions are influenced by underlying assumptions and theoretical frameworks.
• Researchers should employ multiple subtyping methods to assess the robustness and dependency of findings.