Investigating data-driven biological subtypes of psychiatric disorders using specification-curve analysis

Area of Science:

• Psychiatric research
• Data science
• Biomarker discovery

Background:

• Cluster analysis is widely used for data-driven classification in psychiatric research.
• Methodological variations in cluster analysis can compromise the generalizability and replicability of findings.
• Specification-Curve Analysis (SCA) was employed to assess the impact of methodological choices on biomarker-based clustering.

Purpose of the Study:

• To investigate the influence of methodological variations on cluster analysis results using proteomics data from psychiatric patients.
• To evaluate the consistency of clustering results across numerous analytical specifications.
• To understand how data properties, such as cluster number and noise levels, affect SCA outcomes.

Main Methods:

• Proteomics data from 688 patients and 426 healthy controls were analyzed using 1200 k-means and hierarchical clustering combinations.
• Specification-Curve Analysis (SCA) assessed consistency across varying clustering algorithms, fit-indices, and distance metrics.
• Simulated datasets with different cluster numbers and noise levels were used to evaluate data property effects on SCA.

Main Results:

• Real data SCA did not identify robust biological clustering patterns in major depressive disorder (MDD) or combined MDD/healthy datasets.
• Simulations indicated that while the correct number of clusters could be identified consistently, accuracy decreased with more clusters and higher noise levels.
• The study found that SCA results can be inconsistent with complex, noisy real-world biomarker data.

Conclusions:

• SCA offers insights into potential clusters within biomarker data.
• In complex, noisy biomarker datasets, SCA results may be highly dependent on model specification.
• The findings suggest that conclusions about biological clusters in psychiatric patients should be drawn cautiously due to potential methodological influences.