Clusterization of Behavioral and Psychological Symptoms of Dementia as Assessed by Neuropsychiatric Inventory: A Case Against the Use of Principal Component Analysis.

Background: The term Behavioral and Psychological Symptoms of Dementia (BPSD) covers a group of phenomenologically and medically distinct symptoms that rarely occur in isolation. Their therapy represents a major unmet medical need across dementias of different types, including Alzheimer's disease. Understanding of the symptom occurrence and their clusterization can inform clinical drug development and use of existing and future BPSD treatments.

Objective: The primary aim of the present study was to investigate the ability of a commonly used principal component analysis to identify BPSD patterns as assessed by Neuropsychiatric Inventory (NPI).

Methods: NPI scores from the Aging, Demographics, and Memory Study (ADAMS) were used to characterize reported occurrence of individual symptoms and their combinations. Based on this information, we have designed and conducted a simulation experiment to compare Principal Component analysis (PCA) and zero-inflated PCA (ZI PCA) by their ability to reveal true symptom associations.

Results: Exploratory analysis of the ADAMS database revealed overlapping multivariate distributions of NPI symptom scores. Simulation experiments have indicated that PCA and ZI PCA cannot handle data with multiple overlapping patterns. Although the principal component analysis approach is commonly applied to NPI scores, it is at risk to reveal BPSD clusters that are a statistical phenomenon rather than symptom associations occurring in clinical practice.

Conclusions: We recommend the thorough characterization of multivariate distributions before subjecting any dataset to Principal Component Analysis.