BACKGROUND: The heterogeneity of neurodegenerative diseases challenges accurate clinical

diagnosis and clinico-pathological correlation. Traditional univariate methods

often fail to capture the complex relationships between multi-domain clinical

symptoms and distinct neuropathological burden.

 

METHOD: This study utilises data from the National Alzheimer Coordinating Center, a

tertiary referral real-world database comprising clinical and post-mortem

neuropathological data from both cognitively unimpaired and symptomatic

participants. General characteristics and clinical (behavioural, cognitive, and

functional) features documented at the final visit prior to death were jointly

analysed with pathological findings. A multi-block sparse Partial Least Squares

(sPLS) analysis was employed to identify latent clinico-pathological profiles

sharing variance between clinical and pathological blocks. The significance,

stability, predictive performance, and clinical validity were assessed.

RESULTS: Data preprocessing yielded a final sample of 4,951 participants with 40 clinical

and 34 pathological features. The analysis identified distinct clinico-pathological

profiles, with the first three yielding the highest correlations (range, r =

0.41−0.52): Alzheimer’s disease (profile 1), Lewy body disease (profile 2), and

frontotemporal dementia (profile 3). Clinically, cognitive impairments and the

expression of neuropsychiatric symptoms most markedly separated the profiles.

Additionally, demographic factors (age, sex, living arrangements) modulated the

clinical presentation in profiles 2 and 3, but not profile 1. Pathologically, pontine

locus coeruleus degeneration emerged as a disease-specific rather than generalised

feature, contributing prominently to profiles 1 and 2 only, despite their divergent

clinical presentations. All profiles were independent of vascular factors. With

significance, stability, and consistent feature selection frequencies being

confirmed, the profiles' prevailing agreement with the clinician-assigned

diagnosis validated the identified disease axes.

 

CONCLUSION: This unbiased sPLS approach autonomously reconstituted three distinct

neurodegenerative entities from raw variance, independent of prior clinical labels.

By integrating divergent neuropsychiatric symptoms and demographic modulators

with their pathological drivers, this data-driven model captures disease complexity

more accurately than traditional clinical classifications, potentially enabling earlier

diagnosis and treatment.