Multivariate analysis identifies distinct latent components and highlights the contribution of motor cortex neurochemistry to clinical variance in amyotrophic lateral sclerosis

Abstract

Abstract Amyotrophic lateral sclerosis (ALS) is marked by substantial clinical heterogeneity across motor and extramotor domains, yet how different neuroimaging modalities jointly relate to this heterogeneity remains poorly understood. Prior ALS neuroimaging studies have largely relied on univariate approaches, evaluating modalities in isolation rather than characterizing their joint covariance with clinical presentation. We applied a multivariate framework to determine which regional imaging metrics contribute most strongly to clinical heterogeneity and presentation in ALS. Multimodal neuroimaging and clinical profiles were evaluated from seventy-five ALS patients from the multicentre CALSNIC-1 dataset. Imaging metrics included magnetic resonance spectroscopy (neurometabolite ratios), diffusion-weighted imaging (diffusivity measures), T1-weighted texture analysis (autocorrelation), and cortical thickness measures sampled from the primary motor and mesial prefrontal cortices. Partial least squares (PLS) analysis, coupled with permutation testing and bootstrapping, was utilized to isolate latent variables (LVs)/components that maximally explain the joint covariance between the imaging and clinical datasets. PLS analysis identified two LVs/components surviving false discovery rate (FDR) correction, each representing a distinct covariance structure linking neuroimaging to clinical presentation. LV-1 accounted for 56.7% of the shared covariance (r = 0.54, uncorrected p = 0.002) and reflected an overall pattern of motor system integrity. Higher motor cortex neurometabolite ratios, better white matter integrity, and higher texture autocorrelation tracked with superior tapping scores, preserved forced vital capacity, higher ALSFRS-R scores, and lower upper motor neuron (UMN) burden. LV-2 captured 19.8% covariance (r = 0.51, uncorrected p = 0.002), linking higher motor and frontal autocorrelation and diffusivity measures, alongside lower frontal and motor neurometabolite ratios to greater clinical UMN burden and altered finger-tapping. Two additional components (LV-3, LV-4) identified by PLS analysis reached significance under permutation testing but did not survive FDR correction across all twelve extracted latent variables and were not interpreted further. Using a fully data-driven multivariate approach, without prior assumptions about which specific brain-behaviour relationships should exist, these findings demonstrate that multimodal neuroimaging markers do not converge on a single clinical axis but instead capture distinct latent covariance patterns mapping across motor and extramotor domains. These results highlight that neurochemical, microstructural, and structural imaging features jointly capture covariance with clinical presentation in ALS, underscoring the multidimensional nature of brain-behaviour relationships in this disease.