CHARACTERISING PRESYMPTOMATIC ATROPHY PATTERNS THROUGH MULTIVARIATE MACHINE LEARNING

Abstract

Alzheimer's disease (AD) is an insidious process, with pathology slowly building up many years before clinical manifestation of the disease. Individuals who carry mutations known to cause autosomal dominant forms of AD provide an opportunity to study this preclinical phase of AD. Increased atrophy is present before onset of symptoms, but specific atrophy levels may be difficult to identify on an individual basis. Longitudinal MRI data was available for 178 participants from the Dominantly Inherited Alzheimer Network (DIAN): 62 non-carriers, 68 presymptomatic carriers and 48 affected carriers. Images were longitudinally registered, segmented and spatially normalised using SPM12 to produce maps of atrophy (or growth) for each participant. All grey matter voxels from the resulting maps of the non-carriers and affected carriers were used to train a support vector machine (SVM) classifier. Presymptomatic individuals were then used as independent testing data, with the SVM classifier generating a global probability score that represented how similar an individual's atrophy pattern was to those observed in the affected carriers. The resulting probabilities were clustered into four groups: from the highest probability (>75%) to the lowest probability (<25%) of being classified as symptomatic. Presymptomatic carriers within five years to expected onset were more likely to be classified as symptomatic (Figure). Of the 26 presymptomatic carriers within five years to onset, 16 were classified as symptomatic with a probability greater than 50%. 75% of the presymptomatic carriers in the highest probability group and 37.5% in the next highest group progressed to a global CDR score of 0.5 or greater at a subsequent follow-up scan (range of duration before progression=1.0–4.2 years). Conversely, only one of the ten presymptomatic carriers within five years to expected onset and classified in the lower two probability groups progressed to a higher CDR score at follow-up.