Disentangling brain heterogeneity via semi‐supervised deep‐learning and MRI: Application to dimensional representations of Alzheimer’s disease

Abstract Background Heterogeneity of neurodegenerative diseases, including Alzheimer’s disease (AD), has hampered precision diagnosis and prognosis. Machine learning methods are able to dissect neuroanatomical heterogeneity and enable identification of disease subtypes via their imaging signatures. Method We apply a novel semi‐supervised deep‐learning clustering method to derive dimensions of neurodegeneration, and investigate longitudinal paths in this dimensional system. Analyses used structural MRIs (from the Alzheimer’s Disease Neuroimaging Initiative and the Baltimore Longitudinal Study of Aging; total 2,832 participants; 8,146 scans) including controls and patients with Mild Cognitive Impairment (MCI) and AD. Result Smile‐GAN identified 4 neurodegenerative patterns/axes (Figure 1): P1, normal anatomy; P2, mild/diffuse atrophy; P3, focal medial temporal atrophy; P4, advanced neurodegeneration. Further application to longitudinal data revealed two distinct progression pathways: P1→P2→P4 and P1→P3→P4. Clinically, P2 participants showed worse performance in executive function than P3 participants (median ADNI‐EF= ‐0.29 vs. ‐0.1, with p=0.038) but had better memory (ADNI‐MEM= ‐0.13 vs. ‐0.39, with p=0.003) (Figure 2). Progression to P4 occurred more rapidly from P3 than P2, with median survival time to progression of 4.5 years and 3.2 years for P2 and P3 MCI/Dementia participants, respectively. Baseline expression of these patterns predicted the pathway and rate of future neurodegeneration, with high accuracy for determining whether P1 participants would progress and whether they would progress to P2 or P3 (Cox‐proportional‐hazard model average concordance index (CI) on the validation set of 0.8±0.02). Pattern expression offered better performance (CI=0.72±0.02) in predicting clinical conversion from MCI to Dementia compared to amyloid/tau (CI=0.69±0.02), with all three together showing modest improvement over pattern performance (CI=0.76±0.02). A composite risk score comprised of pattern probabilities, Abeta, pTau, presence of APOE E4, and ADAS‐cog score was able to predict survival time with an average concordance index, 0.78±0.02 on randomly split validation sets. (Figure 3). Conclusion Four reproducible volumetric patterns, and 2 major longitudinal pathways were revealed using deep learning, which can be determined on an individual patient basis and has implications for rate of future decline. These pattern‐based biomarkers can enrich the N‐dimension of the A/T/N framework which may inform on clinical trial enrollment and monitoring outcomes in a more personalized manner.