Computational progression subtypes of hypometabolism due to Alzheimer’s disease: towards linking antemortem and postmortem diagnoses

Abstract

AbstractBackgroundPostmortem diagnosis of Alzheimer’s and related pathologies remains the gold standard. Advanced neuroimaging techniques enable assessment of some pathologies in living subjects, but antemortem signatures are often defined post hoc in postmortem‐defined groups. Here we flip this and estimate data‐driven antemortem signatures of pathology using Subtype and Stage Inference (SuStaIn), then predict postmortem classification. SuStaIn is a machine learning algorithm that jointly estimates subtype clusters of data‐driven disease progression signatures to uniquely unravel spatiotemporal heterogeneity using cross‐sectional neuroimaging data. We evaluate SuStaIn on FDG‐PET scans to produce the first AD subtypes of hypometabolism and explore links with LATE‐NC.MethodADNI data was downloaded 06/2022. Model inputs were baseline FDG‐PET SUVR (Desikan‐Killiany parcellation) extracted using FreeSurfer from the N = 1181 (286/791/104 AD/MCI/Ab+CN) available ADNI participants having both T1w MRI and FDG‐PET at baseline. Biomarkers were converted into w‐scores relative to 203 controls (Ab–CN), adjusted for control trends in age and sex using robust linear regression. Postmortem diagnoses (N = 43) of AD/LATE‐NC/mixed were predicted. Modeling: Hypometabolism subtypes were estimated using SuStaIn. A bespoke “disease signal” approach selected informative biomarkers for w‐score events of interest: 1 (mild), 1.5 (moderate), 3 (severe). Average out‐of‐sample log likelihood (10‐fold cross‐validation) selected the final model (#subtypes, max 5). Statistical Analysis: Subtype group comparisons of cross‐sectional outcomes and postmortem pathology were performed using common statistical tests.ResultFigure 1 shows the four subtypes of hypometabolism resulting from our analysis: Typical (limbic+temporal, N = 283: 21/190/72 CN/MCI/AD), Limbic Sparing (N = 261: 19/157/85), Aggressive Typical (early high w‐scores; N = 90: 2/40/48), and Lateral Cortical (N = 62: 1/33/28), with N = 485 (41%) not subtyped. Cognition was worse in the Aggressive Typical/Lateral Cortical subtypes than the Typical/Limbic Sparing (MMSE 25 vs 27; ADAS‐Cog13 19 vs 27) Kruskal‐Wallis test p<10‐3. Figure 2 shows subtyping and staging results (59% subtypeable: N = 696), with AD cases distributed evenly within each subtype and MCI at earlier stages, as expected. Figure 3 is a confusion matrix showing that our model imperfectly predicted postmortem LATE‐NC diagnosis.ConclusionData‐driven antemortem spatiotemporal subtypes of hypometabolism in Alzheimer’s disease were imperfectly predictive of postmortem LATE‐NC/AD classification in a small sample. Validation on larger histopathology samples is necessary to make general conclusions.