How reproducible are data-driven subtypes of Alzheimer's disease atrophy?

Alzheimer’s disease (AD) exhibits substantial clinical and biological heterogeneity, complicating efforts in treatment and intervention development. While new computational methods offer insights into AD progression, the reproducibility of these subtypes across datasets remains understudied, particularly concerning the robustness of subtype definitions when validated on diverse databases. This study evaluates the consistency of AD progression subtypes identified by the Subtype and Stage Inference (SuStaIn) algorithm using T1-weighted MRI data across 5,444 subjects from ANMerge, OASIS, and ADNI datasets, forming four independent cohorts. Each cohort was analyzed under two conditions: one using the full cohort, including cognitively normal controls, and another excluding controls to test subtype robustness. Results confirm the three primary atrophy subtypes identified in earlier studies: Typical, Cortical, and Subcortical, as well as the emergence of rare and atypical AD variants such as posterior cortical atrophy (PCA). Notably, each subtype displayed varying robustness to the inclusion of controls, with certain subtypes, like Subcortical, more influenced by cohort composition. This investigation underscores SuStaIn’s reliability for defining stable AD subtypes and suggests its utility in clinical stratification for trials and diagnosis. However, our findings also highlight the need for improved dataset diversity, particularly in terms of ethnic representation, to enhance generalizability and support broader clinical application.

💡 Research Summary

Alzheimer’s disease (AD) is characterized by substantial clinical and biological heterogeneity, which hampers therapeutic development and trial design. While data‑driven disease progression models have been proposed to disentangle this heterogeneity, the reproducibility of the subtypes they generate across independent datasets has not been systematically examined. In this study, the authors evaluated the robustness of the Subtype and Stage Inference (SuStaIn) algorithm—a probabilistic, unsupervised model that simultaneously infers disease stages and subtypes—using T1‑weighted MRI data from four distinct cohorts derived from the ANMerge, OASIS, and ADNI repositories. The cohorts comprised (1) ANMerge (n = 931), (2) OASIS (n = 1,038), (3) a combined ADNI 1.5 T + ANMerge set (n = 2,084), and (4) a combined ADNI 3 T + OASIS set (n = 1,391). For each cohort, two modeling conditions were created: one that included cognitively normal controls (defined by Clinical Dementia Rating = 0) and one that excluded these controls, yielding eight SuStaIn experiments in total.

Data preprocessing followed the original SuStaIn pipeline: FreeSurfer (versions 5.1–5.3) was used to extract volumes or cortical thicknesses for 14 regions of interest (hippocampus, amygdala, nucleus accumbens, insula, cingulate, caudate, pallidum, putamen, thalamus, entorhinal cortex, and frontal, temporal, occipital, parietal lobes). Left‑right averages were taken after confirming symmetry. Linear regression removed age, sex, and intracranial volume effects, after which each metric was Z‑scored relative to the control group. The sign of the Z‑scores was flipped so that larger values corresponded to greater atrophy, matching the linear Z‑score event model used by SuStaIn.

SuStaIn models were fit using a hierarchical clustering approach with a linear Z‑score progression model. The number of subtypes was selected via 10‑fold cross‑validation, and for each model the algorithm estimated the most probable sequence of regional Z‑score events, the proportion of subjects belonging to each subtype, and each subject’s most likely disease stage. Event thresholds of 2 σ and 3 σ were employed (the 1 σ threshold was omitted to avoid spurious low‑severity events).

Across all cohorts and both control‑inclusion conditions, three primary atrophy subtypes emerged consistently:

1. Typical AD – a pattern dominated by hippocampal, temporal, and frontal atrophy, mirroring the classic AD trajectory reported in neuropathology.
2. Cortical AD – widespread cortical thinning across frontal, temporal, parietal, and occipital lobes, reflecting a more diffuse neurodegenerative process.
3. Subcortical AD – prominent involvement of basal ganglia, thalamus, and brainstem structures, a pattern previously described in post‑mortem studies of atypical AD.

In addition, rare variants resembling posterior cortical atrophy (PCA) were identified in a minority of subjects, demonstrating SuStaIn’s capacity to capture low‑prevalence phenotypes. Importantly, the three main subtypes were robust to the inclusion or exclusion of controls, whereas the Subcortical subtype showed greater sensitivity to cohort composition; its prevalence and event ordering shifted noticeably when controls were omitted, suggesting that this subtype relies more heavily on the reference distribution defined by healthy subjects.

The authors also examined technical factors such as magnetic field strength (1.5 T vs 3 T) and differences in FreeSurfer versions. By harmonizing ROI definitions and confirming that version‑specific processing did not materially affect volume/thickness estimates, they showed that these scanner‑related variables did not undermine subtype reproducibility.

Limitations acknowledged include: (i) the ethnic homogeneity of the three databases (predominantly White/European participants), limiting generalizability to more diverse populations; (ii) reliance on CDR = 0 as the sole control definition, which may be less specific than biomarker‑based criteria (e.g., CSF β‑amyloid negativity); (iii) the cross‑sectional nature of the data, which precludes direct observation of longitudinal progression; and (iv) the restricted set of 14 ROIs, which may omit relevant network‑level information.

Overall, the study provides strong empirical evidence that SuStaIn can reliably recover established AD atrophy subtypes across heterogeneous datasets, field strengths, and modeling conditions. This reproducibility supports the algorithm’s potential for clinical trial stratification, personalized diagnostic pipelines, and mechanistic investigations of AD heterogeneity. Future work should incorporate multi‑ethnic, longitudinal cohorts and multimodal biomarkers (e.g., PET, CSF) to further validate and refine subtype definitions, thereby enhancing the translational impact of data‑driven disease modeling in Alzheimer’s research.