Diagnostic classifications based on cognitive, CSF, and PET biomarkers across ADNI cohorts: cross-sectional and longitudinal variability

Abstract

Abstract Alzheimer’s disease and related neurodegenerative disorders exhibit marked clinicobiological heterogeneity, posing challenges for accurate diagnosis, prognosis and therapeutic development. Using data from multiple cohorts of the Alzheimer’s Disease Neuroimaging Initiative (ADNI), we conducted comprehensive cross-sectional and longitudinal concordance analyses of dementia severity, CSF and PET amyloid and tau. We also examined the stability of biomarker positivity and cognitive status (stable, progressive, reversive and fluctuating) and their inter-relationships. The ability to separate methodological from biological sources of variability is relevant to understanding the robustness of the Amyloid/Tau/Neurodegeneration classification framework, which is particularly important in the era of disease-modifying therapies. Our data revealed both temporal and spatial discrepancies among variables used for classifying the level of cognitive impairment and between the methods used for classifying positive amyloid or tau status (CSF/PET). Methodological discrepancies in region-specific definitions, positivity thresholds and segmentation constraints contributed to PET inconsistencies. Additionally, biomarker status for CSF and PET positivity fluctuated over time, and there was no evidence of temporal precedence for either CSF or PET. There was pronounced spatiotemporal heterogeneity in tau-PET patterns, with implications for the clinical trajectory. Initial tau deposition in the entorhinal cortex, compared to non-limbic meta-temporal regions, was associated with more rapid cognitive decline and greater amyloid burden. Notably, a subset of mild cognitive impairment cases demonstrated diagnostic reversion to cognitively normal status and showed lower amyloid and tau burden and a higher prevalence of affective symptoms, possibly representing individuals without or at an early stage of Alzheimer’s pathology. Results underscore the complexity and variability of Alzheimer’s disease phenotypes and progression and provide a sanguine perspective on current diagnostic frameworks. Recognizing variability sources in biomarkers is particularly important when diagnosis relies on a single biomarker or cross-sectional data. More importantly, given the complex interplay of factors contributing to dementia and cognitive decline, deep learning and artificial intelligence-based algorithms are needed for precise and individualized diagnoses. Harmonized and standardized datasets with clearly defined variables, without excluding atypical cases, are needed for the effectiveness of such machine learning models. Therefore, investigating Alzheimer’s disease biomarker heterogeneity is critical towards maximizing the potential of artificial intelligence in integrating multi-dimensional clinical and biomarker data beyond human-centred interpretation. In this comprehensive work, we provide insights into the observed heterogeneities and their potential contributing factors. We aim to use this insight and continue this effort by harmonizing the conflicting data, making them suitable for diagnostic algorithms and sharing the harmonized dataset in the near future.