Independent validation and outlier analysis of EuroPOND Alzheimer’s disease staging model using ADNI and real-world clinical data

Mandy Melissa Jane Wittens, Diana M. Sima, Arne Brys, Hanne Struyfs, Ellis Niemantsverdriet, Ellen De Roeck, Christine Bastin, Florence Benoit, Bruno Bergmans, Jean‐Christophe Bier, Peter Paul De Deyn, Olivier Deryck, Bernard Hanseeuw, Adrian Ivanoiu, Gaëtane Picard, Éric Salmon, Kurt Segers, Anne Sieben, Evert Thiery, Jos Tournoy, Anne‐Marie Van Binst, Jan Versijpt, Dirk Smeets, Maria Bjerke, Maura Bellio, Neil P. Oxtoby, Daniel C. Alexander, Annemie Ribbens, Sebastiaan EngelborghsAlzheimer 2025 · DOI

Abstract

This study validates the EuroPOND Event-Based Model (EBM) for Alzheimer’s disease (AD) staging using both standardized research data (ADNI) and heterogeneous real-world clinical data (REMEMBER). The model successfully identifies a consistent biological progression from cerebrospinal fluid (CSF) changes to cognitive decline and structural brain atrophy. Notably, the model distinguishes between amnestic and non-amnestic mild cognitive impairment (MCI), offering a tool for precision recruitment in clinical trials.

Introduction: Beyond Coarse Clinical Labels

Clinical diagnosis of Alzheimer’s disease typically relies on broad categories: Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease Dementia (ADD). However, these labels often fail to capture the underlying biological progression. Two patients sharing an MCI diagnosis may exist at vastly different stages of the disease's biological "clock." Furthermore, models developed on "pristine" research datasets like ADNI often struggle with the "messy" reality of clinical practice, where data is frequently missing or non-standardized. This research addresses the need for a robust, biomarker-driven staging system applicable across diverse clinical environments.

Methods: Event-Based Modeling (EBM)

The researchers employed an enhanced version of the EuroPOND EBM toolbox. EBM is a probabilistic, data-driven approach that determines the most likely sequence of biomarker changes—referred to as "events"—without requiring pre-defined clinical cut-offs.

The model tracked nine key variables:

  1. CSF Biomarkers: Amyloid-beta (Aβ1-42), P-tau181, and T-tau.
  2. Clinical Scores: MMSE (global cognition) and RAVLT (memory).
  3. MRI-derived Volumes: Hippocampus, temporal, parietal, and frontal cortices (computed via icobrain dm).

The model produces a 10-stage scale (0–9), where Stage 0 indicates no abnormalities and Stage 9 signifies that all biomarkers have reached an abnormal state.

The maximum likelihood sequence of biomarker changes determined by the EBM, showing the progression from CSF markers to cognitive decline and finally regional brain atrophy.

Key Insights: The Biological Cascade

The model confirmed a consistent "cascade" of events: CSF changes occur first, followed by cognitive decline, and finally structural brain atrophy.

MCI Heterogeneity

A critical finding was the differentiation between amnestic MCI (aMCI) and non-amnestic MCI (naMCI). The EBM revealed that aMCI patients are significantly further along the AD trajectory, with 54% residing in the most advanced stages compared to only 38% of naMCI patients. This suggests that naMCI may frequently be driven by non-AD pathologies, highlighting the model's utility for precision patient screening.

Positional variance diagram showing the probability of each biomarker reaching an abnormal state at each stage of the sequence. The concentration along the diagonal indicates high model confidence in the sequence.

Results and Validation

The model, trained on the ADNI dataset (n=1737), was successfully validated on two independent cohorts:

  • Research Cohort (University of Antwerp): 119 patients with complete multimodal data.
  • Real-world Cohort (REMEMBER): 686 patients from nine Belgian dementia clinics. Despite missing CSF data in the REMEMBER cohort, the model effectively staged patients, demonstrating high generalizability.

Outlier analysis further proved the model's biological accuracy. For instance, "healthy" controls placed in high stages often possessed positive Amyloid-PET scans (preclinical AD), while MCI patients in Stage 0 often had symptoms attributable to non-AD causes, such as depression.

Conclusion and Future Directions

The EuroPOND EBM provides an objective, biomarker-driven framework for AD staging that remains robust even with real-world clinical data. While the current model assumes a single sequence of progression, future work will utilize Sustain (Subtype and Stage Inference) models to account for AD's inherent heterogeneity and different disease "flavors." Incorporating longitudinal data will further enhance the model's predictive power for multi-center clinical trials.

Summary updated

Abstract

This study highlights the generalizability of EuroPOND's AD EBM model across research and real-world clinical datasets, supporting its use in multi-center trials. aMCI subjects generally reside in more advanced stages than naMCI, who may not necessarily have AD, demonstrating utility for precision recruitment/screening.
BibTeX
@inproceedings{wittens2025independent,
  author = {Mandy Melissa Jane Wittens and Diana M. Sima and Arne Brys and Hanne Struyfs and Ellis Niemantsverdriet and Ellen De Roeck and Christine Bastin and Florence Benoit and Bruno Bergmans and Jean‐Christophe Bier and Peter Paul De Deyn and Olivier Deryck and Bernard Hanseeuw and Adrian Ivanoiu and Gaëtane Picard and Éric Salmon and Kurt Segers and Anne Sieben and Evert Thiery and Jos Tournoy and Anne‐Marie Van Binst and Jan Versijpt and Dirk Smeets and Maria Bjerke and Maura Bellio and Neil P. Oxtoby and Daniel C. Alexander and Annemie Ribbens and Sebastiaan Engelborghs},
  title = {{Independent validation and outlier analysis of EuroPOND Alzheimer’s disease staging model using ADNI and real-world clinical data}},
  booktitle = {Alzheimer},
  year = {2025},
  doi = {10.1186/s13195-025-01788-6},
}
Figure