Non-Parametric Mixture Modelling and its Application to Disease Progression Modelling

Nicholas C. Firth, Neil P. Oxtoby, Silvia Primativo, Emilie Brotherhood, Alexandra L. Young, Keir Yong, Sebastian J. Crutch, Daniel C. AlexanderbioRxiv 2018 · DOI

Abstract

Dementia is characterized by the progressive degeneration of cognitive abilities. Understanding these changes is vital for determining how structural brain alterations affect cognition and for facilitating early detection. While disease progression models (DPMs) are often used to reconstruct longitudinal trajectories from cross-sectional data, existing models—primarily designed for brain imaging—frequently rely on Gaussian assumptions that do not translate well to cognitive data. This paper introduces Kernel Density Estimation Mixture Modelling (KDEMM), a non-parametric approach that accurately models non-normally distributed cognitive test data. Our results demonstrate that KDEMM outperforms traditional Gaussian Mixture Models (GMMs) in both simulated and clinical datasets, providing a more accurate reflection of clinical phenotypes in Alzheimer’s disease subtypes.

The Challenge: Why Gaussian Models Fail Cognitive Data

To map the long-term progression of diseases like Alzheimer’s, researchers utilize Event-Based Models (EBMs). These models determine the sequence in which biomarkers transition from a "healthy" to a "diseased" state. Traditionally, EBMs employ Gaussian Mixture Models (GMMs) to define these states.

However, cognitive test scores rarely follow a normal distribution. They are often plagued by "ceiling effects" (where healthy individuals achieve perfect scores) and "floor effects" (where impaired individuals score zero), resulting in highly skewed or bimodal distributions. Forcing this "messy" data into a Gaussian framework often leads to inaccurate disease timelines that contradict clinical reality.

Figure 6: Various data distributions (a-d) encountered in cognitive testing and simulations, highlighting the non-Gaussian nature of clinical biomarkers that KDEMM is designed to handle.

Methodology: The KDEMM Approach

The core innovation of this work is the replacement of parametric GMMs with Kernel Density Estimation (KDE). Unlike GMMs, which assume a specific mathematical shape (the bell curve), KDE creates a smooth silhouette of the data's actual distribution, regardless of its skewness or complexity.

The authors developed an iterative algorithm to fit these models:

  1. Update Step: Fitting a KDE to subsets of the data to define the healthy and diseased distributions.
  2. Assignment Step: Re-assigning individual data points to either the healthy or diseased group based on which KDE provides a better fit.
Figure 4: The iterative convergence of the KDEMM algorithm. (a) The initial update step shows significant overlap, which (b) refines by the second iteration and (c) reaches a stable, converged model by the seventh iteration.

Experimental Validation

The model was evaluated using both synthetic and clinical data:

  • Simulated Data: 5,000 synthetic subjects were generated using Gamma distributions to test the model against a known "ground truth" with varying levels of data skewness.
  • Clinical Data: A dataset from the Dementia Research Centre (London) included patients with Posterior Cortical Atrophy (PCA) (a visual-led variant) and typical Alzheimer’s disease (tAD) (memory-led), alongside healthy controls.

Key Results and Clinical Insights

  • Superior Accuracy: In simulations, KDEMM achieved a significantly higher correlation with the ground truth sequence (Kendall’s $\tau = 0.89$) compared to the GMM ($\tau = 0.50$). It also proved more reliable at staging patients, with a correlation of $\rho = 0.88$ versus the GMM’s $0.83$.
  • Clinical Alignment: In the PCA cohort, KDEMM correctly identified visual impairment events (e.g., failing "dot counting" tests) as occurring earlier in the disease timeline than memory loss. In contrast, the GMM produced a sequence that was clinically nonsensical.
  • Model Stability: Bootstrapping analysis revealed that KDEMM produces much more consistent timelines across different data subsets, whereas GMMs were prone to "jitter," with biomarkers shifting positions unpredictably.

Conclusion

KDEMM represents a significant advancement in disease progression modelling for non-normally distributed data. By moving away from rigid parametric assumptions, researchers can more accurately integrate cognitive test batteries into complex models. Future work will focus on combining KDEMM with imaging and fluid biomarkers to create holistic, multi-modal maps of neurodegenerative decline.

Summary updated

Abstract

Abstract Dementia is characterised by its progressive degeneration of cognitive abilities. In research cohorts, detailed neuropsychological test batteries are often administered to better understand how cognition changes over time. Understanding cognitive changes in dementia is of great importance, particularly in determining how structural changes in the brain may affect cognition and in facilitating earlier detection of symptomatic changes. Disease progression models are often applied to these data to understand how a disease changes over time from cross-sectional data or to disease trajectories from large numbers of individuals. Previous disease progression models used to build longitudinal models from cross-sectional data have focused on brain imaging data; however, these models are not directly applicable to cognitive data. Here we use the novel, non-parametric, Kernel Density Estimation Mixture Modelling (KDEMM) approach and demonstrate accurate modelling of the progression of cognitive test data. We found that using KDEMM resulted in more accurate models of disease progression in simulated data compared to Gaussian Mixture Models (GMMs) for the majority of parameters used to simulate the data. When comparing KDEMM and GMM to cognitive data collected in different Alzheimers Disease subtypes, we found the KDEMM resulted in a model much more in line with clinical phenotype. We anticipate that the KDEMM will be used to integrate cognitive test data, and other non-normally distributed datasets into complex disease progression models.
BibTeX
@misc{firth2018non,
  author = {Nicholas C. Firth and Neil P. Oxtoby and Silvia Primativo and Emilie Brotherhood and Alexandra L. Young and Keir Yong and Sebastian J. Crutch and Daniel C. Alexander},
  title = {{Non-Parametric Mixture Modelling and its Application to Disease Progression Modelling}},
  year = {2018},
  doi = {10.1101/297978},
  howpublished = {bioRxiv},
}
Figure