[Paper Review] Models of Disease Spectra

Iead Rezek, Christian F. Beckmann|arXiv (Cornell University)|Jul 19, 2012

Gaussian Processes and Bayesian Inference25 references3 citations

TL;DR

This paper proposes Gaussian Process (GP) regression to model the continuous spectrum of brain activation across Alzheimer’s Disease (AD) progression, using neuropsychological scores (e.g., MMSE, ACE) as continuous inputs. By directly modeling fMRI phenotypes as a function of these scores, the method provides high-resolution, uncertainty-aware predictions of brain function across unobserved behavioral states, revealing non-linear patterns of neural decline and identifying optimal training data points to improve prediction confidence.

ABSTRACT

Case vs control comparisons have been the classical approach to the study of neurological diseases. However, most patients will not fall cleanly into either group. Instead, clinicians will typically find patients that cannot be classified as having clearly progressed into the disease state. For those subjects, very little can be said about their brain function on the basis of analyses of group differences. To describe the intermediate brain function requires models that interpolate between the disease states. We have chosen Gaussian Processes (GP) regression to obtain a continuous spectrum of brain activation and to extract the unknown disease progression profile. Our models incorporate spatial distribution of measures of activation, e.g. the correlation of an fMRI trace with an input stimulus, and so constitute ultra-high multi-variate GP regressors. We applied GPs to model fMRI image phenotypes across Alzheimer's Disease (AD) behavioural measures, e.g. MMSE, ACE etc. scores, and obtained predictions at non-observed MMSE/ACE values. The overall model confirmed the known reduction in the spatial extent of activity in response to reading versus false-font stimulation. The predictive uncertainty indicated the worsening confidence intervals at behavioural scores distance from those used for GP training. Thus, the model indicated the type of patient (what behavioural score) that would need to included in the training data to improve models predictions.

Motivation & Objective

• To overcome the limitations of traditional case-control fMRI studies that discretize disease stages and lose information through data binning.
• To model brain activation as a continuous function of neuropsychological markers like MMSE and ACE, reflecting the true spectrum of Alzheimer’s Disease progression.
• To provide predictive uncertainty estimates that guide future data collection and cohort design for improved model reliability.
• To reverse the typical GP regression paradigm by predicting brain function from behavioral scores, rather than vice versa, to uncover neurophysiological causes of clinical decline.
• To enable region-specific identification of the most relevant behavioral markers for brain activation patterns across the cortex.

Proposed method

• Uses Gaussian Process regression to model high-dimensional fMRI image phenotypes as a continuous function of low-dimensional behavioral scores (e.g., MMSE, ACE).
• Applies a squared exponential kernel to capture smooth, non-linear dependencies between behavioral markers and brain activation across the whole brain.
• Employs leave-one-out cross-validation for model validation due to small sample size (N=15), despite its known bias.
• Treats neuropsychological scores as continuous, independent variables to avoid information-erosive binning of disease stages.
• Uses Bayesian inference to provide predictive distributions and uncertainty intervals for unseen behavioral scores.
• Compares linear and non-linear GP models to assess the shape of disease progression across brain regions.

Experimental results

Research questions

• RQ1 How can fMRI phenotypes be modeled as a continuous function of neuropsychological behavioral scores, rather than discrete disease groups?
• RQ2 What is the shape of brain activation changes across the AD spectrum—linear or non-linear—with respect to MMSE and ACE scores?
• RQ3 How does predictive uncertainty vary across different behavioral score values, and where should new data be collected to improve model accuracy?
• RQ4 Which brain regions show the most sensitive or distinct responses to changes in cognitive function, and what behavioral markers best explain their activation patterns?
• RQ5 Can a reverse GP regression approach—predicting brain function from behavioral scores—reveal causal neurophysiological trends not captured by standard classification methods?

Key findings

• The non-linear GP model provided a significantly better fit than the linear model across the whole brain, indicating that disease progression is not uniformly linear in neural activation patterns.
• Despite the overall superiority of the non-linear model, the linear model outperformed it in the temporal lobe, suggesting that memory-related activation intensity correlates linearly with MMSE performance.
• Predictive uncertainty increased for behavioral scores farther from the training data range, indicating reduced confidence in extrapolated predictions.
• The model identified specific behavioral score ranges that would most improve prediction accuracy if included in future training data, enabling data-driven study design.
• The approach successfully captured the known reduction in spatial extent of activation in response to reading versus false-font stimulation, validating its biological plausibility.
• The method enables region-specific identification of the most relevant behavioral markers for brain activation, a novel capability not previously demonstrated in fMRI modeling.