On the evaluation of disease progression modelling in Alzheimer's Disease

Abstract

AbstractBackgroundAlzheimer's Disease (AD) is a heterogeneous condition that can be viewed as a continuum involving multiple stages from cognitively normal to mild cognitive impairment and finally dementia. Recent works in Disease Progression Modelling (DPM) have attempted to use biomarkers to construct representations of progression that are more informative of patients' status than clinical diagnosis alone. However, there is no clear consensus (Figure 1) regarding the choice of evaluation methods or metrics to facilitate model comparison. We propose the use a simple, universally applicable biomarker‐based baseline to estimate patient‐realigning time‐shifts and subsequently evaluate it against multiple released DPM methods.MethodWe computed an abnormality score for 2439 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) by averaging the normalised values collected for 12 biomarkers representative of amyloid, tau, neurodegeneration, and cognition. Abnormality ranges were determined from 251 amyloid‐positive patients. These scores were compared to values from published DPM models (Figure 3). Additionally, all methods were benchmarked in different ways, including diagnosis classification and cluster analysis.ResultWe found that, despite its simplicity, our abnormality score reasonably separated patients according to their diagnosis (Figure 2). Computed scores also correlated significantly with those of other methods (Figure 3, Pearson correlations up to 0.891). While our approach only required cross‐sectional information, other methods rely on longitudinal data. Our method performed similarly to others in the cutoff‐based disease classification task, with test F1‐scores of 0.734 and 0.890 in the harder tasks of CN/MCI and AD/MCI, respectively (Table 1). In a three‐class setting, all methods including our baseline showed a good level of agreement, with Rand indices all above 0.6 (Figure 4).ConclusionA significant challenge in DPM for AD lies in estimating a proxy for patient realignment, which is crucial to describe individuals on a common disease continuum. We have proposed a method to calculate an abnormality score at the baseline visit, which can be trivially extended to longitudinal data and help interpret progression in follow‐ups. We hope that future work considers more principled evaluation procedures, facilitating meaningful comparisons between DPMs.