25 July 2018
Neural Network classification of longitudinal cognitive data for prediction of individual-level change
For the first time, Neural Network classifiers have been applied to changes in CANTAB PAL performance between baseline and 10 months to accurately predict the development of MCI at 20 months. Dr Elizabeth Baker, Statistical Scientist at Cambridge Cognition, presented the novel findings at AAIC 2018.
In an attempt to combat the pandemic that is Alzheimer’s disease (AD), clinical trials are now being conducted earlier in the disease process with the aim of delaying the onset, or preventing the progression, of the disease.
With this shift in research focus in mind, it is crucial to develop models that can reliably predict whether an individual with Mild Cognitive Impairment (MCI) will transition to AD.
One potential avenue is Neural Network (NN) classifiers which model the temporal relations between events.
This study is the first to apply NNs to repeated memory assessments with the aim of predicting, at an individual level, whether MCI will progress into Alzheimer’s disease.
106 older adults (57% female, M = 70 years) were assessed on the CANTAB Paired Associates Learning (PAL) memory assessment at baseline, 10 months and 20 months.
Based on PAL performance at baseline, participants were classified as either:
- Non-amnestic MCI if deficits in performance were limited to non-memory tests
- Amnestic MCI if deficits existed in memory tests, or
- Control participants with no measurable deficits
Our algorithms predicted that participants would either “progress” or be “stable” MCI at 20 months, based on only:
- Disease duration, and
- PAL performance at baseline and 10 months
The best performing NN classifiers and Gradient Boosting Machines (GBM) showed excellent sensitivity in using PAL scores to predict MCI progression (NN AUC 0.92, GBM AUC 0.81), as shown in figure 3.
In conclusion, our algorithms show excellent sensitivity for predicting MCI progression at 20 months, using PAL performance at baseline and 10 months.
These findings suggest that using automatic learning of complex temporal relationships in cognitive test scores could support specialists in making early diagnoses of Alzheimer’s disease, which could potentially enhance efficacy for the next wave of clinical trials.
Dr Elizabeth Baker and colleagues