Feature Selection to Forecast Cognitive Decline Using Multimodal Alzheimer's Disease Models

Multimodal medical data (e.g. MR and PET imaging, CSF measurements, clinical assessments) reflect different aspects of Alzheimer's Disease, including early changes in brain structure and function that can occur before the onset of the associated cognitive impairment.

We propose to use a feature selection method within a disease progression model to identify the combinations of imaging and non-imaging biomarkers across modalities that allow the best predictions of the cognitive decline.

We first demonstrate that the chosen non-linear mixed-effect model outperforms all benchmarked methods in the TADPOLE challenge, with increasing performance as various modalities are added.

We then introduce a controlled protocol to compare the added value of each feature for the forecast of cognition, at different stages of the disease, and for varying time-to-predictions.

Notable findings include that the volumes of the ventricles are predictive features at the later AD stages but not at early stages, hippocampal volume is mostly important for intermediate stages and cognitively unimpaired subjects, cortical thickness of temporal cortex is most important for short-term predictions in AD patients at any stages, and cortical summaries of glucose and amyloid PET uptakes are only useful for intermediate AD stages.

These conclusions may inform the design of efficient prognosis scores that have been shown to decrease sample size in clinical trials and can be adapted to the targeted disease stages and the trial duration.