Improved Prediction of Cognitive Outcomes via Globally Aligned Imaging Biomarker Enrichments Over Progressions.

Abstract

Longitudinal neuroimaging data have been widely used to predict clinical scores for automatic diagnosis of Alzheimer's Disease (AD) in recent years. However, incomplete temporal neuroimaging records of the patients pose a major challenge to use these data for accurately diagnosing AD. In this paper, we propose a novel method to learn an enriched representation for imaging biomarkers, which simultaneously captures the information conveyed by both the baseline neuroimaging records of all the participants in a studied cohort and the progressive variations of the available follow-up records of every individual participant. Taking into account that different participants usually take different numbers of medical records at different time points, we develop a robust learning objective that minimizes the summations of a number of not-squared l2-norm distances, which, though, is difficult to efficiently solve in general. Thus we derive a new efficient iterative algorithm with rigorously proved convergence. We have conducted extensive experiments using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Clear performance gains have been achieved when we predict different cognitive scores using the enriched biomarker representations learned by our new method. We further observe that the top selected biomarkers by our proposed method are in perfect accordance with the known knowledge in existing clinical AD studies. All these promising experimental results have demonstrated the effectiveness of our new method. We anticipate that our new method is of interest to biomedical engineering communities beyond AD research and have open-sourced the code of our method online.11The code package of this paper have been made publicly available online at https://github.com/lyujian/Improved-Prediction-of-Cognitive-Outcomes.