Ensemble transfer learning for Alzheimer's disease diagnosis.

There is considerable interest in developing inexpensive, nonintrusive diagnostic tests for Alzheimer's disease (AD), in the hope these tests will facilitate early diagnosis and support design of effective treatments. While tests based on blood-borne biomarkers such as microRNAs have shown promise, work to date indicates these tests often do not generalize well: diagnostic models derived for one patient group are not accurate when applied to a new cohort. This paper presents a novel ensemble-based transfer learning methodology which induces models that provide accurate diagnoses across distinct patient groups without retraining. The algorithm combines information from supervised ensemble learning and unsupervised ensemble clustering to enable robust transfer learning. The efficacy of the approach is illustrated through a case study involving microRNA-based AD diagnosis. In this test, we use our algorithm to learn a diagnostic model on data from one patient group and then apply the model to a different target group, obtaining diagnostic accuracy which actually exceeds that of a state-of-the-art model trained directly on the target group.