A weighted multi-feature transfer learning framework for intelligent medical decision making

Abstract

Abstract Transformative computing provides an emerging technology to data analysis and information processing, but how to effectively connect the data derived from different domains has aroused much of concern. Especially on medical areas, the scarcity of annotated medical data makes it hard to build a robust classification model, thus, the utilization of medical resources from different sources is particularly important. Transfer learning leverages the knowledge gained from the related domain to enhance the computational effectivity on the target domain. In this work, we extend transfer learning with ensemble learning to present a novel Weighted Multi-Feature Hybrid Transfer Learning Framework (W-MHTL) that builds a transformative approach to connect different domains and applies it to medical decision making. Our approach lessens the distribution variances from multiple perspectives by applying variant types of feature-based transfer learning methods. In each feature space, we construct the transfer model by evaluating the correlations and obtain the predicting result from each model. Finally, a feasible ensemble strategy is used to jointly consider each result. We evaluate our approach on synthetic datasets and UCI medical benchmarks, and a cerebral stroke dataset collected from local hospital. The experimental results reveal that our method achieves superior performances with the currently available alternatives.