Abstract:The deep learning–based computer-aided diagnosis (CADx) approaches of dementia often require a lot of manual intervention. Although deep learning has a good effect on feature extraction, the current deep learning methods usually need to set a large number of parameters manually, which is time consuming. Hierarchical extreme learning machine (H-ELM) needs only less manual intervention and can extract features by a multi-layer feature representation framework, which is much faster than the traditional deep learning methods. A CADx framework based on H-ELM, named DCADx, is proposed. As common spatial pattern (CSP) and brain functional network (BFN) have been proven to have better de-redundancy effects on brain data, the DCADx contains two different data redundancy reduction methods: (1) CSP-based DCADx (i.e., DCADx-CSP model) and (2) BFN-based DCADx (i.e., DCADx-BFN model). The experimental evaluation proved the effectiveness of the proposed algorithms. The DCADx-CSP model obtained 83.2% on Alzheimer’s disease and 82.5% on Parkinson’s disease. The DCADx-BFN obtained 89.3% on Alzheimer’s disease and 88.7% on Parkinson’s disease. DCADx can make full use of the feature expression ability of H-ELM to achieve better performance. CSP and BFN can reduce the redundancy to enhance the diagnostic accuracy further.

Computer-Aided Dementia Diagnosis Based on Hierarchical Extreme Learning Machine