Proposal for Feature Enhancement of Bioinformation Using Attractor Pattern and Frequency Analysis
| Autor: | Makoto Kikuchi |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
Artificial neural network
business.industry Computer science Deep learning Feature extraction Pattern recognition 02 engineering and technology Signal Convolution Identification (information) 020401 chemical engineering Attractor 0202 electrical engineering electronic engineering information engineering Feature (machine learning) 020201 artificial intelligence & image processing Artificial intelligence 0204 chemical engineering business |
| Zdroj: | ISMICT |
| Popis: | When recognising signals using deep learning, it is essential to extract features efficiently. In this study, we propose an individual identification method using medical data itself as identification information as part of research on medical malpractice prevention technology. In particular, to efficiently emphasise the features of the signal, we proposed a pretreatment method combining attractor pattern and frequency analysis. Especially, as a biological signal, in this study, we focused on the centre of gravity fluctuation of a standing posture, which is one of the human's whole body movements. This movement is representative biological information and is being studied not only as a diagnosis of diseases and functional disorders however also as an index for evaluating health condition. In this study, we extract the component in one direction from the centre of gravity fluctuation and create the attractor pattern using that signal and its rate of change. Moreover, from the difference in the model, the identification of the subject and the state of the standing posture control system are identified. The attractor pattern is two-dimensionally Fourier transformed to emphasise a part of the characteristics of signals. After filtering the result, it used for supervised machine learning as the input signal of the hierarchical neural network. Furthermore, classification and individual identification of unknown data were performed using the weight space obtained by machine learning. As a result, the effectiveness of the method proposed this time confirmed from the viewpoint of feature extraction. The process of this study has less information loss compared with the case using a convolution layer and has a small computation processing load. Therefore there is a possibility that it can apply to a failsafe system for medical malpractice prevention or a medical diagnosis system. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|