Feature Extraction from Infrared Spectroscopy by Machine Learning with Correlation Coefficient as Error Function

Conventionally when diagnosing diabetes, experts diagnose using infrared spectroscopy of blood in hospital. Recently, however Blood testing apparatus that are easily available at home are attracting a lot of attention as a preventive medicine. To realize this, it must be easy to measure by an ordinary person at any time. Therefore, I aim to develop a system to analyze infrared spectroscopy acquired from blood by machine learning. Neural networks are also used for pattern recognition problems such as biological signals. However, to discriminate the infrared spectrum, the input becomes high-dimensional, and in a device such as those used at home, large errors are included in the data, so the accuracy drops remarkably. In this study, as a basic study of the analysis of infrared spectroscopy by machine learning, I experimented to extract characteristic peaks from artificially generated data by a new machine learning. In the proposed machine learning, learning is performed by an error back propagation using a correlation coefficient for the error function to extract a waveform more than the value magnitude as a feature value. In this paper, I show the result of the experiment of extracting feature value from artificial data simulating infrared spectroscopy.