Time series signal forecasting using artificial neural networks: An application on ECG signal. (July 2022)

• Record Type: Journal Article
• Authors: Ratna Prakarsha, Kandukuri; Sharma, Gaurav
• Is Part Of: Biomedical signal processing and control. Volume 76(2022)
• Publication Date: 2022-07
• Subjects: Time Series Forecasting -- Artificial Neural Networks (ANN) -- FFNN -- Adaptive Filter -- LMS; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103705
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Time Series Forecasting is the prediction of future values of a signal based on the observed past values. It has various applications in signal processing, especially in the medical field which needs high accuracy. This paper presents an MLP (Multilayer Perceptron), a class of FFNN (Feedforward Neural Network) for highly accurate time series forecasting. There are various methods of signal processing that are used in time series forecasting but each method is specific to the particular problem it solves. The current methods involve the use of different types of adaptive filters out of which the most common method is LMS (Least Mean Square) algorithm. Although the adaptive filters give a decent accuracy, but neural networks (NN) give the results more than satisfactory. On performing time series forecasting on a simulated ECG (Electrocardiogram) signal, an accuracy of 95.72% was achieved using ANNs (Artificial Neural Networks) competing with the LMS filter, which gave only 79% accuracy. When the same was implemented on real ECG data of a person suffering from Sleep Apnea, the ANNs offered 98.68% while LMS filter displayed only 91% accuracy. Additionally, the neural network was also denoising the signal while predicting. A signal-to-noise ratio of 29.71 dB and 16.33 dB for Neural Network prediction and LMS filter prediction was attained, respectively. In the case of the real data, the aforementioned values stand at 22.8 dB and 3.8 dB, respectively. Simulated results show that the neural networks give superior performance in time series forecasting than Adaptive Filters.