Designing a model to detect and separate data anomalies caused by sensors and medical wearables using LSTM neural network algorithm

Predicting abnormalities of wearable medical devices plays a very important role in saving the lives and health of patients. This importance has opened new horizons for researchers with the development of newer algorithms. The long-term memory algorithm (LSTM) is one of the most important methods that are a special type of recurrent neural network (RNN) that has a high ability in this field and greatly increases the accuracy of correct and incorrect prediction of these abnormalities. In the current research, by using this algorithm and taking into account different parameters, the anomalies related to the sensors of the research field were determined. The results showed that there are influential parameters in the construction of this architecture, which include ۳ very important factors: the number of neurons in the LSTM layer, the batch size, and the activation function. Also, the LSTM architecture together with the Dropout layer, with parameters Batch size = N = ۱۲۸ and Tanh activation function shows a better performance and the lowest amount of error (MAPE) as well as the amount of the calculated mean square error (RMSE) in determining the anomaly. have sensors in the medical field. Investigations related to the results of ۱۶ repetitions of optimization also showed that the process of reducing errors in the correct and incorrect identification of anomalies in the training phase has reached its lowest level with the increase in the number of tests, which shows the optimality and appropriateness of the work process. Therefore, this algorithm has a very good ability to identify errors in sensors and medical wearables, and it will be of great help in identifying the possible failure of sensors, and critical conditions of the patient, informing and finally helping patients in time.

کلیدواژه ها:

anomaly ، medical sensors ، medical wearables ، Neural Networks ، LSTM algorithm