Attention-enhanced BiLSTM for score forecasting in parkinson’s disease

Abstract

Parkinson’ Disease is the most progressive neurological disease that is caused because of the dying of the neurons or brain cells.It is the most common neurological disease caused around the globe, with most patients in the age greater than 60[1].The disease is characterised with tremors,instability in the walking or posture even problem in sleeping along with speech problems. Speech is one of the earlier trait in diagnosing the disease at the earliest stage[9]. So in this study the speech characteristics or features are taken advantage of to forecast the disease at the earliest.Using a patient dataset that has a clinical PD grading based on speech characteristics could slow the progression of PD by offering a computational prognostic tool for the condition. Based on previously and now recorded speech, it can assist a person with PD in tracking the development of unique symptoms they are currently experiencing. This study uses recurrent neural network, Bidirectional Long- Short term memory or BiLSTM to forecast the time series based output. The traditional BiLSTM framework is used as base model and an attention layer is used on top of that to increase the efficiency or performance of forecasting.The model proposed is called attention-BiLSTM. The performance of the model is compared with other tradition recurrent neural models like LSTM and BiLSTM.Regression models are commonly used in the forecasting models. So, the performance of model is also compared with that of multiple linear regression.A ridge regression is an advanced form of traditional linear regression is also run on the the dataset and is also used for the comparison of the attention based BiLSTM model.Different performance metrics are used to study the performance. Though RMSE is the metrics that is widely used in literature for the performance analysis in the forecasting but in this study along with the RMSE, other metrics like MSE and MAE are also used in the study.The proposed attention enhanced BiLSTM model is showing a result of 7.58%,4.36% and 2.08% for MAE,MSE and RMSE respectively, that is better than the comparing/base models in the study.