recurrent neural network

Autism spectrum is a neurological disorder that manifests itself in the early years of a child's development. People with autism face challenges in regulating emotions and express their emotional states in different ways. The current research presents a vocal activity detection (VAD) system adapted to the voices of autistic children.

The proposed VAD system is a Recurrent Neural Network (RNN) with short-term memory (LSTM) cells. The data includes 25 English-speaking autistic children performing a structured learning activity and was collected as part of the DE-ENIGMA project.

Our experiments show that the pediatric VAD system performs less well than our generic VAD system trained under the same conditions, as we obtain system performance characteristic curve under the curve (ROC-AUC) criteria of 0.662 and 0.850, respectively. The SER results show different performances between capacity and excitation, depending on the VAD system used, with a maximum match correlation coefficient (CCC) of 0.263 and a minimum root mean square error (RMSE) of 0.107.

Although the performance of SER models is generally low, the pediatric VAD system can lead to slightly improved results compared to other VAD systems and especially the VAD-less baseline, which supports the hypothesized importance of pediatric VAD systems in the context under discussion.

Diabetes is a group of metabolic disorders that are the result of untreated high blood glucose greatly. Early diagnosis and continued control of this disease can reduce its effectsTherefore, providing a method for timely diagnosis of this disease is of great importance. Until now, researchers have made many efforts to provide machine learning methods to diagnose diabetes. But most of these models are either based on simple machine learning methods or based on the assumption that the available diabetes data are balanced. Both cases are factors of their complete failure. Therefore, considering the existing challenges as well as the importance of timely diagnosis of diabetes, in this research, a diabetes diagnosis model based on deep recurrent neural networks and SMOTE oversampling algorithm is presented. In this model, several pre-processing steps including quantification of missing values, removal of outlier data and then oversampling have been performed. Three deep recurrent neural networks with three recurrent hidden units including LSTM, GRU and BiLSTM have been used to diagnose diabetes. The results of the model presented on the Pima database received from the UCI repository indicate that the average accuracy in 10 different runs in LSTM, GRU and BiLSTM is 91.21%, 89.61% and 90.99%, respectively. The recurrent network with GRU unit has achieved the highest accuracy of 93.74% on average in 10 different executions. The results of the proposed model show that deep neural networks have a much more successful performance in diabetes diagnosis compared to traditional machine learning methods.