Attention mechanism in neural network for early diagnosis in newborns using cry signals

Abstract

Humans use their voice to communicate, and it is often driven by instinct. On a larger scale, this same instinct can be applied to a newborn infant who tries to express itself, with obviously, his cry. According to past studies made on this subject, there is a strong correlation between an infant's cry and psychological condition and pathology that affects the latter. This study focuses on finding a correlation and identifying the newborn's condition using a neural network. This tool would help experts identify something that they could have potentially ignored and better diagnose a newborn according to the disease affecting it before it is too late. Whether it is in a developed country or a developing country, this solution requires no expensive diagnosis material. Our study uses an Attention mechanism neural network that has known success in speech recognition and text translation in the last two years. The Attention layer learns to focus on different aspects of the input. The Transformer uses Encoder-Decoder architecture to summarize the entire input before outputting the results, which are strongly linked with each input sequence. The idea is to eliminate the dependency in the fixed-length input conducted in the previous studies. The Attention mechanism enhanced LSTM (Long short-term memory) is a recurrent neural network that inherits the self-attention layer from the Transformer. Only the expirations sessions are extracted for each cry sample to generate the matrix of Melfrequency cepstral coefficients (MFCC). These features are then fed into the neural network. Using the data at our disposition, we train the network and test it with new data to compare the performance with the classical LSTM default variant. Several previous studies use the same dataset, so the results are compared to the same criteria to evaluate the performance of this variant for our purpose. The features and parameters are optimized with both variants to obtain a global view of the Attention mechanism for early diagnosis in newborns and conclude if this path can be taken or not.

Date	9 Nov 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Chakib Tadj (Supervisor)