Diagnosis Of Schizophrenia Using LSTM Models

Schizophrenia is a mental disorder which affects 1% of total human population and is often misdiagnosed. This paper is an attempt at addressing the problem with the application of LSTM models and deep Learning

Keywords—Deep Learning, RNN, LSTM, EEG

Introduction

With the introduction of improved computing power, availability of a large amount of data and reliable and cost effective data storage facilities led to the development of deep learning systems. Deep learning models can be applied directly on the available raw data and produce reasonable and accurate functions to account for data and predict its future outcomes. The foundations for deep learning models were laid in as far back as 1950’s. Recently the models have been divided as two – feed forward networks and recurrent neural networks. Conventional feed forward models rely upon the current data and independent nature of both the training and test data sets. Moreover on addition of new data previous data is lost and are limited to fixed length of training vectors. This renders them as unviable candidates for modelling time depending scenarios. Recurrent Neural Networks (RNN) remedies the above problem.

Where Q is the activation function,,, and are weights of input-hidden, hidden-output layer and recurrent connections respectively. Here and are vector biases. As evident from the equations the inputs of an RNN depends upon current data point and those from the hidden stateBut the practical applications of RNNs are limited due to the decay or blowing up at recurrent connections caused by the vanishing and exploding gradient. LSTM LSTM or Long Short Term Memory models are an improvement over normal RNNs. LSTMs was introduced in by Hochreiter and Schmidhuber in the year 1997. The structure of LSTMs resembles a set of connected cells and is comprised for three main gates and two nodes which are used for the propagation of information within the cells of the network. The three main gates are as follows – Forget gate, Input gate and Output gate. Forget gate is used to scale the data based on the weight assigned to it.

This essentially removes unwanted details of pre-existing data to better optimise the future outcomes. The forget gate takes in two inputs andwhich corresponds to the previous data and current data respectively. The inputs are assigned a weight values based on previous outcomes and are passed on to the input cell after undergoing a sigmoid function. The sigmoid function assigns the values from it a weight from 0 to 1, wherein ‘0’ denotes absolute removal of that particular data and ‘1’ would be to store the entire data. The input gate is responsible for controlling the input activation into the memory. Input gates acts similar to forget gate and can be considered as filter to extract max data from and. In order to regulate the values a sigmoid function is often used. Further a vector containing all the values is produced using the tanh function to output values in the range -1 to +1. The regulatory filter value (from sigmoid) and the vector value (from tanh) is multiplied and is added to cell state via addition. Output gate is comprised of 3 steps. The first is the creation of a vector by applying tanh function to the cell state. This renders a value between -1 and +1. A filter is fashioned out using a sigmoid function taking in and to regulate the output values of the vector created. The output is produced my multiplication of the vector and filter values.

However one patient out of 21 subjects involved in the test was tested positive for schizophrenia even though that subject was one among the controls. This account for 4. 7 % of all tested subjects and thus possibility of false detections are also present. With the above results the authors state a positive viability of such an LSTM model for the diagnosis of schizophrenia.

CONCLUSION

The proposed LSTM model can accurately classify the people with schizophrenia with accuracy of 95. 32%. This renders the proposed model a viable candidate for real world diagnosis of schizophrenia. Such device enabled with the model could significantly reduce misdiagnosis to a great extent.