ConvNet and machine learning models with feature engineering using motor activity data for schizophrenia classification

The use of wearable sensors such as smartwatches is becoming increasingly popular allied with their increasing functionality and interest in their outputs. This has led to a corresponding interest and increase by researchers to develop tools to analyse the outputted data. In this research, machine learning and deep learning algorithms are applied to classify the presence of schizophrenia using time series activity data. The dataset was collected from a study about behavioural patterns in people with schizophrenia which contains per minute motor activity measurements for an average of 12.7 days for 54 participants, 22 with schizophrenia and 32 without. New features were developed by firstly generating statistical measures in the time domain and secondly by subdividing the day into 3 separate time categories, representing different portions of the circadian rhythm. Five machine learning models are trained using these features. These models classify participants into the condition group (with schizophrenia) and the control group (without schizophrenia). A deep learning convolutional neural network (ConvNet) was also developed which also utilized time of day categories. The best machine learning model using 10-fold cross-validation achieved an average precision of 97.6% compared to a baseline of 83.6% from the original paper that analysed this dataset. Using Leave One Patient Out (LOPO) as a validation technique the machine learning model gives an accuracy of 86.7%, with the deep learning model giving an average accuracy of 87.6% which is comparable to the state-of-the-art of 88%-92.5%. This is the first time to the best of the researchers' knowledge that a deep learning ConvNet model has been applied to this task.