Comparison of Night, Day and 24 h Motor Activity Data for the Classification of Depressive Episodes

Abstract

Major Depression Disease has been increasing in the last few years, affecting around 7 percent of the world population, but nowadays techniques to diagnose it are outdated and inefficient. Motor activity data in the last decade is presented as a better way to diagnose, treat and monitor patients suffering from this illness, this is achieved through the use of machine learning algorithms. Disturbances in the circadian rhythm of mental illness patients increase the effectiveness of the data mining process. In this paper, a comparison of motor activity data from the night, day and full day is carried out through a data mining process using the Random Forest classifier to identified depressive and non-depressive episodes. Data from Depressjon dataset is split into three different subsets and 24 features in time and frequency domain are extracted to select the best model to be used in the classification of depression episodes. The results showed that the best dataset and model to realize the classification of depressive episodes is the night motor activity data with 99.37% of sensitivity and 99.91% of specificity.

Comparison of Manual and Automated Feature Engineering for Daily Activity Classification in Mental Disorder Diagnosis

“…In [15] only the most basic statistical features are extracted from the signal, such as mean, variance, or interquartile range. Authors of [13] propose feature extraction additionally in the frequency domain, using power spectral density (PSD), calculating, e.g. mean, variance, entropy, or spectral flatness in the frequency domain.…”

“…We extracted basic statistical features from time domain signals based on previous articles using Depresjon and Psykose datasets [13,15,2]. Additionally, since the actigraphy data is a signal, we included frequency domain features [13]. We used Welch's method to estimate the power spectral density (PSD) and extract features from it.…”

“…Activity signals have a lot of variation and noise, but we are interested in the overall long-term effects to detect the mental disorder. For this reason, we perform experiments with resampling data in the time domain before feature engineering, replacing the signal with the mean activity for each hour, same as in [13]. This should be particularly effective with features based on minimal values since for raw signals the minimum is always 0 (there is always a period when someone does not move).…”

“…They also extracted statistical features but without formal mathematical justification. More recent developments focus on using machine learning to automatically perform diagnosis [13,14].…”

“…While neural networks such as Gated Recurrent Unit (GRU) or Long-Short Term Memory (LSTM) can work with it, they are known to require a significant amount of data to train, and available datasets are very small. For those reasons, the approach taken by most papers is a classical 2-step process: extract statistical features and use a regular classifier for tabular data [2,13,15]. This approach can be performed manually, where researchers choose features themselves.…”