Actigraphy-Based Scratch Detection Using Logistic Regression

Abstract

Incessant scratching as a result of diseases such as atopic dermatitis causes skin break down, poor sleep quality, and reduced quality of life for affected individuals. In order to develop more effective therapies, there is a need for objective measures to detect scratching. Wrist actigraphy, which detects wrist movements over time using micro-accelerometers, has shown great promise in detecting scratch because it is lightweight, usable in the home environment, can record longitudinally, and does not require any wires. However, current actigraphy-based scratch-detection methods are limited in their ability to discriminate scratch from other nighttime activities. Our previous work demonstrated the separability of scratch from both walking and restless sleep using a clustering technique which employed four features derived from the actigraphic data: number of accelerations above 0.01 gs, epoch variance, peak frequency, and autocorrelation value at one lag. In this paper, we extended these results by employing these same features as independent variables in a logistic regression model. This allows us to directly estimate the conditional probability of scratching for each epoch. Our approach outperforms competing actigraphy-based approaches and has both high sensitivity (0.96) and specificity (0.92) for identifying scratch as validated on experimental data collected from 12 healthy subjects. The model must still be fully validated on clinical data, but shows promise for applications to clinical trials and longitudinal studies of scratch.