Continuous assessment of schizophrenia using heart rate and accelerometer data

• Type of Material: Article
• Authors: Erik Reinertsen, Georgia Institute of Technology; Maxim Osipov, University of Oxford; Chengyu Liu, Emory University; John M. Kane, The Zucker Hillside Hospital; Georgios Petrides, The Zucker Hillside Hospital; Gari D. Clifford, Emory University
• Language: English
• Date: 2017-07-01
• Publisher: IOP Publishing: Hybrid Open Access
• Publication Version: Author Accepted Manuscript (After Peer Review)
• Copyright Statement: © 2017 Institute of Physics and Engineering in Medicine.
• Final Published Version (URL): http://dx.doi.org/10.1088/1361-6579/aa724d
• Title of Journal or Parent Work: Physiological Measurement
• ISSN: 0967-3334
• Volume: 38
• Issue: 7
• Start Page: 1456
• End Page: 1471
• Grant/Funding Information: This research is funded by Emory University, a Research Council UK Grant EP/G036861/1 (CDT in Healthcare Innovation), a Wellcome Trust Centre Grant No. 098461/Z/12/Z (Sleep, Circadian Rhythms and Neuroscience Institute), and EPSRC grant EP/K020161/1 (Multiscale markers of circadian rhythm changes for monitoring of mental health).
• Keywords: Biophysics; machine learning; CORTISOL-LEVELS; Life Sciences & Biomedicine; mobile health; RHYTHM; Technology; Engineering, Biomedical; Science & Technology; PSYCHOSIS; FUZZY ENTROPY; RATE-VARIABILITY; heart rate; mhealth; MODULATION; support vector machine; PERIOD; ANTIPSYCHOTIC TREATMENT; schizophrenia; entropy; Physiology; STRESS; Engineering
• Research Categories: Psychology, General; Engineering, Biomedical

Abstract

Objective: Schizophrenia has been associated with changes in heart rate (HR) and physical activity measures. However, the relationship between analysis window length and classifier accuracy using these features has yet to be quantified. Approach: Here we used objective HR and activity data to classify contiguous days of data as belonging to a schizophrenia patient or a healthy control. HR and physical activity recordings were made on 12 medicated subjects with schizophrenia and 12 healthy controls. Features derived from these data included classical statistical characteristics, rest-activity metrics, transfer entropy, and multiscale fuzzy entropy. We varied the analysis window length from two to eight days, and selected features via minimal-redundancy-maximal-relevance. A support vector machine was trained to classify schizophrenia from control windows on a daily basis. Model performance was assessed via subject-wise leave-one-out-crossfold-validation. Main results: An analysis window length of eight days resulted in an area under a receiver operating characteristic curve (AUC) of 0.96. Reducing the analysis window length to two days only lowered the AUC to 0.91. The type of most predictive features varied with analysis window length. Significance: Our results suggest continuous tracking of subjects with schizophrenia over short time scales may be sufficient to estimate illness severity on a daily basis.