T14 - Digital Phenotyping In Bipolar Disorder

Abstract

Background Diagnostic categories of psychiatric illnesses based on subjective clinical assessments have been the standard for phenotyping in genetics studies. This is a significant problem for disorders such as bipolar, wherein the course of the illness is an important element of the phenotype. Digit phenotypes, assessed using mobile devices, offer an objective method of gathering data to describe phenotypes for translational research. It has been established that speech patterns are altered in mood disorders. The challenge is to monitor moods over extended periods of time using speech or other mobile-based methods and technology on personal devices. Technology is likely to be incorporated into the ongoing assessment and monitoring of individuals if it is passive and integrated into daily routines; mobile smartphones are ideal. Methods Individuals with rapid cycling bipolar disorder (51) and healthy controls (9) were ascertained from a prospective longitudinal study of bipolar disorder. They were provided with a smartphone pre-loaded with “PRIORI”, an app that securely recorded and encrypted all outgoing speech from all telephone calls sent and received. Two android-based phones were used, the Samsung Galaxy S3 and S5. Weekly assessment calls with a clinician included a depression and mania rating scale (HamD) and (YMRS). The encrypted calls were transferred to a central secure server for processing. Pre-processing included a de-clipping algorithm as it was found that the S3 was “clipping” the audio. The audio was segmented and features of rhythm were extracted and support vector machines were used to classify the speech. This presentation focuses on a subset of 217 acoustic features and solely on assessment calls with the clinician. Results Initial analyses included 37 individuals with two or more episodes wherein HamD > 10 and YMRS 10 and HamD Discussion Digital phenotypes derived from speech captured from mobile devices predict mood states. There are many challenges in addressing the comparability of data collected from across devices with different acoustic sensitivity. We demonstrate that through preprocessing, feature extraction, and data modeling techniques it is possible to mitigate the effects of differing amounts of clipping, loudness, and noise. The goal for a digital phenotyping system is to be passive, requiring no active input from the patient or the clinic. This will greatly improve phenotyping in genetic research over the current methods using structured clinical assessments using standard categorical criteria; digital phenotyping is likely to be able to classify and refine human disease by analyzing physiological patterns (such as speech) over time. However, several confounding factors need to be addressed and the refinement of techniques developed in this study increases device comparability in determining the digital phenotypes.