The long-term goal for our PRIORI study is to use speech and behavioral patterns identified on a smart phone to predict the risk for developing an episode of mania or depression. Detection of risk will generate a warning signal of health state change and trigger an early warning signal (EWS). This EWS alert, in turn, will allow for timely intervention so the episode may be avoided or managed to reduce the severity and consequences including hospitalization and suicide. In practical usage, persons will work with their health care professionals to determine who, in addition to themselves, will receive the EWS alerts when a change in health state is detected. Alerts may to be sent to multiple persons to include specific family member(s), close friends, or a health care provider.
Current Results and Status
Our research team's current focus is to narrow down the elements of the early warning signal in the speech signal. This is important in creating an efficient system that can be processed on the mobile device using minimal energy from the battery. Currently, the speech signal is encrypted and uploaded to a central server for detailed analysis. We initially examined several fundamental features of speech, such as rhythm, in the assessment calls with the clinical researcher (symptom severity assessments) and found that once the computational programs were ‘trained’ on known data, it was possible to correctly classify the mood states with ~70% accuracy. The addition of the personal calls into the analyses was done with a program called “iVector” and provided a background matrix for the individual against which the assessment calls were evaluated; this improved the accuracy somewhat (~78%). While we were encouraged by these findings, we are not satisfied and are striving to improve the approach and methods.
PRIORI.3 is in the final testing stages and will be deployed to study participants very soon. The second version of the program (PRIORI.2) failed due to an inability to meet security standards, but provided a significant learning experience that greatly informed PRIORI.3.
Strategy and Direction
Strategies to improve accuracy are ongoing. Expression of mood states is personal and shows variability between people. There are constituent elements of speech and emotion that can be identified by the human ear (e.g. anger) and we have a cadre of listeners that are reviewing and rating the elements of speech sounds. This is to ‘train’ the computer based on how an actual human listener rates the speech. Speech and behavior occurs in context and therefore we will use the data from the GPS to position the individual, asking if there is variability and predictability that is related to time and place.
We have expanded our technical team and are working with an expert local programming team, Arbor Moon Inc., that will provide ongoing monitoring and updates that are needed in the dynamic mobile environment. We are rigorously testing devices to ensure that we can securely collect personal data from participants to determine the signals that are uniquely predictive of mood state changes.
Over the next two years we have substantial technical tasks that include expanded and accelerated data collection using PRIORI.3, and analyzing these data to systematically narrow down and characterize the personal signal that is most predictive of health states and health state changes.
PRIORI Collaborations
We are constantly pursuing opportunities for funding support through government or institutional agencies. The following projects are currently underway and each services to refine and advance the overall PRIORI goals. The common goal facilitated by PRIORI in these projects is the capacity to predict and anticipate the risk of an episode in sufficient time to intervene.
Active Projects with Federal or State Funding:
- Health Monitoring of Acoustic and Behavioral Patterns in Bipolar Disorder across Cultures. A collaborative grant between the University of Michigan and the Institute for Development, Research, Advocacy, and Applied Care (IDRAAC) of the University of Belamand, Beirut, Lebanon. The goals of this project are to identify fundamental elements of acoustics that associate with mood states regardless of the culture.
- Predicting Suicide: a longitudinal analysis of speech patterns in a high-risk sample. A collaborative grant between Brown University and University of Michigan (R01MH108610.) The goal of this proposal is to use computational speech analysis to determine the predictive utility of patterns of risk variation derived from unstructured speech on suicide risk during the 6-month period following psychiatric hospitalization.
- Integrated Self-Management Apps to Enhance Outcomes for Medicaid Consumers* (Medicaid Match funding). A collaborative project with Amy Kilbourne, Ph.D., of the Health Services Section of the U-M Department of Psychiatry. This project is sponsored in part by the State of Michigan to develop programs that will directly benefit individuals who receive care in the community mental health system.
*The Medicaid Match funded project is the first project for PRIORI that combines with an intervention program, LifeGoals – a program developed by Amy Kilbourne Ph.D., Professor of Psychiatry, and is designed to provide tools for the individual with bipolar to enhance quality of life, setting goals and measuring success. PRIORI and LifeGoals will be provided to the Medicaid consumers and tested to determine the effects on measures of quality outcomes.
Pending applications: We have a federal application pending. The goals of this project are to determine the detailed elements of acoustics that drive the predictive capacity of PRIORI. We are also preparing applications for internal competitions focused around the University of Michigan Precision Health Initiative.
National Collaborations: Considerable interest has been generated around the PRIORI program and discussions are underway with several entities that are both academic and industry-based. We have established a formal relationship with the Pine Rest Christian Mental Health clinics, centered in Grand Rapids, MI, and the University of Illinois at Chicago (UIC). We collaborated with UIC investigators around a program called ‘BiAffect,' designed to measure keyboard related behaviors on a mobile device as a proxy of executive functioning. This project won a national competition sponsored by Apple.
International Collaborations: Further interest at the international level has generated collaborations with the Universities of Bergen and Oslo in Norway. Our Norwegian colleagues wish to incorporate PRIORI into their monitoring program for mental health. We meet by teleconference monthly and in the summer of 2018, two computer scientists will spend time at the University of Michigan to study details of the computational approaches ongoing in our labs. Additional potential projects include a research team in the University of Monterey, Mexico and the Shanghai Mental Health Center, where testing of PRIORI is presently occurring.