Developing a Music Listening mHealth Intervention for Stress Reduction in Early Recovery

Overview

The overarching goal of this study is to develop and examine the feasibility of a music-listening intervention that can be deployed in "real time" to regulate emotions and reduce momentary stress among young adults within the first 12 months of recovery from alcohol use disorder. We design the study with two phases to address three aims: Phase I includes the first two aims. For Aim 1, we will conduct formative research with a sample of young adults (N = 30) who have are within 12 months of recovery to identify features of music selections that are most effective in reducing momentary stress in real-world, ambulatory settings. For Aim 2, we will focus on developing mobile health technology that uses passive sensing and machine learning to automatically predict moments of heightened stress in real-time and suggest specific musical selections when stress is detected. During Phase II (Aim 3), we will test the feasibility of a novel music-listening intervention among a second unique sample of young adults who are within 12 months of recovery from AUD (N = 30). This protocol refers only to Phase I of the larger study, which focuses on observing music listening preferences and physiological and mental stress among people in early recovery from alcohol use disorder.

Participants for this study will be recruited either from regional mental health agencies or from ongoing or completed WSU College of Medicine studies. Study Procedures are as follows. Online pre-screening. Interested potential participants will first be directed to a secure online screening questionnaire that assesses basic inclusion criteria (age between 18 - 35 years, early-stage recovery, and own a smartphone with a data plan. The pre-screening survey will be administered via REDCap. The pre-screening survey will also include preferred contact information, including mobile phone, text, and email address. Interested individuals who pass the initial screening, will be contacted by study staff in order to schedule an in-person visit where the subject will be asked a series of more detailed screening questions to determine their eligibility for the study. Those who do not meet the pre-screening criteria will be informed that they are not eligible for the study. Informed consent and In-person screening. Those who endorse the pre-screening criteria will take part in an in-person study entry interview where they will provide written informed consent and complete additional measures assessing eligibility. The interview will be conducted by a trained research coordinator and will take place in a private room located in private offices reserved for WSU faculty in either Spokane, WA or Pullman, WA. Informed consent will be assembled in writing for each participant to read and take home if they wish. Research coordinators will walk through the informed consent packet with the individual before they begin participation. The consent form will be signed electronically via REDCap, and the participant may take the paper copy home with them. Prior to initiating the informed consent process, potential participants will also be asked to provide breath samples to determine their blood alcohol content (BAC). Participants whose BAC level is greater than 0.00, but less than 0.05, will be asked to either remain in the clinic until breath results indicate otherwise, or given the option of rescheduling their visit. Participants whose BAC results indicate impairment (\> 0.05) will be given the option of either waiting in the clinic until BAC \< 0.5 or, alternately, research staff will offer to schedule a ride-share company (e.g. Uber) to drive them to their home. After reviewing the informed consent document with each participant, the research staff will administer the in-person screening survey. This survey will include the following components: 1) the Patient Health Questionnaire (PHQ-9); 2) the Ask Suicide-Screening Questions (ASQ) tool; and 3) the Alcohol Symptom Checklist (ASC). The nine-item version of the Patient Health Questionnaire (PHQ-9), will be administered to ensure the absence of depressive symptoms. We will exclude individuals who indicate they are experiencing at least severe depressive symptoms, operationalized by a score of 20 or higher on the PHQ-9. In addition to the PHQ-9, determination of imminent risk of suicide risk among potential subjects will be assessed using the Ask Suicide-Screening Questions (ASQ) tool developed by the National Institute of Mental Health (NIMH). The ASQ tool is a set of four brief suicide screening questions that takes less than 5 minutes to administer. If a subject answers "No" to the four questions, screening is complete for that subject and no intervention is necessary. If a subject answers "Yes" to any of the four questions, or refuses to answer, they will be considered a positive screen and an additional assessment will be administered to determine potential risk vs. imminent risk. If imminent risk is identified the clinical staff will be alerted immediately and the subject will be kept in sight. If potential risk is identified, the clinic staff will be notified and will administer a brief suicide safety assessment. The 11-item Alcohol Symptom Checklist is a self-report questionnaire that asks patients whether they have experienced each of the 11 Alcohol Use Disorder (AUD) criteria within the past year. Each of the 11 items on the Alcohol Symptom Checklist maps onto one the 11 criteria for AUD as currently defined by the Diagnostic and statistical manual of mental disorders, 5th edition, published by the American Psychiatric Association. Patients indicate whether each AUD criterion was present or absent within the past year and Alcohol Symptom Checklist scores reflect AUD criteria counts that range from 0-11. Endorsing 2-3 criteria, 4-5 criteria, or 6-11 criteria is consistent with DSM-5 definitions for mild, moderate, or severe AUD, respectively. Participants who endorse at least 2 criteria will be eligible for the study. Baseline Survey and Orientation Session. Participants who meet full eligibility criteria will be directed immediately to the online baseline survey, which will be completed on a laptop computer in the private clinic office using REDCap software. The survey is estimated to take 20 minutes to complete. After completing the baseline survey, the participants will be provided with a detailed explanation of the study procedures for each study component, focusing on instructions for wearing the sensor device and EMA survey components. Training for the wearable wristband will include information about how to wear and remove the wristband, and how to care for the wristband. The participants will also receive instructions about answering the online surveys and using the Spotify app during the orientation session. For participants with an existing Spotify account, this training will emphasize switching to the research Spotify account, rather than their personal account during the study period. As part of this process, participants with an existing Spotify account will be asked to transfer up to 5 of their Spotify playlists from their personal Spotify accounts to the research study Spotify account. To accomplish this, participants will follow these steps: 1) open their personal Spotify account; 2) open the desired playlist; 3) right-click on the playlist and select 'Invite Collaborators'; 4) share the link with the research study Spotify account; 5) log into the research study Spotify account and use the shared link to access the playlist; 6) save the playlist by creating a new personal copy that is not collaborative. The study coordinator will then assist the participant create a new playlist that is composed of the five songs indicated in the baseline survey (Survey item: "Please list below five (5) songs that you would use to calm down in a stressful situation"). For participants who do not have a personal Spotify account, the training will focus first on downloading the Spotify app to their mobile phone, followed by general instructions about using the Spotify app. The study coordinator will then assist the participant create a new playlist that is composed of the five songs indicated in the baseline survey (Survey item: "Please list below five (5) songs that you would use to calm down in a stressful situation"). Ambulatory Session. Following the orientation session, the participants will be asked to wear the sensor device during waking hours for fourteen consecutive days to provide continuous monitoring of physiological stress. Participants will also be asked to listen to music on their smart phone only through the Spotify app, using a Spotify premium account created specifically for the research study, to allow music listening history data to be collected. To identify moments of heightened stress, participants will be instructed to press an "event marker" button on the sensor device at any time they feel "more stressed, overwhelmed, or anxious than usual." The event marker will be time-stamped. Ecological Momentary Assessment. Our current data collection infrastructure, includes a front-end app (called CalmiFy) along with back-end server and databases for collecting self-reported EMA data. A brief EMA-based survey will be administered through the CalmiFy app interface to collect participants' self-reported perceptions of acute stress, negative and positive emotions, and alcohol-related cravings. To capture these dynamic processes, the EMA survey will be administered four times per day between 9am and 9pm across the 14 days, consistent with our previous research. If there is no response to the initial prompt, the respondent will be reminded at 5-minute intervals up to three times. After the 3rd reminder, the EMA survey will become inaccessible (i.e., expires). To reduce participant burden, each EMA survey is expected to require only 3-5 minutes to complete. Brief Check-In. Approximately 7 days after beginning the ambulatory session, the participants will be asked to attend a brief check-in session via Zoom. This session is expected to take between 15-30 minutes and will provide an opportunity to confirm that all study components are satisfactory and problem-solve any issues that need to be addressed. The check-in session will also include administration of a timeline-follow back (TLFB) measure of recent alcohol use. In this procedure, participants will first be presented with a chart of the U.S. Standard Drink definition and then asked to indicate the number of drinks consumed on each calendar day across the 7-day assessment period. A calendar method will also be used to review participants' recollections of stressful events that have occurred and to provide additional context for the machine learning algorithms, such as false positive and negative rates. Structured Qualitative Interview. Within 7 days of conclusion of the 14-day ambulatory session, participants will participate in a structured qualitative debriefing interview that will last about 20-30 minutes and will be audio-recorded for transcription. The interview will be structured to assess significant events that occurred during the study protocol and will be informed by the EMA and physiological data that may suggest particularly difficult or stressful days and/or parts of days. Participants will also respond to questions asking about any problems that may have occurred during the study (including if any songs increased their stress), interference of the study with daily-life routines, and overall satisfaction with the study procedures. The interview session will also include administration of a timeline-follow back (TLFB) measure of recent alcohol use. In this procedure, participants will first be presented with a chart of the U.S. Standard Drink definition and then asked to indicate the number of drinks consumed on each calendar day across the days since their brief check-in session. A calendar method will also be used to review participants' recollections of stressful events that occurred since the brief check-in and to provide additional context for the machine learning algorithms, such as false positive and negative rates. Participants will also be asked to return the Empatica EmbracePlus sensor device at the time of the interview. Statistical Methods Data collected in this study will comprise three types: 1) physiological data obtained from continuous monitoring via the EmbracePlus wearable device; 2) self-reported quantitative data obtained from baseline surveys and music intervention responses; and 3) qualitative data obtained from structured interviews at the conclusion of the pilot feasibility study. To account for missing data in the quantitative components, we will use full information maximum likelihood estimation (FIML) approaches,4 which have been shown to result in unbiased parameter estimates under many missing data situations in the context of longitudinal data, including under some violations of assumptions, which will also be assessed using Little's MCAR test. Self-Reported Data. Prior to main analyses, we will conduct preliminary data screening of the self-reported quantitative data. Descriptive statistics and preliminary Pearson correlation analyses will be conducted to determine the univariate relations among all variables. Attrition analyses will be conducted on study variables and sociodemographic characteristics to determine significant differences between groups. Univariate and multivariate assumptions will also be assessed. Data will be screened for outliers and missing data and analysis decisions adjusted accordingly and as needed. This comprehensive screening will ensure accurate analysis in the later steps of the analysis plan. Primary Outcomes: Physiological Data from Wearable Device. Preliminary steps will also assess the validity of the physiological data collected via the wearable sensor device, including EDA and HRV. We will use the recommended tools and procedure by the Empatica guidelines to remove artifacts and extract features of the EDA and HRV signals to be used in further analyses. The Empatica EmbracePlus computes the heart rate (HR) and the inter-beat intervals (IBI) from BVP (Blood Volume Pulse) signal. We will assess the validity of the IBI which provides heart rate variability (HRV). Next, we will investigate the associations of the EDA signal and the HRV signal with the self-reported outcomes. These association will be studied at the feature level for each of EDA and HRV measurements. The EDA peak detection analysis provides a set of features corresponding to each EDA peak. We will utilize the EDA-Explorer public scripts to detect the EDA peaks.6 Previous studies have shown peaks from the EDA signal correlate with emotional arousal in humans. Important EDA features include (1) EDA: the EDA value at Apex of the peak; (2) rise-time: time that takes the EDA peak to reach its maximum value; (3) max derivative; (4) amplitude (5) decay time: the time that takes the signal to drop from the Apex to the minimum of the peak; (6) SRC-width: the width of the peak (number of the samples); and (8) AUC: the area under the curve. For heart rate signals, we will compute statistical features that are considered time-domain indices. These HRV measures are directly extracted from the IBI/RR interval signals. The RR interval is the interval between two successive heartbeats. We will measure mean of the RR interval (MRR), standard deviation of the RR interval (STDRR), root mean square successive difference of the RR intervals (RMSSD), coefficient of variance of the RR intervals (CVRR), mean of the heart rate (MHR), and standard deviation of the heart rate (SDHR). MRR, STDRR, RMSSD are features that represent the HRV, while, MHR and SDHR are features that are extracted from the heart rate. Qualitative Data from Structured Interviews. The structured interviews will be audiorecorded for transcription and coding. Transcriptions will be verified by at least one team member. Systematic thematic analyses, a method for identifying, analyzing, and reporting patterns (themes) within data will be used to identify relevant themes from the interview data. Findings will be reported at the descriptive level, in which themes and illustrative quotes are provided. The resulting themes will be discussed among the research team to integrate results into development of potential adaptations to the proposed music-listening intervention. Statistical Analyses for Aim 1. (Ambulatory Assessment). We will first distinguish between episodes of high vs. low stress. Based on previous literature and our own pilot study, we defined time intervals as lasting 24 seconds. A time interval will be labeled as "stress" (1) if the participants indicate a moment of heightened stress by pushing the event marker button on the sensor device. All of other time intervals that do not include a stress event will be labeled as "nonstress" (0). Music listening episodes will be identified through both subjective and objective measures. Subjective music listening be coded "1" if the participants indicate that they are listening to music at the time the stress event marker on the sensor device was pushed. Objective music listening will be determined by comparing time stamp information of music listening history from the Spotify API with time stamp information for the stress event. Objective music listening will be coded "1" if music listening of at least 3 minutes occurs within the time of the stress event. Because the Spotify API stores exact dates and times of music listening history, we will generate the duration of music listening in this time frame and the time lag between the stress event marker and the most recent track as continuous variables. Analyses for Aim 1 will begin with basic descriptive statistics and examining the correspondence between music listening intervals (no/yes) and heightened stress intervals (no/yes) using chi-square test of independence. Next, we will identify music characteristics that correspond to intervals defined by moments of heightened stress, compared to non-stressed intervals. In these analyses, we will use paired-sample t-tests to compare mean values of the music features provided by Spotify API (e.g., energy, acousticness) across stress and nonstress intervals. The next set of analyses will account for the nested structure of the data. Thus, we will use generalized linear mixed regression to more formally test hypotheses. Level 1 variables comprise those that are assessed repeatedly, level 2 variables comprise individual-level variables assessed once. In the analyses, both unconditional and conditional models will be specified. Comparison between these models will proceed by calculating the reduction in deviance as a measure of model fit. All models will include participant characteristics, such as biological sex, music training, and alcohol use history as level 2 covariates. The first model will test whether music-listening (no/yes) was associated with the probability of experiencing stress (no/yes). The next models will examine music characteristics provided by the Spotify API (e.g., energy, valence) and duration of music listening as additional level 1 predictors of stress. For these models, only time points that include music listening will be included in the analyses. The music features will be identified from the most recent music selection, relative to the stress event marker. Separate models will be specified for each music feature in the first set of models, followed by a full model that includes all music features entered simultaneously. The music features are level 1 predictors and will be person mean-centered. In the final set of models, stress will be modeled as a function of the emotion regulation strategy (each coded as 0/1, with 1 indicating this reason was chosen). Initial models will include separate models for each regulation strategy entered alone and with covariates. Next, a full model with all regulation strategy predictors entered simultaneously. Statistical Analyses for Aim 2. We will validate our stress estimation algorithm using data collected in Aim1 as well as historical data from our prior study, which involved N=11 patients with AUD, and an ongoing study with N = 28 healthy young adult participants. After training the model, we will evaluate the model on various metrics such as precision, recall, F1-score, Receiver Operating Characteristics (ROC) curve, accuracy, and loss. Recall measures the total number of positive samples among all positive samples that were classified correctly by the classifier. The recall is also called true positive rate (TPR) or sensitivity and represents the classifier's ability to find all the positive samples. Precision is defined as the number of true positives over the number of true positives and false positives. The F1 score is a weighted average of precision and recall with the best value of 1 and the worst score of 0. The ROC curve is a graph of true positive rate vs. false positive rate. ROC shows the performance of a classification model at all classification thresholds. Precision, recall, F1-score, and ROC curve are important model evaluation metrics but become significantly more valuable when we have unbalanced data for training and testing the model. We will measure the performance of the model for cross-subject transfer learning. We will also examine how many labeled instances are needed from each participant in order for the semi-supervised learning approach to achieve 90% accuracy in stress detection using the training dataset labeled through the semi-supervised learning algorithm. Finally, we will implement the model on a smartphone. Our goal will be to optimize the model size such that the algorithm can perform real-time machine learning for a full day prior to recharging the phone and without compromising the accuracy performance.