Can passive measurement of physiological distress help better predict suicidal thinking? There has been growing interest in using wearable physiological monitors to passively detect the signals of distress (i.e., increases in autonomic arousal measured through increased electrodermal activity [EDA]) that may be imminently associated with suicidal thoughts. Before using these monitors in advanced applications such as creating suicide risk detection algorithms or just-in-time interventions, several preliminary questions must be answered. Specifically, we lack information about whether: (1) EDA concurrently and prospectively predicts suicidal thinking and (2) data on EDA adds to the ability to predict the presence and severity of suicidal thinking over and above self-reports of emotional distress. Participants were suicidal psychiatric inpatients (n = 25, 56% female, M age = 33.48 years) who completed six daily assessments of negative affect and suicidal thinking duration of their psychiatric inpatient stay and 28 days post-discharge, and wore on their wrist a physiological monitor (Empatica Embrace) that passively detects autonomic activity. We found that physiological data alone both concurrently and prospectively predicted periods of suicidal thinking, but models with physiological data alone had the poorest fit. Adding physiological data to self-report models improved fit when the outcome variable was severity of suicidal thinking, but worsened model fit when the outcome was presence of suicidal thinking. When predicting severity of suicidal thinking, physiological data improved model fit more for models with non-overlapping self-report data (i.e., low arousal negative affect) than for overlapping self-report data (i.e., high arousal negative affect). These findings suggest that physiological data, under certain contexts (e.g., when combined with self-report data), may be useful in better predicting—and ultimately, preventing—acute increases in suicide risk. However, some cautious optimism is warranted since physiological data do not always improve our ability to predict suicidal thinking. In recent years, there has been increased interest in using wearable devices (e.g., smartwatches) to study psychological constructs of interest in the real world, such as detecting signals of distress that may predict the onset of suicidal thoughts [1,2,3]. Periods of suicidal thinking can occur rapidly and be highly distressing [4,5,6], possibly escalating quickly to a level that interferes with the cognitive resources needed to ask for help or use skills learned in therapy. If a wearable monitor could passively detect this distressing state, it would provide opportunities for the deployment of just-in-time adaptive interventions [7]. Such interventions would be particularly useful for groups of individuals at elevated risk for suicide [8], such as those who have recently discharged from inpatient psychiatric care for suicide risk (the focus of this study).There is a long history of laboratory research supporting the promise of passively detecting distress that may characterize or precede periods of suicide risk. The psychological experience of distress is reliably associated with sympathetic autonomic activity. This activity can be indirectly detected by observing the small increases in perspiration that occur during an autonomic event [9, 10]. Because sweat is a good conductor of electricity, changes in skin conductance (also called electrodermal activity; EDA) [11] signal when an individual is distressed. Increases in EDA are associated with laboratory-induced distress in the form of social comparision [10] or watching a disturbing film [12]. Individuals at risk for suicide exhibit increased physiological reactivity to stress (i.e., increased skin conductance) [13] and this reactivity distinguishes those who are suicidal from those who are depressed but not suicidal (for review see Sarchiapone et al. [14]). This aligns with clinical observation and increasing empirical findings that periods of high suicide risk tend to be characterized by the high arousal affective states (e.g., agitation [15]) that are potentially most easily detectable by monitoring autonomic activity.Although there is optimism about the possibility of using wearable monitors to identify periods of risk for suicidal thoughts among those who are at risk for suicide, there are important questions that must be answered about the predictive ability of autonomic arousal before developing interventions that rely on wearable passive sensing. The goal of this study is to begin answering these questions.Is passively detected distress associated with periods of suicidal thinking?It is currently unknown whether passively detected distress (i.e., increased EDA) measured with a wearable device is associated (either concurrently or prospectively) with periods of suicidal thinking. Importantly, what is found in a controlled laboratory environment—the settings used for the vast majority of such studies to date—may not be as clear in the uncontrolled environment of everyday life. For example, changes in temperature, exercise, being in a hot room, and wearing the device incorrectly can all contribute to noise and may do so to the point of degrading the validity of physiological data [16]. Thus, even though wearable devices are becoming cheaper and easier to use, the technology is still prohibitively expensive to deploy on a large scale based on laboratory research alone. Consequently, it is important to establish whether physiological signals of distress as measured with wearable devices are associated with periods (hours and days) of suicidal thinking.Does passively detected distress add to our ability to predict periods of suicidal thinking over and above self-reports of affect?The ‘gold standard’ for assessing elevations in distress associated with suicide risk involves technology like smartphone-based ecological momentary assessment (EMA) which captures self-reported affective experiences and suicidal thinking in the moment throughout the day [17]. In most initial cases, wearables will not be used as a stand-alone method of detecting risk but will rather be paired with the current ‘gold standard’ of EMA. One key step is to see whether passive detection of distress improves the concurrent and prospective prediction of periods (i.e., hours and days) of suicidal thinking beyond self-report alone. This is important from a scalability perspective because it is still cheaper and easier to deploy monitoring solutions that rely solely on self-report (though certainly, self-report monitoring may require more active engagement for patients/participants). If wearable devices do not add to what self-report can tell us, it may suggest decreased utility of using wearables when self-report is available, and participants are willing/able to respond to these questions.The present studyIn this study, we were interested in (1) the contemporaneous and prospective associations between physiological assessments of distress (i.e., EDA) and self-reported suicidal thinking as well as (2) the incremental predictive validity of physiological assessment of distress above and beyond self-report, across both the presence/absence of suicidal thinking and the severity of suicidal thinking. Since we consider the aims and analyses in this paper to be exploratory, we have few specific hypotheses. Generally, however, we expected that if passively detected physiological distress adds to our ability to predict self-reported suicidal thinking, it would be most likely to do so in cases where the self-report items do not assess states known to be strongly correlated with these same physiological metrics. For example, increases in EDA are more strongly tied to high arousal, rather than low arousal, affect [18, 19]. When self-report items assess high arousal affect, this may contribute to overlap and therefore redundancy between the self-report ratings and the physiological data. Thus, we expect passively-detected distress to improve the characterization and prediction of suicidal thinking when combined with assessments of low arousal versus high arousal affective states, because we hypothesize that passively-detected distress is less redundant with low-arousal affect than with high-arousal affect.Participants were 25 adult inpatients who were hospitalized due to suicidal thoughts or suicidal behavior were recruited from an inpatient psychiatry service from July 2019 until March 2020 at Massachusetts General Hospital as part of a Harvard University IRB-approved registered clinical trial (NCT03950765) testing a novel smartphone intervention. Inclusion criteria were (1) admission due to severe suicidal thinking or a suicide attempt, (2) access to a smartphone, (3) willingness to wear the physiological monitor, and (4) absence of any factor that would preclude capacity to consent (e.g., acute psychosis, drug withdrawal), which was independently confirmed by clinical staff. Our sample size was determined based on power analyses that conservatively assumed 50% compliance (three out of six surveys) over 28 days (i.e., 84 responses per participant). We exceeded this number of assessments (97.08 responses per participant).ProceduresRecruitment/consent/baselineEligible and interested participants provided informed consent and completed baseline measures assessing demographics, history of suicidal thoughts and behaviors, and other trait-level factors. We used only the demographic questionnaire from the baseline session in this study. Participants also were asked to install on their phone a set of apps that allowed us to send surveys and retrieve data from the wearable device.Monitoring periodThroughout their inpatient stay and for 28 days afterward, participants were asked to complete on their smartphone six brief surveys per day. (All participants also received during the inpatient period up to three in-person therapy sessions.) The surveys were hosted on Qualtrics and delivered using the LifeData smartphone app. (We used LifeData to deliver Qualtrics surveys because doing so allowed us to have the benefit of the direct customization over randomization of prompts and the aesthetics of the interface allowed in Qualtrics and the delivery methods (e.g., push notifications) available in LifeData.) The surveys were delivered at random times within pre-specified windows. (Because this was a treatment study, 3 of the 6 daily prompts were randomized to include an opportunity to practice the skills learned in treatment. Before and after each skills practice prompt, participants completed a set of questions assessing a variety of affective states (described below). Because this manuscript is not concerned with the effect of the intervention, we used in these analyses the data from the pre-practice prompts, but not the post-practice prompts. The other three assessments included only the assessment items. Thus, the data we used in this study consisted of the responses to the pre-practice prompts and the assessment prompts.) Participants also were asked to wear on their wrist the Empatica Embrace 2 (www.Empatica.com), a physiological monitoring device that assesses movement (via 3-axis accelerometer), orientation (via a gyroscope), skin temperature, and electrodermal activity. It has been well-validated for its consumer use as an FDA-approved seizure detection device and uses similar technology and sensors as other validated [20,21,22] research-grade wearables made by the same company. The Embrace syncs to a secure cloud server through the Empatica Mate smartphone app. Participants were asked to wear the device 24 h a day, except for when showering or other times when the device could get submerged in water. We suggested participants charge the device while showering.Measures and feature creationAffectAt each prompt, participants were presented with a list of affective state labels and a definition for each state. They were asked to rate each label in regard to the current moment on a 0 (not at all) to 10 (very much) scale. Relevant to this study were five specific negative affect states, categorized into high/low arousal based on the circumplex model of affect [23, 24]. There were three low-arousal negative affect states: (1) fatigued, (2) hopeless, and (3) burdensome and two high arousal negative affect states: (4) agitated and (5) angry.Suicidal thinkingWe used a three-item measure of suicidal thinking assessing in the present moment, which has been used in our prior studies [5]. The items assessed the strength of (1) urge to die by suicide, (2) the intention to kill oneself at some point during the next day, and (3) the ability to resist the urge to die by suicide. All items were on a 0 (not at all) to 10 (very strong) scale. In line with prior EMA studies [5], we averaged these items to create a suicidal thinking composite with high internal consistency (alpha = 0.82) (Internal consistency was calculated according to Nezlek’s [25] approach that uses an unconditional three-level model with responses nested within measurement occasion nested within people.). The item assessing ability to resist the urge to die by suicide was reverse-coded.Autonomic eventsThe Embrace records EDA at 4hz using three stainless steel electrodes mounted on the bottom of the watch case. Once these data are transferred to the server, a proprietary algorithm run on the server classifies autonomic events. Specifically, the algorithm identifies increases in skin conductance level (i.e., tonic EDA) that occur in the absence of increases in temperature and movement (since increases in temperature and movement could be signs that increased EDA is due to being in a hot room or physical activity). The algorithm removes any increase in EDA due to a potential ‘storm’ of rapid EDA changes during sleep [26, 27]. Evaluation of this algorithm in 46 adults found high sensitivity (97%) and a low rate of false positives (0.83/day) (Matteo Migliorini, Ph.D., Empatica s.r.l., Email Communication, April 2021).Data preparationPhysiological data were collected continuously, meaning that autonomic events could have been recorded at any time of the day. Self-report data were collected six times daily over a participant-defined 14-h period (most participants chose a window that lasted from 9 am to 11 pm). This means we would likely not have self-report data at or near the time of most autonomic events. To address this, we aggregated our data into two levels, hourly (i.e., within-day) and daily (i.e., be
https://www.nature.com/articles/s41398-021-01730-y
Can passive measurement of physiological distress help better predict suicidal thinking?
