Wearable sensor-based detection of stress and craving in patients during treatment for substance use disorder: A mixed methods pilot study

Abstract

Aims:

To determine the accuracy of a wearable sensor to detect and differentiate episodes of self-reported craving and stress in individuals with substance use disorders, and to assess acceptability, barriers, and facilitators to sensor-based monitoring in this population.

Methods:

This was an observational mixed methods pilot study. Adults enrolled in an outpatient treatment program for a substance use disorder wore a non-invasive wrist-mounted sensor for four days and self-reported episodes of stress and craving. Continuous physiologic data (accelerometry, skin conductance, skin temperature, and heart rate) were extracted from the sensors and analyzed via various machine learning algorithms. Semi-structured interviews were conducted upon study completion, and thematic analysis was conducted on qualitative data from semi-structured interviews.

Results:

Thirty individuals completed the protocol, and 43% (N=13) were female. A total of 41 craving and 104 stress events were analyzed. The differentiation accuracies of the top performing models were as follows: stress vs. non-stress states 74.5% (AUC 0.82), craving vs. no-craving 75.7% (AUC 0.82), and craving vs. stress 76.8% (AUC 0.8). Overall participant perception was positive, and acceptability was high. Emergent themes from the exit interviews included a perception of connectedness and increased mindfulness related to wearing the sensor, both of which were reported as helpful to recovery. Barriers to engagement included interference with other daily wear items, and perceived stigma.

Conclusions:

Wearable sensors can be used to objectively differentiate episodes of craving and stress, and individuals in recovery from substance use disorder are accepting of continuous monitoring with these devices.

1. Introduction

Treatment options for substance use disorder (SUD) focus largely on a combination of behavioral and pharmacologic therapies, however current literature quotes rates of return to use between 40–90% depending on the substance and treatment method involved.(Back et al., 2014) The ability to objectively identify moments of greatest need and to deliver just-in-time targeted interventions would provide a tremendous advantage over the current model of care. Stress and craving are two important neurobehavioral phenomena that have been linked to return to drug use and are candidates for identification of high-risk periods.

Stress can be defined as a subjective experience to events that are perceived as harmful, threatening, or challenging.(Sinha, 2001) Both clinical and preclinical literature link stress to poor decision making and drug use, including return to use even after periods of prolonged abstinence. However, the directionality of the relationship between stress and drug use is not straightforward; there is also evidence that drug dependence results in neurobiological changes that augment individual response to stress.(Sinha, 2001)

Craving, also linked to return to drug use, is defined as the subjective experience of wanting to use a drug- it is a conscious expression of desire, specific for an individual’s drug of choice.(Tiffany and Wray, 2012) Craving is considered an important feature of SUD, as evidenced by its inclusion in the Diagnostic and Statistical Manual of Mental Disorders (DSM 5) diagnostic criteria of SUD, and it may be one of the strongest predictors of return to use after recovery.(Kavanagh and Connor, 2013) It has been studied in relation to diagnosis, prognosis, and clinical outcomes, and as a target for intervention.(Tapper, 2018; Tiffany and Wray, 2012)

A strong relationship between stress and craving in SUD has been repeatedly demonstrated, most notably whereas stress states increase the risk of craving. Interestingly, some hypothesize that stress physiologically and emotionally resembles a withdrawal-like state, triggering drug craving.(Sinha, 2001) Stress has been positively correlated to craving for opioids, cocaine, and tobacco, and to return to use in individuals undergoing treatment for SUD.(Brady and Sonne, 1999; Preston and Epstein, 2011; Sinha, 2001, 2008; Sinha et al., 1999) Both stress and craving herald high-risk events in recovery, and their identification may predict risk of return to use in susceptible individuals. Additionally, both stress and craving have recognizable physiologic correlates that create opportunities for objective detection. Stress is associated with notable increases sympathetic nervous system activity, which can easily be inferred though noninvasive proxy measures such as increased heart rate, skin conductance and locomotion parameters. Craving demonstrates unique neurobiological changes(Sinha, 2009), and is also associated with stress-like changes in autonomic arousal and reactivity.(Weinstein et al., 1998).

Wearable sensors are powerful mobile health (mHealth) tools that are noninvasive, well received by patients, and leverage the current trend of wearable wellness technology.(Carreiro et al., 2015b; Garbarino et al., 2014) They are under investigation for the detection of other clinically important conditions(Fletcher et al., 2011; Indic et al., 2012; Poh et al., 2012a; Poh et al., 2012b), including episodes of drug use,(Carreiro et al., 2015a; Carreiro et al., 2015b). Wearable sensors have notably been used to detect stress in both laboratory and real-world settings for a number of other populations(Can et al., 2019; Hovsepian et al., 2015; Kaczor et al., 2020; Leonard et al., 2018; Rodrigues et al., 2018). While wearable sensors provide a convenient medium to detect and differentiate stress and craving in real time, to our knowledge, no literature exists on the application of this technology to SUD treatment.

The ability to employ wearable technology would offer those in treatment for SUD and their provider(s) a more objective method of monitoring potential for return to use and prime opportunities to deploy just-in-time interventions in real world settings. As an initial step toward this goal, the present study aims to: 1) Determine the accuracy of a wearable sensor to identify and differentiate drug craving and stress in participants enrolled in an outpatient treatment program for SUD; and, 2) assess acceptability, barriers, and facilitators to continuous sensor-based monitoring within this population.

2. Materials and Methods

2.1. Study Protocol

This was an observational mixed methods pilot study and was approved by the institutional review board at the University of Massachusetts Medical School. Informed consent was obtained from all participants. A convenience sample of participants was recruited from two affiliated outpatient substance use treatment facilities located in New England, with an average total census of 275 clients per site. Potential participants were identified by treatment providers and those that opted to participate were enrolled during routine treatment visits. Inclusion criteria were: age 18 years or older, enrollment in an outpatient treatment program for SUD, and ability to speak English and to provide informed consent. Exclusion criteria were: physical inability to wear a wrist-mounted sensor (i.e. upper extremity amputation or fracture), incarceration, or pregnancy.

2.2. Hardware

The E4 wearable sensor (Empatica, Milan Italy, Figure 1) is similar in form factor to a wristwatch and was used to collect all physiologic data. This noninvasive sensor continuously measured and recorded the following parameters as surrogate makers for changes in sympathetic nervous system (SNS) arousal: electrodermal activity (or EDA, in microsiemens), skin temperature (in degrees Celsius), three-dimensional acceleration (in units g), and heart rate (in beats per minute). The E4 has a push-button interface that allows for data annotation. Physiologic data was stored in the E4’s on-board memory until downloaded during follow-up visits.

Figure 1:

Empatica E4 wrist-mounted sensor

2.3. Data Collection

Each participant was enrolled in the study for four consecutive days. During the enrollment visit a brief (10-minute) training session to demonstrate the proper techniques for wearing, using, and charging the sensor was provided. Participants were then instructed to wear the E4 on their non-dominant wrist during all waking hours for the duration of the study period and to only remove the sensor for charging while sleeping. They were instructed to press the E4’s event marker button to “tag” whenever they perceived stress or craving. Given the goal of capturing episodes of subjective phenomena (craving and stress) in real time, participants were instructed to self-report any time they experienced a sensation of craving or stress that was significant enough to interfere with their regular daily activities. Craving was defined as desire to use their drug of choice. Stress was defined as mental or physical discomfort due to an experience that was perceived as threatening, challenging, or harmful. Participants were also asked to keep a written log of brief descriptions for each tagged event and also to note any periods of exercise or physical activity more strenuous than a brisk walk.

Physiologic parameters continuously collected using the Empatica E4 throughout the study period included: EDA (sampling frequency of 4 Hz), heart rate (sampling frequency of 64 Hz), accelerometry (sampling frequency of 32 Hz), and skin temperature (sampling frequency of 4 Hz).

Demographic and historical data (including medical history, current medications, and substance use history) were collected during the baseline visit. Throughout the study period, data on protocol compliance (number of hours per day of physiologic data collection), number of interactions with the device (number of tags), and number of events logged were tracked. Semi-structured exit interviews were conducted with each participant to learn more about their perceptions of the sensor and experiences using it in this context. Specifically, participants were asked questions regarding usability, acceptability, barriers/facilitators to use and continued use, degree to which the technology was bothersome, perceived degree to which continuous physiologic monitoring affected behavior, and interest in using the system for a longer period of time. Non-physiologic study data were collected and managed using Research Electronic Data Capture (REDCap), a secure, online web platform for data management.(Harris et al., 2009)

2.4. Physiologic Data Analysis

2.4.1. Classification, Segmentation, and Feature Extraction

Forty-minutes of physiologic data (20 minutes pre- and post-event annotations) were segmented from the original recordings and labeled as “stress” or “craving” based on participant annotations. Forty-minute “no-stress” segments were then randomly selected from the remaining data. These “no-stress” segments had a start and end times at least 60 minutes before or after annotated events. To develop a contrasting class for the craving class, data segments from the stress and no-stress classes were mixed, randomized, and designated as the “no-craving” class.

After extracting craving, stress, and no-stress data segments, each segment was subjected to a sliding window operation. This operation utilized a five-minute window that slid through the 40-minute data segments one step at a time, with each step incrementing by one-minute of data. This five-minute windowing operation was applied to each data segment until all the segments’ data points were covered by a five-minute window.

There was a total of 37 windows for each 40-minute data segment. MATLAB (MathWorks, Natick, MA) was used to find the instantaneous amplitude of each five-minute window by subjecting it to the Hilbert transformation. This is a powerful technique in signal processing to estimate the amplitude instantaneously from a rapidly fluctuating (nonstationary) data.(Benitez et al., 2001; Hahn, 1996) The estimated amplitude was fitted to a distribution and the gamma distribution with the scale as well as shape as the characterizing parameters best represented the amplitude distribution. The distribution that best represented the data was a gamma distribution. Hence, the scale (Sc) and the shape (Sh) parameters of the gamma distribution were calculated. Furthermore, the mean (M) and variance (V) of the sensor data for each five-minute window were calculated. In addition, the distance measure, $D_k=\sqrt{S{h}^2+S{c}^2}$ was defined. Thus, for each period containing data from three axes of accelerometry, heart rate, and electrodermal activity sensors, five characterizing parameters were obtained from the average of 37 values from each of: M, V, Sh, Sc and D_k. This yielded a total of twenty-five features ((3 + 2) × 5 = 25). The obtained features were cleaned by removing non-numeric values, empty values, values outside of physiologic range, and opcode values based on the software platform.

2.4.2. Machine Learning Classification, Train/Test, and Feature Selection

The following four classification cases were analyzed for the ability to distinguish between classes: craving vs. no-craving, stress vs. no-stress, craving vs. stress, and craving vs. no-stress vs. stress. Twenty five classification models were tested from the following categories: decision trees, discriminant analysis, logistic regression, naïve Bayes classifiers, support vector machines, nearest neighbor classifiers, and ensemble classifiers. The dataset was not split into a training set and testing set because of its limited size. Instead, to avoid overfitting the model, the dataset was subjected to 10-fold cross validation (separated into 10 folds, trained with nine folds, and tested on the other fold until each fold was used as a test set).

Three variations of sensor feature selection were performed. First, only the features of the accelerometer data from all three axes (x, y, and z) were selected. Second, for exploration purposes, every combination of the accelerometer axes (x, x-y, y-z, x-y-z, etc.) and their respective features were selected. Third, features from all accelerometer, heart rate, and electrodermal activity data were selected. Each classification algorithm was evaluated with the following standard metrics: sensitivity, specificity, accuracy, and area under the curve (AUC) of the receiver operating characteristic (ROC) curve.

2.5. Qualitative Data Analysis

Thematic analysis was used to code and analyze exit interview data. The coding structure was first developed based on deductive codes from the interview questions and then inductive codes from review of the interview transcripts themselves. Once the coding scheme was developed, the interview transcripts were coded by two coders (SC and BC) independently. The two coded transcripts were then compared to ensure comprehensiveness of coding and agreed-upon codes were entered into AtlasTi qualitative data analysis software version 8.2.4 (AtlasTi, Berlin). Conflicts between coders were discussed by both coders until a mutually agreed upon code selection was reached.

3. Results

3.1. Sample Demographics

Thirty-one participants were enrolled and thirty completed the protocol (Table 1). Participants had a mean age of 32 years, were 43% female (N=13), and were predominantly white. Thirty three percent were on medication assisted treatment (MAT, including methadone or naltrexone) at the time of the baseline visit. Half of the participants identified greater than one primary substance of use. Of the 15 participants who reported a single primary substance of abuse, 73% (N=11) reported a history of alcohol use and 27% (N=4) reported history of heroin use. Co-occurring psychiatric conditions were common, with 21 participants (70%) self-reporting a psychiatric diagnosis in their medical history.

Table 1:

Participant Characteristics (N = 30)

		Mean	Range
Age		32.3	18–67
Number of substances reported		3.9	1–10
		N	% of total
Female Gender		13	43%%
Race	White	25	83%
	Black/African American	0	0
	Asian	2	7%
	Other	3	10%
Ethnicity	Hispanic/Latino	2	7%
Duration of Abstinence from primary substance	< 1 month	4	13%
	1–6 months	13	43%
	6 months-1 year	10	33%
	1–5 years	3	10%
Primary Substance Use History	Single substance	15	50%
	Multiple Substances	15	50%
Primary Substance(s) Used*	Alcohol	17	56%
	Heroin	15	50%
	Other Opioids	7	23%
	Cocaine	9	30%
	Methamphetamine/Stimulants	3	10%
On MAT**	Any MAT	10	33.30%
	Naltrexone	8	26.70%
	Methadone	2	6.60%

^*Sum is > 100%, 50% of participants identified > 1 primary substance

^**MAT = Medication Assisted Treatment

3.2. Protocol Compliance and Event Annotation

Participants wore the sensor for a mean of 4.4 days (range 3–6), with a mean total recording time of 50.3 hours (range 18.5–71.5). On average, each participant tagged six events (range 1–17), with a mean of 4.5 (range 0–16) stress events and a mean of 1.5 (range 0–9) craving events. In total, 1,510 hours of physiologic data were captured from 30 participants, with 178 tagged events (46 craving episodes and 132 stress). Of note 4 episodes were tagged by participants as both stress and craving; these were excluded from the analysis due to insufficient amount of data to create a new classification category.

3.3. Machine Learning

After cleaning the data and removing events where data was truncated, there were 41 craving events and 104 stress events available for analysis. One hundred and ninety-eight “no-stress” segments were extracted for comparison. The craving class was designated as the positive class for most of the machine learning analyses. The authors previous work (Carreiro et al., 2016; Chintha et al., 2018) found that the accelerometer data features contain reliable information to distinguish the effect of drug use from baseline conditions. Hence, the first analysis explored whether similar accelerometry features could be employed to detect craving and stress events experienced by these participants. Additional expanded analysis included the features of heart rate and electrodermal activity.

3.3.1. Three-Axis Accelerometer Analysis

To determine which set(s) of accelerometer axes would yield the best classification performance, the classifiers were first trained on every combination of the accelerometer’s multi-axis data. As shown in Table 2, the trained machine learning models performed best by utilizing all the features from each of the three accelerometer axes. All analyses were trained and tested using 10-fold cross validation, and in all four classification schemes, a Fine Gaussian Support Vector Machine (SVM) model using all three axes (X, Y and Z) provided the best performance.

Table 2:

Best performing models for accelerometry based event classification

Accelerometer feature combination		Stress Vs No-Stress	Craving Vs No-Craving	Craving Vs Stress	Craving Vs Stress Vs No-Stress
X Axis	Accuracy	60.30%	66.00%	68.30%	46.50%
	Model	Cubic KNN	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM
Y Axis	Accuracy	56.90%	59.10%	60.50%	43.30%
	Model	Cubic KNN	Fine Gaussian SVM	Medium KNN	Weighted KNN
Z Axis	Accuracy	56.90%	60.70%	60.90%	43.60%
	Model	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM
X & Y Axes	Accuracy	69.20%	71.30%	72.80%	54.20%
	Model	Fine KNN	Fine Gaussian SVM	Fine Gaussian SVM	Weighted KNN
X & Z Axes	Accuracy	68.50%	70.20%	73.00%	53.60%
	Model	Weighted KNN	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM
Y & Z Axes	Accuracy	67.80%	67.50%	69.20%	52.30%
	Model	Fine KNN	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM
X, Y, & Z Axes	Accuracy	74.50%	75.70%	76.80%	59.30%
	Model	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM	Fine Gaussian SVM

SVM = Support Vector Machine

KNN= K Nearest Neighbors

Stress vs. No-Stress Analysis.

Designating the stress data set as the positive class, the sensitivity was 73.0% and the specificity was 76.0% (Figure 2A). The AUC of the model’s ROC was 0.81 (Figure 2B), and the model’s accuracy was 74.5%.

Figure 2:

Stress vs. No-Stress Analysis A) Confusion Matrix B) ROC Curve

Craving vs. No-Craving Analysis.

With craving as the positive class, the sensitivity and specificity of the model were both 76.0% (Figure 3A). The AUC of the model’s ROC was 0.82 (Figure 3B), and the model’s accuracy was 75.7%.

Figure 3:

Craving vs. No-Craving Analysis A) Confusion Matrix B) ROC Curve

Craving vs. Stress Analysis.

With craving marked as the positive class, the model had a sensitivity of 77.0% and a specificity of 76.0% (Figure 4A). The model’s AUC was 0.82 (Figure 4B), and its accuracy was 76.8%.

Figure 4:

Craving vs. Stress Analysis A) Confusion Matrix B) ROC Curve

Craving vs. No-Stress vs. Stress Analysis.

Classification performance for the craving, no-stress, and stress data sets was further explored by training the machine learning classifiers with all three of the classes. With craving set as the positive class, the sensitivity and specificity of the model were 68.0% and 82.0% respectively (Figure 5A). The AUC of the ROC curve was 0.82 (Figure 5B), and the accuracy of the model was 59.3%.

Figure 5:

Craving vs. No-Stress vs. Stress Analysis A) Confusion Matrix B) ROC Curve

3.3.2. Accelerometer, Electrodermal Activity, and Heart Rate Analysis

The classification accuracies for models trained using only accelerometer features were checked against the accuracies of models trained using accelerometer, EDA, and heart rate (HR) features. As shown in Table 3, based on the classification loss function, the models with ACC, EDA and HR performed better than the model with ACC only. We found that Linear Gaussian SVM was the best model among each of the classification sets tested.

Table 3.

Model Accuracies For Accelerometer only vs Multi-sensor Classification

Classification Case	Accuracy, Best Model and Loss Function
	ACC, EDA, & HR	ACC Only
Stress vs. No-Stress	81.30%	74.50%
	Fine Gaussian SVM	Fine Gaussian SVM
	0.08	0.13
Craving vs. No-Craving	82.10%	75.70%
	Fine Gaussian SVM	Fine Gaussian SVM
	0.09	0.13
Craving vs. Stress	80.70%	76.80%
	Fine Gaussian SVM	Fine Gaussian SVM
	0.09	0.12
Craving vs. Stress vs. No-Stress	66.90%	59.30%
	Fine Gaussian SVM	Fine Gaussian SVM
	0.20	0.26

3.3.3. Analysis of validity of No -Stress segments

While the 40 minute segments for craving and stress were defined by participant annotation, the no-stress segments were extracted based on the absence of participant annotations. While we assume that the no-stress segments do not have any stress or cravings, there may be scenarios where the subjects were feeling stress or cravings during these periods, however failing to report. To further understand the probability of these segments being misclassified, we re-tested the 198 no-stress segments by applying machine learning classification models with only acceleration to Stress vs. No-Stress, Cravings vs. Stress along with a new classification model Craving vs. No-Stress. A given no-stress segment was considered as positive (correctly labeled), if both the models, Stress vs. No-Stress and Craving vs. No-Stress classified the segment as no stress. Based on this analysis we found that out of 198 No-Stress segments, 131 (66.16%) were classified as no-stress with both the models, whereas 78 % were classified as no stress by the craving vs no stress model, and 76% were classified as no stress by the stress vs no stress model. The reduced detection rate with both the models suggests that some of the no stress segments may represent either cravings or stress, which we obtain as 17.17% (cravings), 15.15% (stress) and 1.51% undetermined. Thus in the craving vs no stress classification, if any of the no stress segments are stress segments it is more likely to classify as no stress, similarly in the stress vs no stress classification if any of the no stress segments are craving segments it is most likely to classify as no stress, thereby increasing the no stress detection rate in the two class models.

3.4. Content Analysis of User Experience Data

Based on exit interviews with the 30 study participants that completed the protocol, several themes arose in the following areas:

3.4.1. Sensor Appearance was Important

The majority of participants reported they would be willing to wear more than one sensor simultaneously, but the majority would prefer a more streamlined version to do so. Seven participants reported they would not wear more than one sensor (for example, one on each wrist) because it would look awkward and suspicious. Participants most common issue with the sensor appearance was that it was bulky, and they would prefer a more streamlined sensor for long-term wear. Even though the sensor was commonly mistaken for a fitness tracker by family and friends, four participants expressed concern that it would be perceived as a law enforcement tracking device, and three were actually asked if it was.

3.4.2. The Sensor had an Effect on Participants

The largest emerging theme with regard to effects of the sensor was an increase in mindfulness, or the process of being attentive to, and accepting of, one’s current experience.(Lindsay and Creswell, 2017) This came up with greater than two thirds of participants and was uniformly regarded by participants as positive and therapeutic. Twenty-two participants noted that they were more conscious of stress and/or craving while they were wearing the sensor. Participants reported that this made them think harder about their options for responding to the situation and that pressing the button on the sensor (to indicate stress or craving) was cathartic and helped them “let it go.”

“I pressed the button and was more aware of stress, which got my mind more focused to get off the craving.”

−25-year-old male in recovery from polysubstance use

“Pressing the button alleviated stress…and let it go- like meditation.”

−38-year-old male in recovery from opioid use disorder

The sensor-based interaction was also therapeutic in other ways, beyond mindfulness. Participants reported they felt “connected” by wearing the sensor (despite the fact that it was not yet connected to an application), and liked the idea that in the future, it could potentially notify their clinician/someone from their support system if they needed help. Two participants also cited a sense of accountability that came with continuous monitoring from the sensor, and that this helped them cope with stress in a positive way.

“I noticed feelings and mind/body reaction, pressing [the button] took care of feelings. Kind of like telling on yourself by acknowledging.”

−28-year-old female in recovery from opioid use disorder

“…took the time to think about what was going on, like a safety blanket. I rationalized my thoughts better with [the sensor] on.”

−24-year-old male in recovery from cocaine and alcohol use disorders

In contrast to the majority response that the sensor made them more mindful, two participants paradoxically described the sensor as a distraction in a positive way. Interacting with the sensor and logging their experiences provided a distraction from the situation, which was perceived as therapeutic.

Participants also reported they felt forced to quantify their stress for the protocol. Only one participant noted this as a negative experience (i.e. that it was “more stressful to think about stress”). The majority of participants reported that this was positive and helped them realize that the stress/craving event was less significant than originally perceived.

“I was more mindful of emotions and stress level, which helped me handle stress better and decide how important a stressor was by categorizing it. It made me realize that my day wasn’t as stressful as I thought.”

−44-year-old female in recovery from alcohol use disorder

“It made me question what really counts for stress/craving- knowing the boundary. Like what is enough to mark a point? Or where is this enough that it counts? Like a pain score. Where do you draw the line?”

−26-year-old male in recovery from alcohol and opioid use disorders

3.4.3. Positive Attributes of Sensor Technology

In addition to the positive effects mentioned previously, most participants report that the sensor was easy to use and described it as “simple, and straightforward.” The majority cited sensor use during the study period as an enjoyable (N=19) or neutral (N=11) experience; no participants described their overall experience as negative. Over two thirds of the participants reported that the device was noninvasive. Some even reported that they were surprised the sensor was so unobtrusive, and in fact, participation made them want to continue using a wearable device in the future.

“[I] liked the fact that you can notify someone if you are having a bad day.”

−33-year-old female in recovery from opioid use disorder

“It made me realize that I really want a FitBit.”

−24-year-old male in recovery from alcohol use disorder

3.4.4. Barriers to Sensor Use

Participants identified several barriers to adoption of the sensor for daily use. There were some concerns that it would interfere with work, either physically, or due to co-workers/clients’ perception of what the device was for. Additionally, some participants noted that they were not allowed to wear bracelets/watches during work and had to request an exception for the study.

“I wondered if my passengers thought I escaped from somewhere with a sensor strapped on.”

−26-year-old male ride-share service driver in recovery from alcohol and opioid use disorders

Three participants reported the sensor itself was physically difficult to use (i.e. the button was hard to press, the charger was hard to snap in place). Seven subjects questioned durability of the sensor because it was not waterproof, or they were generally concerned for risk of damage during daily activity. There were a small number of participants that reported having technical difficulty with the sensor, and that the learning curve to become comfortable with device functionality was one day.

The sensor occasionally interfered with things normally worn on the wrist (such as watches, bracelets, or fitness trackers). Participants expressed a desire for additional features to be incorporated into the study sensor to replace some of these items. Specifically, one third of participants expressed frustration that the device did not function as a watch.

“I kept looking at [the sensor] thinking it was a watch. Get me one that tells time.”

−67-year-old male in recovery from alcohol use disorder

3.4.5. Thoughts and Attitudes Toward Wearable Sensor Research

Participants report that bystanders were excited about the study, particularly those that were supportive of their recovery process (such as others in recovery, family, and friends).

“People thought [the study/sensor] was cool, especially the other people in recovery that I told about it. Most other people didn’t notice [the sensor].”

−28-year-old female in recovery from opioid use disorder

“I did wonder if the device would pick up on stress that I didn’t notice. The more people I told about [the study/sensor], the more I thought this will change the future of addiction treatment.”

−44-year-old female in recovery from alcohol use disorder

Participants expressed a desire for real time feedback from the device, which is a main objective for future iterations of the system. Three participants specifically stated that they would want to receive a notification from the device when stress or craving was detected.

“…[the device needs an] immediate feedback/reward system, or a special tone that signals a need to check in with different levels of stress…[should have] different interventions based on level of stress.”

−56-year-old male in recovery from alcohol use disorder

The majority of participants (N=28) were willing to wear the sensor beyond the study period, especially if a more streamlined sensor was available. Participants noted that a system like this would be most helpful to them early in their recovery, specifically in the first year

4. Discussion

Using machine learning, stress and craving were identified and differentiated from baseline conditions on the basis of three-axis accelerometry data alone, with accuracy ranging from 75–77% under binary classification conditions. The addition of heart rate and EDA data to the algorithm increased accuracy of all models. These additional features were notably most helpful in the case where accelerometry performed the poorest, which was the three-category classification attempt to simultaneously differentiate among craving, stress, and non-stress conditions.

An important consideration for future iterations of this work will be feature selection-specifically which sensor derived features are crucial to measure and include, and which can be omitted. In this work we used features easily measured with a commercially available research grade device-namely accelerometry, heart rate and EDA. Accelerometry is a particularly interesting parameter as it may capture subtleties in movement patterns that correlate with individual behavior; where heart rate and EDA can increase or decrease, accelerometry can capture both the direction (increase or decrease) and the characteristics (short vs long amplitude, axis dominance, repetitive behaviors, etc.) that are more complex. Accelerometry data, which is readily available on simple smartwatches and fitness trackers, was used alone to produce detection accuracies of 74–76% in the three binary classification cases. While adding additional parameters of heart rate, and EDA increased accuracy, they also come with increase complexity and cost of the hardware, increase processing time and memory use. This tradeoff between accuracy and complexity will need to be considered in future work on optimal feature selection.

Further work is needed to refine the algorithms and improve their accuracy. A particular scenario that may lead to misclassifications is the co-occurrence of stress and craving-either a particularly stressful event that precipitates a craving, or an unprovoked craving causing significant distress. Our data included several annotations indicating simultaneous stress and craving, however the overall number was too small to analyze this as a distinct state. Of note, other research suggests that craving and stress often occur in proximity to one another and are likely even harder to distinguish during a crisis. While conventional behavioral treatments may help clients to differentiate these feeling states, mHealth technologies might be even more effective as they offer objective information during times of heighted risk

Participants in treatment for SUD engaged in the protocol by wearing the sensor and self-reporting episodes of stress and craving. Overall, acceptability was high. All participants were informed that the sensor may also be able to detect physiologic changes associated with drug use (if it occurred during participation), which was not a deterrent to study enrollment. Participants generally reported the sensor was not bothersome and were willing to engage with a wearable-based detection system for a longer time period. Interestingly, some reported that interacting with the sensor increased their ability to recognize and positively acknowledge/respond to stress and craving, suggesting that simply the concept of continuous monitoring and feedback is therapeutic. Participants reported that, in order to adopt a sensor-based system long-term, the device needs to be streamlined, aesthetically pleasing, and multi-purpose. Specifically, the sensor needs to incorporate additional functionality to replace their current devices (i.e. watches and fitness trackers). This is critical to the design of future iterations of sensors, not only for SUD applications, but for mHealth applications in general.

The ultimate goal of this work is to use the digital biomarkers generated from this study to trigger mobile-based interventions. Such interventions could include real time prevention of return to use or even well-timed mindfulness exercises, however, that was not the goal of the present pilot study. Interestingly, despite the fact that the sensor was not yet connected to an application or intervention, participants described a sense of mindfulness that came from simply wearing a device that could potentially identify and differentiate their self-reported episodes of stress and craving. Participants uniformly regarded this as positive and helpful, which is consistent with recent literature demonstrating that mindfulness based interventions (MBIs) decrease stress and craving during recovery(Davis et al., 2018; Sancho et al., 2018), and attenuate the relationship between craving and return to drug use.(Enkema and Bowen, 2017) Some hypothesize that mindfulness practice corrects some of the neurocognitive processes that become dysregulated during the process of addiction.(Garland and Howard, 2018) In fact, regular MBIs have not only been associated with behavioral changes, but with neuroplastic changes detectable on functional neuroimaging.(Garland and Howard, 2018) This further supports the premise of using mindfulness-based interventions in response to craving and stress detection.

5. Strengths/Limitations

To our knowledge, this is the first study to deploy wearable sensors to detect and differentiate self-reported stress and craving in a natural environment. This gives our findings more generalizability to real world clinical applications than cue-lab based measurements of these phenomena. Although this study was not powered to evaluate differences across participant characteristics, such as gender and age, in this small sample, these differences may exist and will need to be further replicated and evaluated. Also, the study demographic was largely Caucasian, raising the possibility that racial differences could be missed. In addition, as co-occurring mental illness is highly prevalent in this population, future work is needed to understand its potential impact on stress and craving detection algorithms.

Perhaps most challenging and problematic is the subjective nature of stress and craving, creating a situation where the threshold for this outcome is individually defined. This protocol relied on individual participant identification of events they considered “significant” or “important” to continuously capture events in real time, as opposed to validated measurement tools that would capture measures of these constructs at isolated intervals. Several participants were nervous that they were under or over reporting and struggled with the decision whether or not to report particular events. The authors acknowledge that the threshold to report varied among individuals. This is viewed as both a strength and a limitation: a limitation in the lack of standardization, but a strength in the instantaneous, personalized, and continuous nature of events captured. Other related limitations include the unknown number of unreported stress or craving episodes, and the relatively small number of episodes per participant which may have had an impact on model development.

6. Conclusion

In a population of individuals in recovery from SUD, a wearable sensor-based detection method for stress and craving was acceptable to participants, and accurate in detection capabilities. Tri-axial accelerometry data was useful to develop models with high accuracy, and this was increased with the addition of heart rate and EDA data. Future research is needed to further refine the detection algorithm capabilities in real time, and to develop connected interventions to respond to detected stress and craving with behavioral interventions to promote sustained recovery from SUD.

Highlights.

• Wearable sensors have to potential to dramatically impact SUD evaluation and interventions

• Wearable sensors can detect self-reported episodes of stress and craving with accuracy of 75–77%

• Patients in recovery from SUD are highly accepting of continuous monitoring via wearable sensors