Substance use disorder (SUD) is a major public health concern worldwide. The harmful use of alcohol is responsible for 5.1% of the global burden of disease (7.1% for males and 2.2% for females) [1]. The prevalence of heavy episodic drinking, defined as ≥60 g of pure alcohol on at least one occasion at least once per month, was 18.2% in 2016 among the total population [2]. Alcohol use disorders are more prevalent in high-income countries [2]. About 5.5% of the global population aged 15-64 years had used psychoactive drugs in the previous year, and about 35 million people are estimated to be affected by drug use disorders [3]. However, treatment access has been limited because of the increasing costs of the health care system, patient time demands, and concerns about stigma [4-7]. Additionally, alcohol and other substance use increased during the COVID-19 pandemic [8-10]. The need for accessible and effective health care services has increased.

Advances in technology have the potential to reduce barriers to treatment services in this field. Broadly known as digital health or mobile health (mHealth), research in these areas uses mobile and computer software apps and wearable biosensor devices to understand and treat health conditions better [11-15]. In the field of SUD, various mobile apps have been developed to provide information on substance use and preventive health care services to continuously track specific health data such as substance use or support chronic condition management. Additionally, mHealth programs can help bridge treatment gaps by allowing patients to communicate with their physician or health care team without meeting face-to-face. One of the typical approaches of mHealth in this area is the ecological momentary assessment (EMA) method. The EMA approach has been developed to explore real-time substance use and motives, cravings, or triggers of substance use in daily life [12,16-19]. EMA has more benefits in capturing data on situations of substance use in the real world than cross-sectional or laboratory setting assessment because substance use is an episodic behavior influenced by immediate environmental factors (eg, external triggers) and internal factors such as stress and craving, which can be difficult to practically and ethically duplicate in an experiment environment [15-17]. Moreover, self-reported substance use can be biased by retrospective recall [20]. In a typical EMA, signaled prompts are sent to participants on a portable device (eg, smartphone) several times per day. The timing of the prompts depends on each study. There are generally the following 3 patterns: random time, fixed time, and event contingent. When participants receive these prompts, they are required to record their current thoughts, behaviors, and feelings on their devices.

Although EMA has advantages in assessing substance use accurately and repeatedly, compliance and missing data are still challenges [18,21]. Participants are likely to be unwilling to complete assessments during work, while studying, or when spending time with friends [22]. In people who use substances, noncompliance may occur because of an unstable life, comorbid psychopathology, and social pathology [17,23]. Missing data, especially systematically missing data (not at random), can lead to low statistical power and biased findings [18,24]. Uncomplicated EMA methods to minimize a participant’s engagement with a collection tool are needed to increase compliance with the EMA protocol. Moreover, it is difficult to confirm the validity of the data collected by self-reported assessment [25], even when using EMA methods, because social desirability bias influences subjective judgment when people report situations regarding substance use, especially illicit drug use [26,27].

One of the solutions to the challenges of EMA might be the use of a wearable activity tracker (eg, Fitbit and Apple Watch), which can automatically collect participants’ data. Although these products are not specifically made for the data collection on substance use, researchers can repeatedly gather various objective data, such as behavioral data (eg, steps and sleep) and biological data (eg, heart rate, blood pressure, and degree of stress). These devices have been used in observational studies to collect accurate real-time data [28-30], as well as in intervention studies to promote healthy behavior and improve health-related outcomes [31-33]. Also, numerous data collected repeatedly over minutes, hours, or days for a certain period enable the prediction of health outcomes using machine learning methods. For example, acute exacerbation of chronic obstructive pulmonary disease [34], depression [35], movement types [36], sleep stage [37], and COVID-19 symptom exacerbation [38] were detected by analyzing data collected using wearable devices.

If a wearable activity tracker and a self-monitoring tool that records variables related to substance use are combined in a study, researchers can collect behavioral and biological data with minimal reliance on participant input. Moreover, researchers can analyze the data collected by a wearable activity tracker in many ways because these data could be potential predictive factors of substance use or subsequent outcomes following substance use. Daily self-monitored behavioral data along with momentary biological data collected by wearable activity trackers have the potential to bridge important data gaps.

This paper aims to introduce a procedure for implementing wearable tracker technologies as a method of EMA. We describe the methods to analyze data on substance use and psychological data using a self-monitoring app and objective behavioral and biological data using a wearable activity tracker (Fitbit). We will develop a predictive model of substance use by analyzing combined data using machine learning methods.

This study is an ongoing observational study using a Fitbit and a self-monitoring app (Figure 1). Participants were required to wear the Fitbit for 8 weeks as well as open the Fitbit app every day in order to sync the Fitbit data with their smartphones. Participants were also required to record daily entries in a self-monitoring app for 8 weeks and answer an in-app survey at the time of participation, 1-month, and 2-month follow-up. After the study was finished, participants received the Fitbit along with a gift card as compensation for participating in the study.

Participants were outpatients with alcohol or methamphetamine use disorders from 2 psychiatric hospitals in Japan. A psychiatrist specializing in SUD screened potential participants based on inclusion and exclusion criteria, and patients were referred to the study if applicable. General residents with alcohol problems were recruited through a research company. Considering realistic circumstances at each facility and the approximate number of participants required to adequately perform preliminary model testing using machine learning, 20 people with methamphetamine use problems and 40 people with alcohol problems were recruited for this feasibility study. Inclusion criteria were as follows: (1) alcohol consumption with health risks in the past year (Alcohol Use Disorders Identification Test [AUDIT]: 8-19 points) [44,45], or methamphetamine use with health risks in the past year (Drug Abuse Screening Test [DAST]-10: 1-8 points) [46,47]; (2) owning a smartphone (iOS14 or Android 7.0 or later); (3) being able to wear the Fitbit constantly for approximately 8 weeks and record the self-monitoring app daily; and (4) being 20 years of age or older. Exclusion criteria were as follows: (1) not being able to speak or write in Japanese; (2) not being able to access the internet or complete the web-based survey; (3) not being able to participate for the entire 8-week research period, and (4) attending an outpatient clinic and having been deemed unsuitable to participate in the study by their attending physician.

Fitbit’s application programming interfaces capture data for heart rate per minute, sleep duration per day, sleep stages per day, the number of steps per day, and the amount of activity per day (hereafter referred to as “Fitbit data”). We finished collecting the data in April 2021 and are preparing to conduct data analysis. In order to determine substance use based on the Fitbit data, 3 steps were applied, as shown in Figure 3, and are described in further detail as follows:

First, the Fitbit data will be visualized to confirm typical Fitbit data patterns for individual users. There are many patterns to detect drug or alcohol use; for example, the heart rate may remain high while drinking alcohol. Accordingly, data visualization is beneficial for confirming such patterns. If, for example, we can detect a typical pattern as mentioned above, signals other than that pattern may be considered noise for this analysis. Subsequently, noise removal methods will be applied to the Fitbit data. For example, average heart rate or heart rate variability may also be influenced by exercise or walking rather than drug or alcohol use. Therefore, we will use step counts or activity amounts on Fitbit data to specifically filter out such movement or workout periods. Other possible methods for noise removal, in terms of signal processing, may include signal smoothing methods, for example, moving average and Gaussian filter. Methods will be chosen to retain typical patterns but remove the other signals.

Second, the Fitbit data will be analyzed by using statistical analysis methods in order to predict drug or alcohol use on a relatively long-term basis, such as within a day or a week. By using self-monitoring data, we can separate the Fitbit data into data reflecting substance use and data from normal activity. Some descriptive statistical methods, for example, averaging, variance, self-correlation, or frequency-domain analysis, will be applied to each parameter in the separated data sets to determine features that include the days a user has declared drug or alcohol use. In this analysis, survey data containing the participants’ relation to drug or alcohol use, along with psychological and behavioral data points, may also be added as complements through the descriptive statistics. After transforming the data into its statistics, we would like to apply clustering methods to provide a quantitative evaluation of typical patterns reflecting drug or alcohol use.

There are 2 approaches to determining an optimal detection method for drug or alcohol use. If we can successfully obtain distinctive features by using the above statistical analysis, we will accordingly construct a detection model based on it. However, constructing such a model may be insufficient to isolate use with high accuracy and precision. Therefore, data analysis using both machine learning and statistical analysis methods will be performed in order to create a model that solves the problem as indicated above, based on the combined Fitbit and self-monitoring data. The primary candidate for machine learning methods will be Long Short-Term Memory. Long Short-Term Memory is considered one of the most suitable methods for analyzing time series data, which applies to this study. Meanwhile, other deep learning methods will also be considered to fit the data. The deep learning model will be carefully selected after we perform the statistical analyses, and then this model will be tested based on 5-fold cross-validation. Partitioning data depends on the degree of temporal precision with which drugs and alcohol consumption is estimated. Further preprocessing and machine learning methods will be conducted based on the preliminary results of the validation.

This study was approved by the Ethics Committee of the Faculty of Medicine and Graduate School of Medicine of Tokyo Medical and Dental University (M2020-189) and the Institutional Review Boards of each recruiting hospital. Data collected from the self-monitoring app were automatically stored in a protected database. Fitbit data were archived from the Fitbit server and preprocessed on a cloud-based, protected database without personal information attached. These data were managed by Humanome Lab, Inc. Both self-monitoring data and Fitbit data were then matched using participants’ ID numbers. Patients or participants provided their written informed consent in person before participating in this study.