This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Mental Health, is properly cited. The complete bibliographic information, a link to the original publication on http://mental.jmir.org/, as well as this copyright and license information must be included.
Mobile health technology has demonstrated the ability of smartphone apps and sensors to collect data pertaining to patient activity, behavior, and cognition. It also offers the opportunity to understand how everyday passive mobile metrics such as battery life and screen time relate to mental health outcomes through continuous sensing. Impulsivity is an underlying factor in numerous physical and mental health problems. However, few studies have been designed to help us understand how mobile sensors and self-report data can improve our understanding of impulsive behavior.
The objective of this study was to explore the feasibility of using mobile sensor data to detect and monitor self-reported state impulsivity and impulsive behavior passively via a cross-platform mobile sensing application.
We enrolled 26 participants who were part of a larger study of impulsivity to take part in a real-world, continuous mobile sensing study over 21 days on both Apple operating system (iOS) and Android platforms. The mobile sensing system (mPulse) collected data from call logs, battery charging, and screen checking. To validate the model, we used mobile sensing features to predict common self-reported impulsivity traits, objective mobile behavioral and cognitive measures, and ecological momentary assessment (EMA) of state impulsivity and constructs related to impulsive behavior (ie, risk-taking, attention, and affect).
Overall, the findings suggested that passive measures of mobile phone use such as call logs, battery charging, and screen checking can predict different facets of trait and state impulsivity and impulsive behavior. For impulsivity traits, the models significantly explained variance in sensation seeking, planning, and lack of perseverance traits but failed to explain motor, urgency, lack of premeditation, and attention traits. Passive sensing features from call logs, battery charging, and screen checking were particularly useful in explaining and predicting trait-based sensation seeking. On a daily level, the model successfully predicted objective behavioral measures such as present bias in delay discounting tasks, commission and omission errors in a cognitive attention task, and total gains in a risk-taking task. Our models also predicted daily EMA questions on positivity, stress, productivity, healthiness, and emotion and affect. Perhaps most intriguingly, the model failed to predict daily EMA designed to measure previous-day impulsivity using face-valid questions.
The study demonstrated the potential for developing trait and state impulsivity phenotypes and detecting impulsive behavior from everyday mobile phone sensors. Limitations of the current research and suggestions for building more precise passive sensing models are discussed.
ClinicalTrials.gov NCT03006653; https://clinicaltrials.gov/ct2/show/NCT03006653
Mobile health (mHealth) technology has demonstrated the ability of smartphone apps and sensors to collect high-fidelity and high-frequency data pertaining to patient activity, behavior, symptoms, cognition, and context [
One construct that has not been rigorously examined is impulsivity and impulsive behavior. Impulsivity is a multidimensional construct primarily characterized by the inability to inhibit acting on short-term temptations despite long-term consequences or loss of potential gains. Consequently, it is the hallmark feature of self-regulation failures that lead to poor health decisions and outcomes, making understanding and treating impulsivity one of the most important constructs to tackle in building a culture of health [
Potential behavioral biomarkers of impulsive behavior are intuitively present in most interactions with digital technology. Mobile sensing may be especially useful for assessing impulsive behavior indicative of digital addiction, such as loss of control over mobile phone use, interference with other activities, and repeated phone checking. Objectively quantifying phone usage can further help inform the debate on the existence of digital addiction [
We developed a mobile sensing system—mPulse—to remotely monitor impulsivity on both Apple operating system (iOS) and Android platforms. Our system was designed based on data that are pervasive and available across both iOS and Android platforms and can be used to measure signals of daily activities, social interactions, and digital addiction. We selected call logs, battery charging, and screen checking as the mobile sensor data sources. We conducted a 3-week exploratory study with 26 participants as part of a larger mHealth study of impulsive behavior called the Digital Marshmallow Test (DMT) [
The DMT study by Sobolev et al [
Of the 116 participants enrolled in the DMT study, a subsample of 26 participants enrolled in this passive sensing study. The subsample included 14 females, 10 males, and 2 participants who refused to disclose, and the average age of the participants was 39.1 (SD 14.16) years. Twenty-two participants owned Apple (iOS) phones (ie, iPhones) and 4 owned Android phones. We compared the baseline subjective trait assessments of trait impulsivity and impulsive behavior between the current subsample of participants and the full sample and found no significant differences between the groups.
The DMT study included three main data sources, which we used as dependent variables in this study: (1) subjective, self-reported trait impulsivity assessments performed at baseline in the lab; (2) behavioral and cognitive active tasks performed daily on the DMT mobile app; and (3) self-reports, ecological momentary assessments (EMAs), and the Photographic Affect Meter (PAM) performed daily on the DMT mobile app.
The DMT study included the two most popular self-report generalized impulsivity trait assessments collected in a lab setting: the 15-item short form of the Barratt Impulsiveness Scale (BIS-15) and the UPPS.
The BIS-15 [
The UPPS impulsive behavior scale [
The DMT app included an adaptation of three exploratory, lab-based behavioral and cognitive measures related to impulse control to mobile devices, called “active tasks”: (1) a mobile Balloon Analogue Risk Task (mBART [
The mBART measures how individuals balance the potential for reward and loss via a simulated test where the participant can earn virtual money by pumping a balloon. It is based on the BART [
The mGNG is a measure of attention and response control. It is based on the GNG task [
The mDD task is used to measure the ability to delay immediate, smaller, and shorter monetary and time-based rewards for longer, time-lapsed, but larger rewards. It is based on DD tasks that were used in research on addiction [
The DMT app included self-reports, EMAs, and PAM.
EMAs were based on a semantic differential scale and questions consisted of two opposite feelings, thoughts, or behaviors [
Self-reported questions were also based on a semantic differential scale [
PAM was designed for momentary response where users choose an image that best represents their emotion at a given time [
We analyzed the correlations between different self-reports (BIS-15 and UPPS) and behavioral measures (BAR and GNG) in the full sample of the DMT study (N=116) because it provides better estimates than the subsample of 26 participants in this study. Overall, our results corresponded to previous research on impulsivity by demonstrating high correlations between different self-reports but low correlations between behavioral measures and self-reports [
AWARE Framework is an open-source framework used to develop an extensible and reusable platform for capturing context on mobile devices [
Our goal was to create sensing models that can effectively transform raw sensor data collected from mobile phones into measurable outcomes of clinical interest. We focused on data that are pervasively available across both iOS and Android platforms while minimizing battery consumption beyond the normal use of mobile devices and protecting user privacy. Therefore, despite the relevance of data sources such as accelerometers and location data for physical activity, mobility, and motor impulsivity, we elected not to include these data sources in the passive sensing model in this study. Eventually, three types of sensor data were identified and implemented in the mPulse system (
Conceptual framework of passive sensor data and inferred behavior.
Call logs are indicators of social interactions [
Battery logs are an indicator of daily activities [
Screen checking can serve as an indicator of digital and mobile addiction. For example, a previous study demonstrated that individuals with smartphone addiction presented with some symptoms common to substance- and addictive-related disorders such as compulsive behavior, tolerance, and withdrawal [
From the passive data, we extracted the same set of features for all sensor data, namely usage, frequency, entropy, mean, and standard deviation. This resulted in 15 passive features for the analysis:
Use duration and frequency per hour: normalized duration and frequency for each hour—that is, the summation of sensor event durations and occurrences divided by total hours of data collected from each individual, respectively. For example, screen unlocks use duration per hour (denoted as screen_Use in
Use mean and standard deviation: used to measure individual usage baselines and variances. We calculated the means and standard deviations of the event durations (unit in hours) across the study for each participant. For example, screen_Mean=0.1 means that the average screen unlock duration was 0.1×60=6 minutes.
Entropy: calculated from the possibility distribution of event occurrences over 24 hours. The intuition is that if the occurrences of the events distribute more uniformly across the day, the pattern is more random (higher entropy); otherwise, if the events occur more frequently at certain hours of the day, the pattern is more controlled (lower entropy). This was inspired by the use of the entropy feature in prior mobile sensing research to measure variability of time the participant spent at the location clusters [
Means and standard deviations across individuals for the mobile sensing features are presented in
Descriptive statistics of mobile sensor data and features.
Descriptive statistics | Battery charging, mean (SD) | Call logs, mean (SD) | Screen checking, mean (SD) |
Usage (per hour) | 0.30 (0.12) | 0.02 (0.01) | 0.17 (0.08) |
Frequency (number per hour) | 0.20 (0.15) | 0.38 (0.28) | 1.97 (1.22) |
Mean (duration per activity in hours) | 2.02 (1.33) | 0.06 (0.04) | 0.11 (0.05) |
Deviations (duration per activity in hours) | 2.81 (1.21) | 0.15 (0.18) | 0.17 (0.07) |
Entropy | 2.65 (0.24) | 2.52 (0.25) | 2.89 (0.10) |
In this section, we evaluate the value of mobile sensing in explaining and predicting trait impulsivity. We first examined the correlations between mobile sensing features and different components of trait impulsivity. Next, we compared the goodness of fit for regression models using mobile sensing features as predictors. Finally, we validated the predictive power of such models using leave-one-subject-out (LOSO) cross-validation.
We found significant correlations between passive data and five of the components of trait impulsivity: (1) motor positively correlated with the entropy features extracted from screen checking (
Correlation between the 15 features of mobile sensor data and trait impulsivity scales (15-item short form of the Barratt Impulsiveness Scale [BIS-15] and UPPS) and subscales. Ent: entropy; Freq: frequency per hour; Mean: use mean; SD: use deviations; Use: use duration per hour.
We performed a multivariate regression analysis to examine the power of extracted mobile sensing features from day-to-day phone usage to explain components of trait impulsivity. Features were standardized across samples. Given our small sample size, we first used Lasso regularization to prevent overfitting by selecting the most important features. The same penalty threshold was used across all models (α=.05). We then used a linear regression model with ordinary least squares to estimate the trait impulsivity scores from the selected features. Model performance was evaluated against adjusted
Our analysis discovered four significant models: (1) sensation seeking (
Descriptive statistics of laboratory subjective impulsivity and impulsive behavior trait, and regression analysis of mobile sensor data as predictors of impulsivity trait scales and subscales.
Scale and subscale | Descriptive statistics, mean (SD) | Regression summary | Significant features | |
|
1.77 (0.36) | None | ||
Motor | 1.74 (0.45) | Screen entropy (β=.24; |
||
Nonplanning | 1.84 (0.54) | Screen deviations (β=.33; |
||
Attention | 1.66 (0.49) | None | ||
|
2.04 (0.36) | Call entropy (β=−.21; |
||
Urgency | 2.07 (0.66) | Call entropy (β=−.39; |
||
Lack of perseverance | 1.57 (0.42) | Screen deviations (β=.22; |
||
Lack of premeditation | 1.73 (0.35) | None | ||
Sensation seeking | 2.66 (0.64) | Battery frequency (β=.27; |
LOSO cross-validation was performed to further examine the predictive power of the passive sensing features for out-of-sample data. We trained a separate linear support vector regression model for each set of passive features for 25 participants and tested it on the 1 remaining participant. We ran the same procedure 26 times to obtain predicted scores for all 26 participants. Model performance was evaluated against mean absolute error (MAE) and Pearson
Descriptive statistics on daily variables used for prediction of state impulsivity are presented in
List of features from ecological momentary assessments and active tasks.
Features | Description | Descriptive statistics, mean (SD) | |
|
|||
Focused–distracted | Present moment distracted score | AM: 3.23 (2.45); PM: 3.71 (2.74) | |
Intentional–impulsive | Present moment impulsive score | AM: 3.86 (2.75); PM: 4.47 (2.93) | |
Cautious–thrill-seeking | Present moment thrill-seeking score | AM: 3.63 (2.20); PM: 3.63 (3.68) | |
Engaged–bored | Present moment bored score | AM: 3.24 (2.11); PM: 3.23 (2.33) | |
Determined–aimless | Present moment aimless score | AM: 2.74 (2.04); PM: 3.08 (2.19) | |
|
|||
Positive–negative | Previous day negativity score | 2.59 (2.11) | |
Intentional–impulsive | Previous day impulsive score | 3.95 (2.92) | |
Productive–unproductive | Previous day unproductive score | 2.47 (1.99) | |
Relaxed–stressed | Previous day stressed score | 4.64 (2.84) | |
Healthy–unhealthy | Previous day unhealthy score | 3.92 (2.50) | |
|
|||
Positive affect | Positive affect score from PAM | 9.25 (3.50) | |
Negative affect | Negative affect score from PAM | 5.79 (3.66) | |
|
|||
Risk-taking | Average number of pumps across all trials | 3.89 (1.09) | |
Total gains | Average total gain across all trials | 10.31 (2.73) | |
|
|||
Response latency | Average response time across all trials | 423.99 ms (67.70) | |
Commission error | Proportion of “go” errors across all “go” trials | 0.02 (0.06) | |
Omission error | Proportion of “no-go” errors across all “no-go” trials | 0.02 (0.03) | |
|
|||
Present bias | Average propensity to choose immediate reward across all trials | 0.34 (0.18) |
aMeasured on a scale from 0-10, with 0=most positive and 10=most negative.
bPAM: Photographic Affect Meter.
cmBART: mobile Balloon Analogue Risk Task.
dmGNG: mobile go/no-go task.
emDD: mobile delay discounting task.
We used a generalized estimating equation (GEE) model to take into account the intraclass correlations for individual differences. We performed a multivariate regression analysis for five daily semantic differentials and positive and negative affect measures. We further performed a binary classification task by labeling samples with 1=higher than the median value and 0=lower than the median value for each daily measure. We used a logistic regression model and LOSO cross-validation. The full results are reported in
Our analysis discovered three significant models for morning and evening semantic differentials: (1) focused–distracted (AM:
Regression analysis and classification of mobile sensor data as predictors of daily ecological momentary assessment questions for semantics differentials and the Photographic Affect Meter (PAM).
Features | Generalized estimating equation regression summary (Pearson |
Classification accuracy (SD) across individuals | |
|
|||
Focused–distracted | AM: |
AM: 0.83 (0.21); PM: 0.74 (0.26) | |
Intentional–impulsive | AM: |
AM: 0.80 (0.28); PM: 0.64 (0.29) | |
Cautious–thrill-seeking | AM: |
AM: 0.86 (0.17); PM: 0.87 (0.16) | |
Engaged–bored | AM: |
AM: 0.86 (0.14); PM: 0.84 (0.18) | |
Determined–aimless | AM: |
AM: 0.94 (0.12); PM: 0.91 (0.15) | |
|
|||
Positive–negative | 0.84 (0.17) | ||
Intentional–impulsive | 0.68 (0.28) | ||
Productive–unproductive | 0.92 (0.10) | ||
Relaxed–stressed | 0.63 (0.22) | ||
Healthy–unhealthy | 0.76 (0.21) | ||
|
|||
Positive affect | 0.72 (0.15) | ||
Negative affect | 0.72 (0.15) |
We used a GEE model to take into account the intraclass correlations for individual differences. We performed an exploratory multivariate regression analysis for six features from the three behavioral and cognitive active tasks: mBART, mGNG, and mDD. We further performed a binary classification task by labeling samples with 1=higher than the median value and 0=lower than the median value for each daily measure. We used a logistic regression model and LOSO cross-validation. The full results are reported in
Our analysis discovered five significant models that varied greatly in classification accuracy: (1) total gains from mBART (
Regression analysis and classification of mobile sensor data as predictors of daily active behavioral and cognitive tasks.
Active tasks | Generalized estimating equation regression summary (Pearson |
Classification accuracy (SD) across individuals | |
|
|||
Risk-taking | 0.48 (0.23) | ||
Total gains | 0.59 (0.27) | ||
|
|||
Response latency | 0.58 (0.31) | ||
Commision error | 0.89 (0.16) | ||
Omission error | 0.87 (0.13) | ||
|
|||
Present bias | 0.84 (0.33) |
amBART: mobile Balloon Analogue Risk Task.
bmGNG: mobile go/no-go task.
cmDD: mobile delay discounting task.
This exploratory study examined the potential of detecting and monitoring state impulsivity and impulsive behavior in daily life using continuous and ubiquitous mobile sensing. We explored the predictive power of the mobile sensing system and model we developed (mPulse). We discovered relationships between passive mobile sensor data and self-reported impulsivity traits, EMA of impulsive behavior, and mobile behavioral and cognitive active tasks of risk-taking, attention, and time preference.
This is the first study to examine the relationship between passive mobile phone data, daily self-reports and self-report measures of trait impulsivity, and exploratory, objective, active mobile measures of impulsivity. Overall, our findings suggest that passive measures of mobile phone use such as call logs, battery usage, and screen on-off metrics can predict different facets of impulsivity and impulsive behavior in nonclinical samples. This study adds to the emerging literature on mobile phone phenotyping using ubiquitous sensor data as well as to the measurement of impulsive behavior in daily life [
First, we investigated the relationship between mobile sensing features and impulsivity traits on the individual level. Our regression models significantly explained variance in sensation-seeking, nonplanning, and lack of perseverance traits, but failed to explain motor, urgency, lack of premeditation, and attention traits. Passive sensing features from call logs, battery charging, and screen checking were particularly useful in explaining and predicting the sensation-seeking trait. The regression model indicated that overall battery charging frequency and screen-checking usage were significant positive predictors of sensation seeking, while call entropy was a significant negative predictor. Cross-validation further confirmed the validity of these mobile sensing features for predicting sensation seeking.
Sensation seeking in itself has multiple facets from thrill-seeking to boredom proneness to disinhibition. Therefore, due to the rewarding nature of interacting with mobile devices, one would expect to discover digital biomarkers of sensation seeking in mobile sensor data. Our results suggest that individuals high in sensation and thrill-seeking may be more prone to repeated phone checking and more intense interactions with their devices when they are using them (eg, less entropy). Previous studies have yielded mixed findings on the relationship between sensation seeking and psychopathology. For example, in a meta-analysis of the UPPS subscales, sensation seeking demonstrated the strongest associations with alcohol and substance use but an overall lower relationship with other clinical conditions than other UPPS traits [
Second, we explored the use of mobile sensing features to discover measures that assess state impulsivity and impulsive behavior in daily life. Our mobile sensing model successfully predicted objective behavioral measures, such as present bias in a delay discounting task, commission and omission errors in a cognitive attention task, and total gains in a risk-taking task. Our models also successfully predicted daily EMA questions on positivity, stress, health, and affect. Perhaps most intriguingly, our model failed to predict daily EMA questions designed to measure previous day and present moment impulsivity directly.
This finding indicates that it might be easier to predict constructs related to trait impulse control than self-reported state impulsivity itself in our sample. While studies have revealed that trait impulsivity is highly related to state impulsivity [
Passive mobile sensing can be particularly useful for detecting signs of digital addiction and problematic phone usage. Digital addiction and excessive phone usage are considered other negative consequences of impulsivity and self-regulation failures [
Our inability to predict traits such as attention and urgency, which should theoretically correlate with mobile sensing features, indicates the challenge of predicting impulsivity using the sensors chosen for the current study. Similarly, our models struggled the most with predicting the EMA question that directly asked participants to self-report the general state impulsivity in the present moment and in the previous day. We suspect this finding might be due to the multidimensional nature of impulsivity and the complex interaction between trait and state impulsivity [
One of the primary goals of this study was to design a mobile sensing system and model, supporting both iOS and Android platforms. The majority of foundational research on mobile sensing was examined on a single platform, which limits the generalizability and real-life applicability of the findings. Cross-platform research services more diverse populations and offers different opportunities for passive and active assessment. Given differences between the two operating systems, compromises are required when considering passive sensor data sources to only collect the subset of sensor data that are available on all devices. Android devices in particular offer a wider range of passive sensing modalities, such as app usage and keyboard typing, compared with iOS devices. The mobile sensing capabilities of different platforms, however, continue to evolve and new restrictions might limit future research and replicability of our findings. Passive sensing can only be useful if the environments used to collect the data do not cause the user more burden than other methods of data collection.
More comprehensive sensing suggests greater privacy concerns, as more data related to a person’s life and behavior can be quantified, transmitted, and stored. The intention of collecting passive sensing active behavioral tasks and EMA data was to build and validate digital biomarkers that can assess impulsivity for future intervention and management, and the preliminary results show the promise of such data. Yet, there exist very real possibilities for such data to be used to exploit a user, for example through stimulated impulsive purchasing [
There are several limitations to the study design that may have affected the performance of passive sensing models. One of these limitations is that the passive sensor data collection was noisy in the sense that user intentions were not fully captured by the current system. For example, it is potentially useful to distinguish screen checks in response to notifications from screen checks initiated by the users. Another limitation is that this study was based on a small sample size, as was the case with previous exploratory passive sensing studies. In addition, due to the cross-platform (iOS and Android) implementation of the mPulse system, the passive sensing and range of mobile sensing modalities were limited. Relevant data sources, such as keyboard and SMS logs, could potentially be used to examine behaviors but were not included in this study because they were only available on the Android platform. Another limitation is that our preference to protect user privacy and reduce battery drain led to the exclusion of relevant mobile sensor data sources, such as location and accelerometer data for motor impulsivity.
Future work should pursue replication of promising measures as well as explore novel sensing modalities with larger samples. Mobile sensor data sources, such as global positioning systems and accelerometers, can be explored to detect mobility and physical activity as predictors of motor impulsivity. Such future work should directly address technical limitations, including battery drain, privacy concerns with regard to location sharing, and the generalizability of mobile sensing models to both iOS and Android platforms. Similarly, physiological sensing modalities from wearable devices, such as heart rate variability, can provide multimodal sensing capabilities. These explorations can reveal more information and improve the prediction accuracy of state impulsivity and impulsive behavior.
We developed a mobile sensing system called mPulse for both iOS and Android smartphones to remotely detect and monitor state impulsivity and impulsive behavior as part of the DMT study. The design of our mPulse system was based on data that are pervasively available across both iOS and Android platforms: call logs, battery charging, and screen checking. In the exploratory study, we used mobile sensing features to predict trait-based, objective behavioral, and ecological momentary assessment (EMA) of impulsivity and related contacts (ie, risk-taking, attention, and affect).
Our findings suggest that passive sensing features of mobile phones can predict different facets of trait and state impulsivity. For trait impulsivity, the models significantly explained variance in sensation, planning, and lack of perseverance traits but failed to explain motor, urgency, lack of premeditation, and attention traits. On the daily level, the model successfully predicted objective behavioral measures such as present bias in a delay discounting task, commission and omission errors in a cognitive attention task, and total gains in a risk-taking task. Our models also successfully predicted daily EMA questions on positivity, stress, health, and affect. Overall, the study highlights the potential for continuously, passively, and remotely assessing impulsive behavior in daily life to advance the science of self-regulation and awareness.
15-item short form of the Barratt Impulsiveness Scale
Digital Marshmallow Test
ecological momentary assessment
generalized estimating equation
leave-one-subject-out
mean absolute error
mobile Balloon Analogue Risk Task
mobile delay discounting task
mobile go/no-go task
mobile health
Photographic Affect Meter
Positive and Negative Affect Schedule
HW, MS, and FM wrote the manuscript. FM, DE, and JPP designed the study. JK and HW implemented the mobile app for the study under the supervision of JPP and DE. HW, RPV, and MS conducted all statistical analyses. All authors reviewed the final manuscript.
This study was supported by a Robert Wood Johnson Pioneer Portfolio grant entitled The Digital Marshmallow Test: Multiple PI: FM and DE.
None declared.