Digital solutions for shaping mood and behavior among individuals with mood disorders

Abstract

Mood disorders present on-going challenges to the medical field, with difficulties ranging from establishing effective treatments to understanding complexities of one’s mood. One solution is the use of mobile apps and wearables for measuring physiological symptoms and real-time mood in order to shape mood and behavior. Current digital research is focused on increasing engagement in monitoring mood, uncovering mood dynamics, predicting mood, and providing digital microinterventions. This review discusses the importance and risks of user engagement, as well as barriers to improving it. Research on mood dynamics highlights the possibility to reveal data-driven computational phenotypes that could guide treatment. Mobile apps are being used to track voice patterns, GPS, and phone usage for predicting mood and treatment response. Future directions include utilizing mobile apps to deliver and evaluate microinterventions. To continue these advances, standardized reporting and study designs should be considered to improve digital solutions for mood disorders.

Introduction

Mood disorders, including bipolar disorder (BP) and major depressive disorder (MDD), challenge an individual’s ability to maintain a normal emotional state. Despite the availability of several treatments, including medication and psychotherapy [1], patients have a difficult time managing their illness. Similarly, physicians are faced with the challenge of responding to their patients’ mood on a daily basis. Managing symptoms is critical, as the consequences of mood disorders can be severe: suicidal ideation, irregular sleep, anhedonia, and hospitalization [2].

Recent research has trended towards using digital technology, such as activity trackers and mobile apps, for monitoring and managing mood disorders. These devices recover self-reported and physiological measures of mood and related concerns (e.g., heart rate and sleep), which provide insight into an individual’s current and future mood. To better treat and predict how an individual’s mood state will be impacted by the many biological and environmental factors that contribute to it, one must be able to identify all of these factors and how they are interrelated (Fig 1). The present review aims to discuss how systems biology research currently incorporates wearable and mobile data to shape mood and behavior among individuals with mood disorders.

Fig 1.

Digital solutions in mood disorders. Mobile and wearable devices interact with an individual in an effort to shape better outcomes. Data collected from several sources (e.g., GPS, voice) are used as digital markers of psychosocial targets, which could uncover illness-related dynamics and predictions.

Engaging users

Before digital data can be used to improve mental health, it must first be collected. This feat is not without complications. Low adherence is a common barrier to digital health, and mood disorders are no exception [3,4]. There is hope that individuals with mood disorders may be especially motivated given that their disorder can be debilitating and self-monitoring is clinically recommended [5]. A recent study looked at how to engage individuals with BP in digital self-monitoring of their mood. Participants were prompted via push notifications to record their symptoms twice each day in a mobile app and were asked to wear a Fitbit to track sleep, heartrate, and activity. This study found that at least 50% of self-report responses and 12 hours of activity data could be collected for about 80% of days over 6 weeks [6]. They also noted daily adherence declined over 6 weeks by 18.1% for the Fitbit but only by 6.1% for the app, even though at follow up more participants reported they would monitor symptoms with wearable device for over a year than with an app (72% vs 47%) [6]. Another study found 80% of participants had over 4 weeks of active participation in an 8 week smartphone-based intervention [7]. It will be important for digital solutions for mood disorders to keep this time frame in mind, and/or to take active steps to prolong engagement.

Passive data collection is one proposal for enhancing user engagement. Mobile apps can run in the background while collecting voice patterns (c.f., MONARCA [8] or PRIORI [9]), GPS [10], phone keystroke data [11], or more general phone metadata [12]. Wearables such as a Fitbit also collect physiological measures passively. The assumption is that passive data collection leads to better adherence than active data collection, but this is not always true. In the engagement study discussed above, for example, average percent days that a Fitbit was worn for at least half the day was lower than average percent days that at least half of in-app survey responses were provided (78% vs. 82%). [6]. Users must remember their device and not mind the invasion of privacy [13] — although these issues are addressable through reminders or advanced security.

Even if passive data collection led to better adherence, active data collection processes should not be discarded. Current “ground truth” for clinical diagnoses and mood is based on information collected actively from a patient. Data collected actively could yield information more consistent with this ground truth. Another reason that researchers may want to consider active data collection, or avoid active data collection depending on their goals, is that mood is influenced by active measurement (just like Schrodinger’s cat). That is, an individual who records their mood becomes aware of their mood, which may then alter their mood and behavior. Presumably, this observation is why monitoring symptoms is clinically recommended [14] or why raising self-awareness was endorsed as the primary reason for engagement among individuals with BP [6].

Engagement is also deeply tied to a person’s context. Participants have been shown to engage better depending on their age and gender [6] as well as time of day or location [15]. Additionally, engaging in smartphone monitoring could lead to reduced risks of mania, yet increases in risk of depression [16]. This suggests possible harmful effects of mobile self-reporting and the necessity for more studies to examine that issue. Further advances would benefit from standardized reporting to most effectively and rapidly achieve a method of monitoring mood [17].

Uncovering dynamics

Once digital data is available, mood symptoms and behavior can be examined for underlying temporal patterns or dynamics. Dynamics may reveal computational phenotypes of individuals with mood disorders, i.e. a quantitative, data-driven description of an individual. Computational phenotypes would hopefully guide treatment decisions and enrich basic research. A classic example is a rapid-cycler, an individual with BP experiencing a pronounced episode count over a year. Digital data, however, makes it possible to go beyond episode count to define computational phenotypes that account for symptoms at a finer resolution of time, type, and severity.

What resolution of time should be reflected in different phenotypes? Clinical practice would suggest weeks to reflect the duration of mood episodes, but this precedence is based on diagnostic criteria that require a week-long duration of symptoms and is likely motivated by practical reasons. Meaningful shifts in mood occur in days, or even hours, and at a level below diagnostic criteria. Indeed, current and future diagnoses of mood disorders were associated with mobile statistics of intra-day mood dynamics [18,19].

Mathematical models can help synthesize digital data onto mood dynamics [20]. This approach consists of fitting a model to data and using fitted parameters to phenotype individuals. One study, for example, fit a Bayesian hierarchical model of Markov chains to weekly patterns of mood in BP [21]. Fitted parameters revealed three latent subtypes, which differed significantly in their suicidality and disability. Classic models of dynamical diseases would rely on oscillators or multiple stable points (Fig 2), which might not be suitable for mood. An empirical study of BP, for example, did not find statistical support for mood being driven by multiple stable points or period oscillations [22]. Perhaps internal dynamics are overwhelmed by external factors. Instead, mood is popularly described as trending towards some baseline value except for random fluctuations (illustrated under Assumption 2 in Fig 2) [23–25]. Such a model captures dynamics with three parameters (baseline level, variability, and rate of return), which could be used to guide treatment.

Fig 2.

Computational modeling and phenotyping of mood dynamics. Top: Possible modeling assumptions for mood dynamics. Bottom: Computational phenotypes recovered from model parameters fit to describe longitudinal patterns of mood.

Predicting symptoms and behavior

Prediction is another target for digital solutions to mood disorders (Table 1). Predicting when depression or mania is going to happen, for example, could mean psychiatric services could be quickly shifted to the individual, perhaps even before a mood episode occurs. Given how quickly mood can shift, predicting future moods is difficult [26], but additional prediction-related goals can be approached.

Table 1.

Connection between digital measurements and psychosocial targets for mood disorders.

Source	Measurement	Psychosocial target
Sleep tracker	Sleep duration, sleep efficiency, time to fall asleep, time to get out of bed, bed or wake time	Difficulty falling asleep, hypersomnia, decreased need for sleep, sleep regularity
GPS	Time at home, places visited, distance travelled, time at work or school	Disorganized behavior, expansive mood, change in social behavior, anhedonia, depressed mood, catatonia
Voice recordings	Pitch, speed, timing, nonverbal expressions	Rapid speech, expansive mood, talkative, incoherent thought patterns, change in social behavior, depressed mood, suicidal ideation
Activity tracker	Step count, time sedentary, exercise time, circadian phase	Increased energy, fatigue, catatonia, anhedonia
Heartrate tracker	Resting heartrate, heartrate variability	Resilience, stress, circadian rhythms
Phone metadata	Number, duration, and type (incoming vs outgoing) of calls and texts	Expansive mood, change in social behavior, depressed mood
Keystrokes	Errors, reaction time	Disorganized behavior, fidgeting
Self-report	Self-reported mood, thoughts, or behavior	Any mood, thought, or behavior

One goal of digital health is to try to predict current mood, which avoids bias from reporting retrospectively or from lack of self-awareness (e.g., knowledge of one’s thoughts and emotions) and could provide missing moods from inconsistent reporting. Many studies have braved the task of predicting current mood from voice recordings and other characteristics of phone usage [16,27]. Machine learning algorithms from random forests and neural networks to more traditional regressions are applied to make predictions from digital data. In the PRIORI study, voice recordings are mined for a rich set of acoustic features of pitch, speed, timing, and expressions, among others. These features are shown to predict mood during clinical assessment phone calls and natural phone conversations from individuals with bipolar disorder – though which features are predictive depends on the type of phone call [28].

Since predictions of mood or phenotype largely serve as an intermediate step in treatment decisions, a newer endeavor is to predict treatment response or timing [28,29]. A primary goal of this work is to predict what type of interventions would be most successful for which particular mood states, both within and between participants. For example, digital data could be used to predict mood responsivity to micro-interventions (e.g., going on a walk, resting). Using random forests, it was shown that intervention success both within and between participants could be predicted, adding to the wealth of potential tools clinicians can use to prevent mood episodes [29].

Researchers also try to predict who belongs to a clinical mood phenotype [30]. One study differentiates BP from controls based on daily self-evaluations, phone usage data, and voice recordings [12]. Another differentiates individuals recently hospitalized for suicidal ideation from controls based on voice recordings [31]. Such digital predictions could support clinical diagnoses, especially when a diagnosis is inaccessible, illustrating yet another advantage of real-time predictions.

A reoccurring theme is shifting away from a single prediction model for a population to multiple prediction models, one personalized to each individual. It is thought that variability in features between individuals in mood disorders overwhelms any predictive information in these features, rendering population-level prediction models unsuccessful [32]. This variability can be reduced by normalizing features to each individual, but still using a single population prediction model based on these normalized features [27]. Alternatively, normalization is not needed if one simply builds a prediction model for each individual. Such an approach was found to greatly enhance predictions over a population-level approach. That is, when using their entire sample to train random forest prediction models, they found that passive data collected on phone usage and mobility had no added value over baseline characteristics. However, models trained to a single individual’s passive data could perform better than random chance for about 80.6% of subjects [32]. Of course, this study points out the limitation of personalized models: small sample sizes!

Newer directions examine other passive data streams. GPS data relates mood-related symptoms and behavior to movement patterns such as where they visit (e.g., work, home, coffee shop) and how far they go. For example, distance travelled was correlated with a change in mood symptoms among individuals with schizophrenia [33]. Distance from home, time spent home, and other GPS data was also related to negative symptoms and motivational deficits in schizophrenia [10]. However, limitations with using GPS data have been found. For example, Android phones appeared to have significantly less GPS coverage than iOS phones [34], which would potentially affect quality of the predictive model.

Physiological data from wearables – heart rate, sleep patterns, and body temperature – are also thought to indicate mood. Heart rate variability, for one, has predicted mental resilience [35] and depressive symptoms in healthy individuals [36]. Wearables have also been shown to more accurately predict stress levels in individuals than approaches to modify behaviors, although the study did not focus specifically on those with mood disorders [37]. Additionally, wearable data was collected over 2 years from Android light sensors and Fitbits and processed into 130 daily features related to sleep (e.g., sleep onset, sleep efficiency), circadian rhythms (e.g., parameters from cosinor analysis of heartrate data), and activity (e.g., step count). Random forests were then trained to predict mood in MDD and BP from these daily features [38].

Shaping mood & behavior

Ultimately, digital solutions to mental health promise to shape mood and behavior. Of the plethora of mental health apps available, many seek to create mobile versions of in-person psychosocial interventions, but few are supported by randomized clinical trials (RCTs) that demonstrate efficacy in mood disorders [39]. Among mobile versions of cognitive behavioral therapy (CBT) and mindfulness therapy with RCTs, a meta-analysis found they can significantly improve depression and anxiety symptoms over controls [40].

Even more compelling than stand-alone mobile therapy services is the use of therapist-supported smartphone-based interventions. One such app, Ascend, used CBT approaches and followed participants post-intervention and still observed clinically significant reductions in depression and anxiety after six and twelve months of follow-up [7]. The relevance of retaining long-lasting reductions in mood symptoms circles back to prevention – CBT has been shown to reduce the risk of relapse in depression [41].

We want to echo two concerns raised by the authors in [40]. First, it is not clear what control group should be used for comparison against a mobile intervention group when evaluating effectiveness in a RCT. Control groups in such studies currently include those on waitlists, assessment only groups, treatment as usual, or providing individuals with educational resources only. A better control group may involve the ua mobile app to better isolate the direct effect of the mobile intervention from some other mobile component. Second, the RCT should be designed to handle user disengagement to prevent non-causal biases when analyzing treatment arms. Intent-to-treat analyses, for example, require that each user has an outcome available, which may not be possible if the outcome variable needs to be collected digitally from users and many users do not engage in providing digital data. These concerns are largely avoided, however, with a new RCT-like design called micro-randomized trials. Micro-randomized trials involve repeated randomization of individuals to intervention groups throughout the study. The intervention is then assessed based on its immediate effect on outcomes (e.g., current mood). Since everyone uses the mobile device, the micro-randomized trial avoids the issue of deciding on a suitable control group. Further, repeated randomization allows these trials to determine who to deliver what intervention when, as opposed to if an intervention is effective. Thus, these trials were designed to develop personalized mobile interventions, but require additional exploration within the mental health space [42].

Conclusion

Addressing the aspects of engagement, dynamics, prediction, and apps for shaping behavior that need to be further studied and modified, primarily due to issues with sample size, methodology, and data quality, will require collaborative work across teams and fields of discipline. Tremendous progress has been made toward creating tools for patients to use supplementary to in-person care, including evidence suggesting that digital phenotyping is best when coupled with receiving clinical care [43]. Wearables and mobile device approaches holds tremendous promise for pointing researchers, clinicians, and patients toward better approaches to tackle the challenges of stabilizing a normal mood state.

Highlights.

• Wearables and mobile data can help to shape mood and behavior

• Digital engagement poses unique risks and benefits to mood disorders

• Mathematical models synthesize digital patterns of mood

• Data streams (e.g., GPS) provide a rich feature space for machine learning

• Mobile apps promise to deliver interventions for timely alleviation of symptoms