An integrated framework for using mobile sensing to understand response to mobile interventions among breast cancer patients

An integrated framework for using mobile sensing to understand response to mobile interventions among breast cancer patients

Journal Pre-proof An Integrated Framework for Using Mobile Sensing to Understand Response to Mobile Interventions among Breast Cancer Patients Lihua C...

2MB Sizes 0 Downloads 18 Views

Journal Pre-proof An Integrated Framework for Using Mobile Sensing to Understand Response to Mobile Interventions among Breast Cancer Patients Lihua Cai, Mehdi Boukhechba, Matthew S. Gerber, Laura E. Barnes, Shayna L. Showalter, Wendy F. Cohn, Philip I. Chow PII:

S2352-6483(19)30050-9

DOI:

https://doi.org/10.1016/j.smhl.2019.100086

Reference:

SMHL 100086

To appear in:

Smart Health

Please cite this article as: Cai L., Boukhechba M., Gerber M.S., Barnes L.E., Showalter S.L., Cohn W.F. & Chow P.I., An Integrated Framework for Using Mobile Sensing to Understand Response to Mobile Interventions among Breast Cancer Patients, Smart Health, https://doi.org/10.1016/j.smhl.2019.100086. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Elsevier Inc. All rights reserved.

An Integrated Framework for Using Mobile Sensing to Understand Response to Mobile Interventions among Breast Cancer Patients Lihua Caia , Mehdi Boukhechbaa , Matthew S. Gerbera , Laura E. Barnesa , Shayna L. Showalterb , Wendy F. Cohnc , Philip I. Chowd,⇤ a Department

of Engineering Systems and Environment, University of Virginia, United States b Department of Surgery, University of Virginia, United States c Department of Public Health Sciences, University of Virginia, United States d Department of Psychiatry and Neurobehavioral Sciences, University of Virginia, United States

Abstract Despite the proliferation of mobile interventions that are delivered through smartphones, and their potential impact on promoting mental health in chronic health populations, little research has examined how to leverage the advanced technological capabilities of smartphones to also monitor response to app-based interventions. Current methods used to evaluate response to digital interventions rely on static and retrospective self-report measures to infer crucial behavioral and affective patterns, which provide little contextual information and limits generalizability. Importantly, this approach can also increase burden among chronic health populations that are already experiencing a heavy treatment load. Using data collected from 7 recently diagnosed breast cancer patients undergoing active cancer treatment (average time from initial diagnosis to study enrollment is 13 days), we propose a framework for integrating self-report surveys and fine-grained mobile sensing data. Preliminary results demonstrate the feasibility and utility of this framework to reduce burden and improve detection of mood in response to an app-based intervention among chronic health populations. Keywords: Mobile sensing, breast cancer patients, mHealth, mobility, mental health.

1. Introduction In the United States, where breast cancer is the most common form of cancer and the second leading cause of cancer-related deaths in women, nearly 50 percent of the 266,120 women that were diagnosed with invasive breast cancer in 2018 will report clinically significant levels of depression or anxiety within the first year of their diagnosis [1, 2, 3, 4], which can lead to low quality of life [5], poorer compliance with cancer treatment [6], and decreased survival [7]. There is, and will continue to be, a considerable need for accessible psycho-social care that can address mental health needs and quality of life in breast cancer survivors undergoing active cancer treatment. Despite the existence of a number of efficacious psychological interventions that have been shown to reduce distress in breast cancer survivors [8], the provision of these treatments is limited by the severe shortage of trained therapists and clinicians that specialize in psychologically-based treatment [9], leading to almost half of breast cancer survivors reporting unmet supportive care needs [10]. Compounding to this overreliance on in-person delivery models for distress management, studies find that distress management is often overlooked during the cancer treatment process [11]. To overcome these barriers, health researchers are increasingly turning to digital interventions, such as mobile applications (or app-based interventions), to increase treatment accessibility [12]. For example, a meta-analysis [13] of 19 randomized controlled trials found that computer-based interventions produced large and significant effects ⇤ Corresponding

author Email addresses: [email protected] (Lihua Cai), [email protected] (Mehdi Boukhechba), [email protected] (Matthew S. Gerber), [email protected] (Laura E. Barnes), [email protected] (Shayna L. Showalter), [email protected] (Wendy F. Cohn), [email protected] (Philip I. Chow) Preprint submitted to Smart Health

July 5, 2019

on anxiety symptoms versus waitlist and placebo conditions (ds = (0.49, 1.14)), and these effects were similar to interventions that were delivered face-to-face via a trained therapist (ES = 0.60, CI = ( 0.001, 0.61)). The percentage of U.S. adults who own a smartphone has steadily increased since 2011, from 35 to 77 percent in 2018 [14]. Cancer survivors report a strong preference for receiving care through digital means, for example, a cross-sectional study (N=102) found that 41.6% of young adults with cancer rated digital communication as essential to their daily lives, and 66.3% reported a desire to receive clinical information online [15]. A qualitative study of interviews with cancer survivors (N=30) found themes of: wanting to monitor symptoms through technology, access information online, and receive personalized advice and tailored care [16]. Thus, smartphones apps are an ideal platform from which to deliver brief, empirically supported interventions to anyone that needs them. In particular, mental health apps are cheap, scalable, and portable, able to be deployed where and when they are most needed [17]. Research shows that breast cancer survivors are more than twice as likely of seeking cancer information online than other cancer survivors, and those recently diagnosed or still on active cancer treatment are the most likely to seek support online [18]. Despite the proliferation of app-based interventions, a critical limitation of prior work is a sole reliance on selfreported outcomes to gauge the effectiveness of interventions. Typically, studies evaluating app-based interventions collect self-reported outcomes only before and after receiving the intervention, through a battery of online or paperand-pencil questionnaires. Thus, data is only collected at 2 time points, which cannot inform researchers of the day-to-day impact of an intervention. Using retrospective self-report instruments to infer crucial affective behavioral patterns, such as daily mood and social isolation, is also prone to inaccurate reporting and recall bias. While feasible for some populations (e.g., students, healthy adults), repeated collection of self-report data on distress and daily behaviors can be burdensome and unrealistic for breast cancer survivors that are receiving active cancer treatment, and who must deal with the inevitable sequelae of anti-cancer care, including treatment side effects, time constraints, and conflicts with work and outside activities [19]. In addition to delivering app-based interventions, technological advances in smartphones now make it possible to continuously and unobtrusively monitor where someone is and what they are doing without needing to ask. Exploring ways to infer mental status based on passively sensed data could minimize measurement burden and enhance understanding of how breast cancer survivors and other chronic health populations respond to digital app-based interventions. The goal of the present study is to present a framework for linking self-reported mood data with objective behavioral data in breast cancer patients undergoing active cancer treatment, before and after receiving an app-based mental health intervention. To our knowledge, no published work has linked mood and passively sensed behavioral data in breast cancer patients undergoing active cancer treatment, to evaluate response to an app-based intervention. We use self-report surveys, before and after receiving an app-based intervention, as ground truth for mood and mental health status. Collecting passively sensed data before and after receiving an app-based intervention can advance understanding of how breast cancer patients respond to mobile interventions in daily life, while minimizing measurement burden, as well as inform how daily changes in mood may influence long-term distress outcomes such as depression and anxiety symptoms. 2. Related Works Health functioning and symptoms can be indirectly assessed through both subjective (e.g., self-report surveys and interviews) and objective methods (physiological variables such as heart rate). Such methods have largely been employed in clinical or laboratory settings, which are valuable, but burdensome to obtain and may have limited ecological validity. To increase generalizability of findings, researchers have tried to understand mental health through non-invasive and real-time data collected from people’s everyday lives [20, 21]. Embedded smartphone sensors (e.g., accelerometers, light sensors, GPS) are now advanced enough to allow for passive and continuous data collection, which are increasingly being used to enhance understanding of the relationship between objective behavior and health status [22, 23, 24, 25, 26]. This research is motivated by previous studies in psychology and engineering that have established a connection between data from smartphone sensors and mental health. To our knowledge, no published work has examined this issue among breast cancer patients actively receiving cancer treatment, and who are potentially unique from other populations. Our work aims to propose a framework for using fine-grained, passively collected data from smartphones before and during an app-based intervention for distressed cancer patients. The theoretical underpinning of this work 2

Figure 1: Framework to study cancer patients.

and related works is based on a layered, hierarchical model for translating sensor data into discernible markers of mental health [27]. This model postulates a connection between people’s mental status and their behaviors from raw sensor data. Mental status (e.g., distress) is believed to manifest through overt behaviors, which are collected and extracted through smartphone sensors [27]. Several studies have focused on using smartphone sensors to approximate mood in general and college student populations. In a general sample of adults, one study found that location data from GPS sensors could predict depression level with impressive accuracy [22]. Prior work in college students has found that passively sensed location information can accurately predict anxious [28, 29] and depressive symptoms [30] with college students, and researchers are increasingly exploring passively sensed indicators of stress and health behaviors in college students [23]. For example, another recent study found that individuals with higher levels of social anxiety were more likely to report negative affect during the day, which predicted spending more time at home at subsequent Ecological Momentary Assessment (EMA) prompts [31]. Multiple studies (e.g., [32, 33, 34]) have examined the relationship between mental health and patterns of mobile phone usage and communication, such as patterns of text messaging and phone calls. In general, these studies provide evidence that it is possible to approximate mental health status using smartphone features. However, none of these studies have examined fine-grained behavioral dynamics that might be linked to mental health status of cancer patients, as it pertains to response from an app-based mental health intervention, and are therefore limited in their ability to inform our understanding of how breast cancer patients might respond to a digital intervention. Finally, most studies that evaluate response to digital or app-based interventions have relied on single pre- and post-intervention assessments, typically via self-report instruments, and are therefore unable to inform researchers of the behavioral and affective patterns of individuals before and after an intervention is introduced. Importantly, the ability to obtain a more nuanced understanding of functioning before and during an intervention can enhance clinical decision making as well as targeted interventions that adjust to a user’s state. 3. Sensing Framework for Monitoring Cancer Patients In order to integrate passive sensing with active sensing to monitor cancer patients’ mental health over time, we design a framework that combines Sensus [35], a cross-platform multipurpose sensing system, with an SMS tool from 3

a survey platform (e.g., Qualtrics), as shown by Figure 1. Specifically, the Sensus mobile app is configured on a user’s personal smartphone, and is responsible for recording passive sensing data from most smartphone embedded sensors. Researchers can flexibly develop sensing plans through protocols, to control what sensors should be deployed and how data are recorded and transmitted to the cloud (e.g., AWS Simple Storage Service or S3). Active sensing through EMAs are performed through a survey platform, in this case via the SMS tool from the Qualtrics Survey Platform. EMAs can be delivered through a text message link, which will automatically connect user to a browser survey. This mechanism is robust and independent of mobile OS. Both passive and active sensing data will be transmitted to the cloud servers and made available for pre-processing and data analyses in both online and offline applications. 4. Participants and Procedures To demonstrate the use and feasibility of this framework, we designed and implemented a single-group pilot study that recruited breast cancer patients from a clinic and provided them with mental health apps. To limit barriers to entry, inclusion criteria are limited to the following: 1) breast cancer patient; 2) at least 18 years of age; 3) proficient in English at a 6th-grade level; 4) has a smartphone or is willing to carry one around if provided. Participants were not required to have a minimum level of familiarity with mobile devices or technology. Participants received a USD 50 dollar gift card for providing user feedback at the end of the study. We report on data collected from 7 breast cancer patients, diagnosed with stages 0-3 breast cancer, undergoing active cancer treatment. All participants were White, female, and the average age was 60 years (range=47-75 years). The average amount of time from primary cancer diagnosis to study enrollment was 13 days (range=6-27 days). This study was approved by the local Institutional Review Board (IRB) and all study procedures were in compliance with the Helsinki Accords. Breast cancer patients were recruited from a breast care clinic. Surgical oncologists and nurses handed out a study flyer to breast cancer patients during a normal scheduled visit. Patients had an opportunity to speak to a research staff member, who provided more details about the study. Eligible patients that expressed interest in participating were led through the consenting process by a research staff member. Critical to the focus of the current paper, which is on evaluating response to app-based interventions in chronic health populations, participants downloaded the Sensus app on their phones, which collected location and activity data throughout the study period. As seen in Figure 1, participants received a daily survey at 8pm every day for 1 week, before the app-based intervention commenced. The purpose of daily surveys is to obtain a reliable pattern of psychological functioning before the app-based intervention is introduced. A 30-minute coaching call [36] was scheduled to take place roughly a week later. Coaching calls were used to orient the patients to the intervention apps and to answer questions about their usage [36], and marked the beginning of the app-based intervention period. The app suite that was tested, IntelliCare, is publicly available and is composed of 12 treatment apps that each focuses on a different aspect of mental health (e.g., identifying negative thoughts, reducing worry). IntelliCare [37] utilizes an elemental, skills based approach to improving mental health. Similar to other empirically supported app-based interventions, the IntelliCare app content is based on evidence-based approaches in cognitive behavioral therapy (CBT), as well as concepts from mindfulness and positive psychology. Patients were instructed to try two new apps every week, and to retain the ones they found useful. During the app intervention phase of the study, participants responded to 6 weekly surveys over the 7 week intervention period. Similar in function to the daily surveys, the purpose of the weekly surveys is to understand patterns in psychological functioning during the app-based intervention, while trying to minimize user burden and reduce noise during the intervention period. Both daily and weekly surveys assessed aspects of mood, social functioning, and health related behavior. In the current study, we present data based on the items assessing mood (e.g., ”Overall, how did you feel today?” on a 5point sliding scale), anxiety (e.g., ”How much did you feel nervous, anxiety, or on edge today?” on a 5-point Likert scale), and depression (e.g., ”How much did you feel down or hopeless today?” on a 5-point Likert scale). In total, responses from 41 daily surveys over one week and 30 weekly surveys over six weeks were recorded in order to demonstrate the feasibility of the proposed framework. GPS and activity data (using Android/iOS activity recognition API) were collected during the same period.

4

Table 1: Semantic location classes. Class

Explanation

Home:

Our algorithm has been trained to recognize Home as the place having a house OSM tag (e.g., apartment, dormitory, house, etc.) where a subject spent the most time between 10 pm and 9 am; see [38] for more details about OSM tags. All houses other than Home; this includes family and friends’ places. For example, malls and grocery stores. For example, pubs, cinemas, and coffee shops. Work was defined as the place (other than home) where participants spent the most time between 9am and 5pm. Restaurants including fast food joints. For example, health clinics and hospitals. For example, sports facilities for sports and exercises. All places of worship, including churches, mosques, cathedrals, synagogues, temples, etc.

Other house: Store: Leisure: Work: Restaurant: Hospital: Gym: Religious:

5. Analyses In this section, we present some basic correlation analyses using Pearson correlations and 95 percent confidence intervals to demonstrate the feasibility of linking passively sensed behavioral features (e.g. proportion of time spent at each semantic location and activity level) to self-reported mental health metrics (e.g., mood and anxiety) among newly diagnosed breast cancer patients undergoing active cancer treatment. 5.1. Feature Extraction First, we learned the semantic locations that are visited by the breast cancer patients using the same algorithm presented in [28] with the raw GPS trajectories as input. Participants’ raw GPS data were parsed by semantic locations (eg, restaurant, campus area, and shops) by combining a spatiotemporal clustering algorithm and OpenStreetMap (OSM) geodatabase. Specifically, we first clustered participants’ GPS locations using time and space dimensions, and then, each cluster was associated with a semantic location using OSM data. After detecting participants’ visited places and labeling them using OSM, we computed the cumulative staying time in each type of location. Given a type of location and a specific participant, this metric characterized the percentage of total time that the participant spent at one type of location during a specific time window (e.g., 24 hours prior to the survey response). We use the proportion of time instead of the actual time because of missing data during continuous sensing due to various reasons (e.g., out of battery and software bugs). Semantic labels being used in the current study include Home, Store, Leisure, Work, Restaurant, Health, Gym, Other house, and Religious. The semantic data obtained from OSM is then classified into one of the following categories shown in Table 1 Sensus also collected activity types data. This data were obtained using Google Activity Recognition API and the Apple CMMotionActivity API for Android and iOS smartphones, respectively, to detect types of participants’ motions using raw sensors data (e.g. accelerometer and gyroscope). The proportion of time being in each activity level (e.g., Still, Walk, InVehicle, OnBicycle, and Running) in the designated time window (e.g., 24 hours prior to the survey response) was calculated as the activity features. 5.2. Associations between Mental Health and Passively Sensed Location Features Table 2 and 3 show the Pearson correlation coefficients between the passively sensed behavioral features and the various mental health metrics. Because the focus of this study is to establish the feasibility of our framework in a small pilot study, these findings should not be over interpreted. To increase transparency of our findings, we also report 95 percent confidence intervals in Table 2 and 3. Several findings may highlight the potential of passive behavioral markers for detecting mental health problems. For location features, there were associations between proportion of time spent at home and the three mental health metrics collected from daily surveys (rmood = 0.58, p-value = 0.001, ranxiety = 0.36, p-value = 0.036; rdepression = 0.54, p-value = 0.001, respectively), such that staying home for a longer period of time each day was associated with better mood, and lower anxiety and depression. Although this finding is seemingly inconsistent with previous findings in college student populations [31], it is important to note the dramatic differences between college students and breast cancer patients (with an average age of 60) undergoing active treatment. For example, while college students 5

Table 2: Correlation analyses between proportion of time spent at each semantic location and the various mental health metrics, including mood, anxiety, and depression collected from both daily and weekly surveys. All correlation coefficients are computed using the Pearson method. Correlations marked by a star indicating a p-value below the 0.05 threshold. The numbers contained in parentheses below the the correlations are the corresponding 95% confidence interval. Metric

Daily

Weekly

Store

Leisure

Work

Restaurant

Hospital

Gym

Family/Friend

Mood

Home

0.584* (0.26,0.79) Anxiety -0.361* (-0.62,-0.03) Depression -0.543* (-0.74,-0.25)

0.521* (0.18,0.75) -0.364* (-0.63,-0.03) -0.518* (-0.73,-0.22)

-0.276 (-0.59,0.12) -0.318 (-0.59,0.02) 0.523* (0.22,0.73)

0.145 (-0.25,0.5) -0.225 (-0.52,0.12) -0.073 (-0.4,0.27)

-0.62* (-0.81,-0.31) 0.268 (-0.08,0.56) 0.4* (0.07,0.65)

-0.798* (-0.9,-0.6) 0.672* (0.43,0.82) 0.541* (0.25,0.74)

0.374 (-0.01,0.66) -0.287 (-0.57,0.06) -0.378* (-0.64,-0.05)

0.451* (0.09,0.71) -0.375* (-0.63,-0.04) -0.451* (-0.69,-0.13)

Mood

0.538* (0.09,0.8) 0.172 (-0.28,0.56) 0.366 (-0.18,0.74)

0.186 (-0.31,0.6) 0.291 (-0.16,0.64) -0.424 (-0.77,0.11)

0.196 (-0.3,0.61) -0.015 (-0.44,0.42) -0.46 (-0.79,0.07)

-0.011 (-0.48,0.46) 0.354 (-0.09,0.68) -0.019 (-0.53,0.5)

-0.053 (-0.51,0.42) 0.407 (-0.03,0.71) -0.387 (-0.75,0.16)

0.427 (-0.05,0.75) -0.225 (-0.6,0.23) 0.2 (-0.35,0.65)

-0.464 (-0.77,0) 0.218 (-0.24,0.59) 0.202 (-0.65,0.35)

-0.01 (-0.47,0.46) Anxiety -0.577* (-0.81,-0.19) Depression 0.187 (-0.36,0.64)

are generally expected to frequently engage in social interactions that take place outside their home environment, recently diagnosed breast cancer patients may relish the opportunity to spend time with family members and friends at home, rather than receiving medical care at a hospital. These preliminary findings also suggest that more time spent in hospitals and health clinics was associated with poorer mood, as well as higher anxiety and depression levels, from the daily surveys, in breast cancer patients (rmood = 0.80, p-value < 0.001, ranxiety = 0.67, p-value < 0.001; rdepression = 0.54, p-value = 0.001, respectively), which may be attributed to the impact of cancer treatment burden on newly diagnosed breast cancer patients. Findings also suggest that spending more time at locations of friends and family members is associated with better mood, and lower levels of anxiety and depression based on the daily surveys (rmood = 0.45, p-value = 0.02, ranxiety = 0.38, p-value = 0.03; rdepression = 0.45, p-value = 0.01, respectively). Figure 2 shows the visualizations of these temporal profiles linked to the correlation analyses. We see that the pink group, which corresponds to the highest self-reported daily mood scores on the right most daily mood starplot, spent the highest proportion of time at home on average on those days. On the opposite, on the right most daily anxiety and depression starplots, the same pink group, which corresponds to the highest self-reported daily anxiety and depression levels, spent smaller proportion of time at home on average on those days. The potential of our framework is further illustrated by findings between passively sensed data and the weekly Table 3: Correlation analyses between proportion of time being at each activity level and the various mental health metrics, including mood, anxiety, and depression collected from both daily and weekly surveys. All correlation coefficients are computed using the Pearson method. Correlations marked by a star indicating a p-value below the 0.05 threshold. The numbers contained in parentheses below the the correlations are the corresponding 95% confidence interval.

Daily

Metric

Still

Walk

InVehicle

OnBicycle

Running

Mood

0.33 (-0.01,0.6) 0.005 (-0.3,0.31) -0.384* (-0.62,-0.09)

-0.347* (-0.61,-0.01) 0.16 (-0.16,0.45) 0.431* (0.14,0.65)

-0.285 (-0.57,0.06) -0.086 (-0.38,0.23) 0.284 (-0.03,0.54)

-0.057 (-0.39,0.29) 0.145 (-0.17,0.43) 0.048 (-0.26,0.35)

-0.187 (-0.49,0.16) -0.343* (-0.59,-0.04) 0.461* (0.18,0.67)

0.386 (-0.07,0.71) -0.1 (-0.51,0.35) 0.55* (0.07,0.82)

-0.414 (-0.72,0.03) -0.291 (-0.64,0.16) -0.109 (-0.57,0.41)

0.066 (-0.39,0.49) 0.39 (-0.05,0.7) -0.479 (-0.79,0.02)

-0.372 (-0.7,0.08) -0.249 (-0.61,0.21) -0.399 (-0.75,0.12)

-0.48* (-0.76,-0.05) -0.149 (-0.55,0.3) 0.616* (-0.85,-0.17)

Anxiety Depression Mood

Weekly

Anxiety Depression

6

Figure 2: Temporal profiles of semantic locations on various self-reported mental health metrics. The proportion of time of the location features is computed based on a 24-hour window prior to the daily survey. The missing sub-figure on the daily anxiety is due to absence of self-reported daily anxiety at the ”Extremely” level.

surveys, which were completed during the app-based intervention period. For example, the associations between the proportion of time spent at home and the three mental health metrics collected from weekly surveys appears to be less consistent than at the level of the daily surveys, which may be attributed to the the app-based intervention that teaches coping and mental health skills. For example, it may be that breast cancer patients were more able to regulate their mood and cope with distress, leading to a lower dependence on being at home in order to maintain their mood. This is consistent with finding that the associations between proportion of time spent at hospital and clinic settings appears to be consistently weaker among weekly mood, anxiety, and depression (during the app-based intervention), relative to the same metrics assessed from the daily surveys (before the intervention). However, it is important to stress that firm conclusions should not be drawn from this feasibility pilot study. For example, due to a small sample size of the weekly surveys, from Figure 3, we see that occasions represented by the first two weekly mood startplots, which correspond to the lowest self-reported weekly mood score levels, indicate either spending all time at home or at other house in those weeks, thereby necessitating a larger sample that utilizes this framework in future work. 5.3. Associations between Mental Health and Passively Sensed Activity Features For activity features, overall we were able to demonstrate the feasibility of linking surveys with fine-grained behavioral data from smartphones, thereby providing support for our framework. Once again, although there were several strong correlations, we refrain from drawing any firm conclusions due to the nature of this pilot feasibility study. Several findings may highlight the potential of passive activity markers for detecting mental health status. For example, from the daily surveys, more time being still was associated with a lower level of depression (rdepression = 0.384, p-value = 0.002), whereas that association is reversed while receiving the app-based intervention from the weekly surveys (rdepression = 0.55, p-value = 0.002), potentially due to acquiring psychological skills that promote physical fitness. It should be noted that activity features are low level mobility modes, and may not reflect the more important high level activities that are directly linked to the breast cancer patients’ mental state, which may be influenced by their unique circumstances. Researchers are encouraged, therefore, to based their hypotheses and findings of activity features with a strong theoretical underpinning. Figure 4 visualizes some of these inconsistencies in this particular set of findings. For example, while one might expect to find that more walking is generally associated with better mood and less anxiety and depression symptoms, this may not be generalizable to breast cancer patients since more walking 7

Figure 3: Temporal profiles of semantic locations on various self-reported mental health metrics. The proportion of time of the location features is computed based on the past week prior to the weekly survey. The missing sub-figures on both the weekly anxiety and depression are due to absence of self-reported weekly anxiety and depression at the ”Extremely” level.

could indicate more trips to see health care specialists. More studies are needed and with larger samples, in order to further contextualize these findings among chronic health populations. 6. Discussion Mental health and support apps have the potential to address a critical healthcare gap among breast cancer patients and other chronic health populations, though few studies have evaluated the potential utility of mobile sensing to directly evaluate the impact of mental health apps before and during an intervention. The proposed framework can help to address this weakness of evaluating app-based interventions in chronic health populations, by leveraging builtin smartphone sensors. In this paper, we demonstrate the feasibility of linking self-reported mental health data with passively sensed behavioral data, to ultimately inform researchers and clinicians of how cancer patients are responding to a mobile app-based intervention. Although we report on data collected from seven newly diagnosed breast cancer patients, overall, our results support the feasibility of a framework that uses passively sensed location and physical activity data to approximate mood response to a digital intervention. Specifically, findings suggest that behavioral metrics can be correlated with self-reported mood data, which served as ground truth in this study. Importantly, this framework can be applied broadly to other chronic health populations; for example, to monitor distress in diabetes and other forms of cancer. By reducing measurement burden, a major contribution of the proposed framework is to enable researchers and clinicians to better understand how breast cancer patients respond to a behavioral intervention by integrating objective, passively sensed data streams. Relative to student and healthy adult populations, newly diagnosed cancer patients undergoing active cancer treatment experience significantly greater levels of distress [1, 2, 3, 4]. In addition to enduring an intense treatment load, patients must endure the impact of receiving a cancer diagnosis on their psychological and emotional functioning, as well as their daily lives (e.g., disruption in occupation, finances, family role). Thus, recruitment of chronic health populations that are receiving active medical treatment can be slow and costly. In the current study, working closely with a large team of clinicians, nurses, and research assistants, we were able to recruit 2-4 breast cancer patients per month on average, using an intensive clinic-based recruitment strategy. Therefore, it is critical to minimize all forms of measurement burden for cancer and other chronic health populations, to promote a high level of engagement and to avoid undesirable consequences (e.g., burnout, study dropout). 8

Figure 4: Temporal profiles of activity levels on various self-reported mental health metrics. The ”Still” level is left out in these figures, and the proportions of time of all activity levels add up to 1. The proportion of time of the activity features is computed based on a 24-hour window prior to the daily survey, while on the past week prior to the weekly survey.

6.1. Limitations and Future Directions There are certain limitations to the present work that could inform future studies. As noted throughout, our focus was on collecting data that would support the feasibility of the proposed framework. Thus, the current sample size is not sufficiently large enough to draw solid conclusions about the relationship between self-reported mood variables and passively sensed data, which is why we not only report significance values but also 95 percent confidence intervals to increase the transparency of the reported findings. We caution readers against over interpreting these findings until they can be replicated in a larger sample. We also had some missing data that could be the result of breast cancer patients turning off their phone or Sensus app, or due to unknown factors that disrupt data collection (e.g., errors during data transmission, dead battery). To minimize study burden on our participants, our study team refrained from contacting participants with technical issues. However, future studies should evaluate whether it is acceptable to contact their participants regarding technical issues. Because our participants had been diagnosed with breast cancer less than two weeks before study enrollment on average, our decision was informed by the current load and circumstances of the individuals in the study. Future iterations of this work should integrate continuous data collection with real-time analysis, in order to inform clinicians how best to adjust interventions to maximally benefit their patients. 7. Conclusions Our findings suggest that it is possible to integrate self-reported survey data with fine-grained sensor data to approximate mood response to a digital intervention among breast cancer patients and potentially other chronic health populations. The proposed framework is the first of its kind among breast cancer patients receiving active cancer treatment, and will ultimately minimize user burden while improving detection of mood symptoms in daily life. References [1] C. Burgess, V. Cornelius, S. Love, J. Graham, M. Richards, A. Ramirez, Depression and anxiety in women with early breast cancer: five year observational cohort study, Bmj 330 (7493) (2005) 702. [2] J. Zabora, K. BrintzenhofeSzoc, B. Curbow, C. Hooker, S. Piantadosi, The prevalence of psychological distress by cancer site, Psychooncology 10 (1) (2001) 19–28.

9

[3] B. Grabsch, D. M. Clarke, A. Love, D. P. McKENZIE, R. D. Snyder, S. Bloch, G. Smith, D. W. Kissane, Psychological morbidity and quality of life in women with advanced breast cancer: a cross-sectional survey, Palliative & supportive care 4 (1) (2006) 47–56. [4] I. Henselmans, V. S. Helgeson, H. Seltman, J. de Vries, R. Sanderman, A. V. Ranchor, Identification and prediction of distress trajectories in the first year after a breast cancer diagnosis., Health Psychology 29 (2) (2010) 160. [5] M. Reich, A. Lesur, C. Perdrizet-Chevallier, Depression, quality of life and breast cancer: a review of the literature, Breast cancer research and treatment 110 (1) (2008) 9–17. [6] B. F. d. Souza, J. A. d. Moraes, A. Inocenti, M. A. d. Santos, A. E. B. d. C. Silva, A. I. Miasso, Women with breast cancer taking chemotherapy: depression symptoms and treatment adherence, Revista latino-americana de enfermagem 22 (5) (2014) 866–873. [7] A. A. Onitilo, P. J. Nietert, L. E. Egede, Effect of depression on all-cause mortality in adults with cancer and differential effects by cancer site, General hospital psychiatry 28 (5) (2006) 396–402. [8] K. Tatrow, G. H. Montgomery, Cognitive behavioral therapy techniques for distress and pain in breast cancer patients: a meta-analysis, Journal of behavioral medicine 29 (1) (2006) 17–27. [9] A. S. Davis, D. E. McIntosh, L. Phelps, T. J. Kehle, Addressing the shortage of school psychologists: A summative overview, Psychology in the Schools 41 (4) (2004) 489–495. [10] S. Aranda, P. Schofield, L. Weih, P. Yates, D. Milne, R. Faulkner, N. Voudouris, Mapping the quality of life and unmet needs of urban women with metastatic breast cancer, European journal of cancer care 14 (3) (2005) 211–222. [11] B. D. Bultz, L. E. Carlson, Emotional distress: the sixth vital signfuture directions in cancer care, Psycho-Oncology 15 (2) (2006) 93–95. [12] D. D. Luxton, R. A. McCann, N. E. Bush, M. C. Mishkind, G. M. Reger, mhealth for mental health: Integrating smartphone technology in behavioral healthcare., Professional Psychology: Research and Practice 42 (6) (2011) 505. [13] M. A. Reger, G. A. Gahm, A meta-analysis of the effects of internet- and computer-based cognitive-behavioral treatments for anxiety, Journal of Clinical Psychology 65 (1) 53–75. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/jclp.20536, doi:10.1002/jclp.20536. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/jclp.20536 [14] Demographics of mobile device ownership and adoption in the united states (Feb 2018). URL http://www.pewinternet.org/fact-sheet/mobile/ [15] E. Abrol, M. Groszmann, A. Pitman, R. Hough, R. M. Taylor, G. Aref-Adib, Exploring the digital technology preferences of teenagers and young adults (tya) with cancer and survivors: a cross-sectional service evaluation questionnaire, Journal of Cancer Survivorship 11 (6) (2017) 670–682. [16] S. Lubberding, C. F. van Uden-Kraan, E. A. Te Velde, P. Cuijpers, C. R. Leemans, I. M. Verdonck-de Leeuw, Improving access to supportive cancer care through an e h ealth application: a qualitative needs assessment among cancer survivors, Journal of clinical nursing 24 (9-10) (2015) 1367–1379. [17] T. Donker, K. Petrie, J. Proudfoot, J. Clarke, M.-R. Birch, H. Christensen, Smartphones for smarter delivery of mental health programs: a systematic review, Journal of medical Internet research 15 (11). [18] Y. Jiang, B. T. West, D. L. Barton, M. R. Harris, Acceptance and use of ehealth/mhealth applications for self-management among cancer survivors, Studies in health technology and informatics 245 (2017) 131. [19] W. Sun, K. Chen, A. Terhaar, D. A. Wiegmann, S. M. Heidrich, A. J. Tevaarwerk, M. E. Sesto, Work-related barriers, facilitators, and strategies of breast cancer survivors working during curative treatment, Work 55 (4) (2016) 783–795. [20] M. Boukhechba, J. Gong, K. Kowsari, M. K. Ameko, K. Fua, P. I. Chow, Y. Huang, B. A. Teachman, L. E. Barnes, Physiological changes over the course of cognitive bias modification for social anxiety, in: 2018 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI), IEEE, 2018. doi:10.1109/bhi.2018.8333458. [21] M. K. Ameko, L. Cai, M. Boukhechba, A. Daros, P. I. Chow, B. A. Teachman, M. S. Gerber, L. E. Barnes, Cluster-based approach to improve affect recognition from passively sensed data, in: 2018 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI), IEEE, 2018. doi:10.1109/bhi.2018.8333461. [22] S. Saeb, M. Zhang, C. J. Karr, S. M. Schueller, M. E. Corden, K. P. Kording, D. C. Mohr, Mobile phone sensor correlates of depressive symptom severity in daily-life behavior: An exploratory study, Journal of medical Internet research 17 (7). [23] R. Wang, F. Chen, Z. Chen, T. Li, G. Harari, S. Tignor, X. Zhou, D. Ben-Zeev, A. T. Campbell, Studentlife: assessing mental health, academic performance and behavioral trends of college students using smartphones, in: Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2014, pp. 3–14. [24] L. Canzian, M. Musolesi, Trajectories of depression: unobtrusive monitoring of depressive states by means of smartphone mobility traces analysis, in: Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2015, pp. 1293– 1304. [25] M. N. Burns, M. Begale, J. Duffecy, D. Gergle, C. J. Karr, E. Giangrande, D. C. Mohr, Harnessing context sensing to develop a mobile intervention for depression, Journal of medical Internet research 13 (3). [26] A. Gruenerbl, V. Osmani, G. Bahle, J. C. Carrasco, S. Oehler, O. Mayora, C. Haring, P. Lukowicz, Using smart phone mobility traces for the diagnosis of depressive and manic episodes in bipolar patients, in: Proceedings of the 5th Augmented Human International Conference, ACM, 2014, p. 38. [27] D. C. Mohr, M. Zhang, S. M. Schueller, Personal sensing: understanding mental health using ubiquitous sensors and machine learning, Annual review of clinical psychology 13 (2017) 23–47. [28] M. Boukhechba, P. Chow, K. Fua, B. A. Teachman, L. E. Barnes, Predicting social anxiety from global positioning system traces of college students: Feasibility study, JMIR mental health 5 (3). [29] M. Boukhechba, A. R. Daros, K. Fua, P. I. Chow, B. A. Teachman, L. E. Barnes, Demonicsalmon: Monitoring mental health and social interactions of college students using smartphones, Smart Health 9 (2018) 192–203. [30] Y. Huang, H. Xiong, K. Leach, Y. Zhang, P. Chow, K. Fua, B. A. Teachman, L. E. Barnes, Assessing social anxiety using gps trajectories and point-of-interest data, in: Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2016, pp. 898–903. [31] P. I. Chow, K. Fua, Y. Huang, W. Bonelli, H. Xiong, L. E. Barnes, B. A. Teachman, Using mobile sensing to test clinical models of depression,

10

social anxiety, state affect, and social isolation among college students, Journal of medical Internet research 19 (3). [32] A. Enez Darcin, S. Kose, C. O. Noyan, S. Nurmedov, O. Yılmaz, N. Dilbaz, Smartphone addiction and its relationship with social anxiety and loneliness, Behaviour & Information Technology 35 (7) (2016) 520–525. [33] Y. Gao, A. Li, T. Zhu, X. Liu, X. Liu, How smartphone usage correlates with social anxiety and loneliness, PeerJ 4 (2016) e2197. [34] M. Boukhechba, Y. Huang, P. Chow, K. Fua, B. A. Teachman, L. E. Barnes, , in: Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM International Symposium on Wearable Computers on - UbiComp '17, ACM Press, 2017. doi:10.1145/3123024.3125607. URL [35] H. Xiong, Y. Huang, L. E. Barnes, M. S. Gerber, Sensus: a cross-platform, general-purpose system for mobile crowdsensing in human-subject studies, in: Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2016, pp. 415–426. [36] D. C. Mohr, P. Cuijpers, K. Lehman, Supportive accountability: a model for providing human support to enhance adherence to ehealth interventions, Journal of medical Internet research 13 (1). [37] D. C. Mohr, K. N. Tomasino, E. G. Lattie, H. L. Palac, M. J. Kwasny, K. Weingardt, C. J. Karr, S. M. Kaiser, R. C. Rossom, L. R. Bardsley, et al., Intellicare: an eclectic, skills-based app suite for the treatment of depression and anxiety, Journal of medical Internet research 19 (1). [38] C. Keler, OpenStreetMap, in: S. Shekhar, H. Xiong, X. Zhou (Eds.), Encyclopedia of GIS, Springer International Publishing, Cham, 2015, pp. 1–5, dOI: 10.1007/978-3-319-23519-6 1654-1.

11