End-to-end deep learning models are increasingly applied to safety-critical human activity recognition (HAR) applications, e.g., healthcare monitoring and smart home control, to reduce developer burden and increase the performance and robustness of prediction models. However, integrating HAR models in safety-critical applications requires trust, and recent approaches have aimed to balance the performance of deep learning models with explainable decision-making for complex activity recognition. Prior works have exploited the compositionality of complex HAR (i.e., higher-level activities composed of lower-level activities) to form models with symbolic interfaces, such as concept-bottleneck architectures, that facilitate inherently interpretable models. However, feature engineering for symbolic concepts-as well as the relationship between the concepts-requires precise annotation of lower-level activities by domain experts, usually with fixed time windows, all of which induce a heavy and error-prone workload on the domain expert. In this paper, we introduce X-CHAR , an eXplainable Complex Human Activity Recognition model that doesn't require precise annotation of low-level activities, offers explanations in the form of human-understandable, high-level concepts, while maintaining the robust performance of end-to-end deep learning models for time series data. X-CHAR learns to model complex activity recognition in the form of a sequence of concepts. For each classification, X-CHAR outputs a sequence of concepts and a counterfactual example as the explanation. We show that the sequence information of the concepts can be modeled using Connectionist Temporal Classification (CTC) loss without having accurate start and end times of low-level annotations in the training dataset-significantly reducing developer burden. We evaluate our model on several complex activity datasets and demonstrate that our model offers explanations without compromising the prediction accuracy in comparison to baseline models. Finally, we conducted a mechanical Turk study to show that the explanations provided by our model are more understandable than the explanations from existing methods for complex activity recognition.
Activity-oriented cameras are increasingly being used to provide visual confirmation of specific hand-related activities in real-world settings. However, recent studies have shown that bystander privacy concerns limit participant willingness to wear a camera. Researchers have investigated different image obfuscation methods as an approach to enhance bystander privacy; however, these methods may have varying effects on the visual confirmation utility of the image, which we define as the ability of a human viewer to interpret the activity of the wearer in the image. Visual confirmation utility is needed to annotate and validate hand-related activities for several behavioral-based applications, particularly in cases where a human in the loop method is needed to label (e.g., annotating gestures that cannot be automatically detected yet). We propose a new type of obfuscation, activity-oriented partial obfuscation, as a methodological contribution to researchers interested in obtaining visual confirmation of hand-related activities in the wild. We tested the effects of this approach by collecting ten diverse and realistic video scenarios that involved the wearer performing hand-related activities while bystanders performed activities that could be of concern if recorded. Then we conducted an online experiment with 367 participants to evaluate the effect of varying degrees of obfuscation on bystander privacy and visual confirmation utility. Our results show that activity-oriented partial obfuscation (1) maintains visual confirmation of the wearer's hand-related activity, especially when an object is present in the hand, and even when extreme filters are applied, while (2) significantly reducing bystander concerns and enhancing bystander privacy. Informed by our analysis, we further discuss the impact of the filter method used in activity-oriented partial obfuscation on bystander privacy and concerns.
Many breast cancer survivors are prescribed daily oral medications called endocrine therapy that prevent cancer recurrence. Despite its clinical importance, maintaining consistent daily adherence remains challenging due to the dynamic and interrelated influences of behavioral, physiological, and psychological factors. While prior studies have explored adherence prediction using mobile sensing, they often rely on single-modality data, limited temporal granularity, or aggregate-level modeling-limiting their ability to capture short and long-term behavioral variability and to facilitate deeper understanding of non-adherence and tailored interventions. To address these gaps, we propose a multimodal sensing framework that explicitly models daily adherence dynamics using temporally adaptive inputs. We recruited a sample of breast cancer survivors (N = 20) and collected longitudinal data streams including wearable-derived physiological features (Fitbit), medication event monitoring system (MEMS) data, and ecological momentary assessments (EMAs). Using multimodal data across varying time windows, we examined whether recent patterns in behavioral, physiological, psychological, and environmental factors improve the prediction of next-day endocrine therapy adherence. Our results demonstrate the feasibility of using multimodal sensing data to predict daily adherence with moderate accuracy. Moreover, models integrating multimodal data consistently outperformed those relying on a single modality. Importantly, we observed that the predictive value of each modality varied depending on the temporal proximity of the input signals, underscoring the importance of modeling immediate and longer-term behavioral patterns. The findings offer valuable insights for advancing adherence monitoring systems, suggesting that incorporating personalized and temporally adaptive data fusion strategies may significantly enhance the effectiveness of intervention design and delivery.
Recent advancements in sensing techniques for mHealth applications have led to successful development and deployments of several mHealth intervention designs, including Just-In-Time Adaptive Interventions (JITAI). JITAIs show great potential because they aim to provide the right type and amount of support, at the right time. Timing the delivery of a JITAI such as the user is receptive and available to engage with the intervention is crucial for a JITAI to succeed. Although previous research has extensively explored the role of context in users' responsiveness towards generic phone notifications, it has not been thoroughly explored for actual mHealth interventions. In this work, we explore the factors affecting users' receptivity towards JITAIs. To this end, we conducted a study with 189 participants, over a period of 6 weeks, where participants received interventions to improve their physical activity levels. The interventions were delivered by a chatbot-based digital coach - Ally - which was available on Android and iOS platforms. We define several metrics to gauge receptivity towards the interventions, and found that (1) several participant-specific characteristics (age, personality, and device type) show significant associations with the overall participant receptivity over the course of the study, and that (2) several contextual factors (day/time, phone battery, phone interaction, physical activity, and location), show significant associations with the participant receptivity, in-the-moment. Further, we explore the relationship between the effectiveness of the intervention and receptivity towards those interventions; based on our analyses, we speculate that being receptive to interventions helped participants achieve physical activity goals, which in turn motivated participants to be more receptive to future interventions. Finally, we build machine-learning models to detect receptivity, with up to a 77% increase in F1 score over a biased random classifier.
Smoking is the leading cause of preventable death worldwide. Cigarette smoke includes thousands of chemicals that are harmful and cause tobacco-related diseases. To date, the causality between human exposure to specific compounds and the harmful effects is unknown. A first step in closing the gap in knowledge has been measuring smoking topography, or how the smoker smokes the cigarette (puffs, puff volume, and duration). However, current gold-standard approaches to smoking topography involve expensive, bulky, and obtrusive sensor devices, creating unnatural smoking behavior and preventing their potential for real-time interventions in the wild. Although motion-based wearable sensors and their corresponding machine-learned models have shown promise in unobtrusively tracking smoking gestures, they are notorious for confounding smoking with other similar hand-to-mouth gestures such as eating and drinking. In this paper, we present SmokeMon, a chest-worn thermal-sensing wearable system that can capture spatial, temporal, and thermal information around the wearer and cigarette all day to unobtrusively and passively detect smoking events. We also developed a deep learning-based framework to extract puffs and smoking topography. We evaluate SmokeMon in both controlled and free-living experiments with a total of 19 participants, more than 110 hours of data, and 115 smoking sessions achieving an F1-score of 0.9 for puff detection in the laboratory and 0.8 in the wild. By providing SmokeMon as an open platform, we provide measurement of smoking topography in free-living settings to enable testing of smoking topography in the real world, with potential to facilitate timely smoking cessation interventions.
Recent developments of novel in-vehicle interventions show the potential to transform the otherwise routine and mundane task of commuting into opportunities to improve the drivers' health and well-being. Prior research has explored the effectiveness of various in-vehicle interventions and has identified moments in which drivers could be interruptible to interventions. All the previous studies, however, were conducted in either simulated or constrained real-world driving scenarios on a pre-determined route. In this paper, we take a step forward and evaluate when drivers interact with in-vehicle interventions in unconstrained free-living conditions. To this end, we conducted a two-month longitudinal study with 10 participants, in which each participant was provided with a study car for their daily driving needs. We delivered two in-vehicle interventions - each aimed at improving affective well-being - and simultaneously recorded the participants' driving behavior. In our analysis, we found that several pre-trip characteristics (like trip length, traffic flow, and vehicle occupancy) and the pre-trip affective state of the participants had significant associations with whether the participants started an intervention or canceled a started intervention. Next, we found that several in-the-moment driving characteristics (like current road type, past average speed, and future brake behavior) showed significant associations with drivers' responsiveness to the intervention. Further, we identified several driving behaviors that "negated" the effectiveness of interventions and highlight the potential of using such "negative" driving characteristics to better inform intervention delivery. Finally, we compared trips with and without intervention and found that both interventions employed in our study did not have a negative effect on driving behavior. Based on our analyses, we provide solid recommendations on how to deliver interventions to maximize responsiveness and effectiveness and minimize the burden on the drivers.
Mobile sensing is a ubiquitous and useful tool to make inferences about individuals' mental health based on physiology and behavior patterns. Along with sensing features directly associated with mental health, it can be valuable to detect different features of social contexts to learn about social interaction patterns over time and across different environments. This can provide insight into diverse communities' academic, work and social lives, and their social networks. We posit that passively detecting social contexts can be particularly useful for social anxiety research, as it may ultimately help identify changes in social anxiety status and patterns of social avoidance and withdrawal. To this end, we recruited a sample of highly socially anxious undergraduate students (N=46) to examine whether we could detect the presence of experimentally manipulated virtual social contexts via wristband sensors. Using a multitask machine learning pipeline, we leveraged passively sensed biobehavioral streams to detect contexts relevant to social anxiety, including (1) whether people were in a social situation, (2) size of the social group, (3) degree of social evaluation, and (4) phase of social situation (anticipating, actively experiencing, or had just participated in an experience). Results demonstrated the feasibility of detecting most virtual social contexts, with stronger predictive accuracy when detecting whether individuals were in a social situation or not and the phase of the situation, and weaker predictive accuracy when detecting the level of social evaluation. They also indicated that sensing streams are differentially important to prediction based on the context being predicted. Our findings also provide useful information regarding design elements relevant to passive context detection, including optimal sensing duration, the utility of different sensing modalities, and the need for personalization. We discuss implications of these findings for future work on context detection (e.g., just-in-time adaptive intervention development).
Smart ear-worn devices (called earables) are being equipped with various onboard sensors and algorithms, transforming earphones from simple audio transducers to multi-modal interfaces making rich inferences about human motion and vital signals. However, developing sensory applications using earables is currently quite cumbersome with several barriers in the way. First, time-series data from earable sensors incorporate information about physical phenomena in complex settings, requiring machine-learning (ML) models learned from large-scale labeled data. This is challenging in the context of earables because large-scale open-source datasets are missing. Secondly, the small size and compute constraints of earable devices make on-device integration of many existing algorithms for tasks such as human activity and head-pose estimation difficult. To address these challenges, we introduce Auritus an extendable and open-source optimization toolkit designed to enhance and replicate earable applications. Auritus serves two primary functions. Firstly, Auritus handles data collection, pre-processing, and labeling tasks for creating customized earable datasets using graphical tools. The system includes an open-source dataset with 2.43 million inertial samples related to head and full-body movements, consisting of 34 head poses and 9 activities from 45 volunteers. Secondly, Auritus provides a tightly-integrated hardware-in-the-loop (HIL) optimizer and TinyML interface to develop lightweight and real-time machine-learning (ML) models for activity detection and filters for head-pose tracking. To validate the utlity of Auritus, we showcase three sample applications, namely fall detection, spatial audio rendering, and augmented reality (AR) interfacing. Auritus recognizes activities with 91% leave 1-out test accuracy (98% test accuracy) using real-time models as small as 6-13 kB. Our models are 98-740× smaller and 3-6% more accurate over the state-of-the-art. We also estimate head pose with absolute errors as low as 5 degrees using 20kB filters, achieving up to 1.6× precision improvement over existing techniques. We make the entire system open-source so that researchers and developers can contribute to any layer of the system or rapidly prototype their applications using our dataset and algorithms.
The transition from high school to college is a taxing time for young adults. New students arriving on campus navigate a myriad of challenges centered around adapting to new living situations, financial needs, academic pressures and social demands. First-year students need to gain new skills and strategies to cope with these new demands in order to make good decisions, ease their transition to independent living and ultimately succeed. In general, first-generation students are less prepared when they enter college in comparison to non-first-generation students. This presents additional challenges for first-generation students to overcome and be successful during their college years. We study first-year students through the lens of mobile phone sensing across their first year at college, including all academic terms and breaks. We collect longitudinal mobile sensing data for N=180 first-year college students, where 27 of the students are first-generation, representing 15% of the study cohort and representative of the number of first-generation students admitted each year at the study institution, Dartmouth College. We discuss risk factors, behavioral patterns and mental health of first-generation and non-first-generation students. We propose a deep learning model that accurately predicts the mental health of first-generation students by taking into account important distinguishing behavioral factors of first-generation students. Our study, which uses the StudentLife app, offers data-informed insights that could be used to identify struggling students and provide new forms of phone-based interventions with the goal of keeping students on track.
Wearable sensors can provide reliable, automated measures of health behaviors in free-living populations. However, validation of these measures is impossible without observable confirmation of behaviors. Participants have expressed discomfort during the use of ego-centric wearable cameras with first-person view. We argue that mounting the camera on different body locations with a different lens orientation, gives a device recording affordance that has the effect of reducing surveillance and social discomfort compared to ego-centric cameras. We call these types of cameras "activity-oriented" because they are designed to capture a particular activity, rather than the field of view of the wearer. We conducted an experiment of three camera designs with 24 participants, collecting qualitative data on participants' experience while wearing these devices in the wild. We provide a model explaining factors that lead to an increase in social presence and social stigma, which, therefore, create social and surveillance discomfort for the wearer. Wearers' attempts to reduce this discomfort by modifying their behavior or abandoning the device threatens the validity of observations of authentic behaviors. We discuss design implications and provide recommendations to help reduce social presence and stigma in order to improve the validity of observations with cameras in the wild.
Recent advances in wearable sensor technologies have led to a variety of approaches for detecting physiological stress. Even with over a decade of research in the domain, there still exist many significant challenges, including a near-total lack of reproducibility across studies. Researchers often use some physiological sensors (custom-made or off-the-shelf), conduct a study to collect data, and build machine-learning models to detect stress. There is little effort to test the applicability of the model with similar physiological data collected from different devices, or the efficacy of the model on data collected from different studies, populations, or demographics. This paper takes the first step towards testing reproducibility and validity of methods and machine-learning models for stress detection. To this end, we analyzed data from 90 participants, from four independent controlled studies, using two different types of sensors, with different study protocols and research goals. We started by evaluating the performance of models built using data from one study and tested on data from other studies. Next, we evaluated new methods to improve the performance of stress-detection models and found that our methods led to a consistent increase in performance across all studies, irrespective of the device type, sensor type, or the type of stressor. Finally, we developed and evaluated a clustering approach to determine the stressed/not-stressed classification when applying models on data from different studies, and found that our approach performed better than selecting a threshold based on training data. This paper's thorough exploration of reproducibility in a controlled environment provides a critical foundation for deeper study of such methods, and is a prerequisite for tackling reproducibility in free-living conditions.
Just-In-Time Adaptive Intervention (JITAI) is an emerging technique with great potential to support health behavior by providing the right type and amount of support at the right time. A crucial aspect of JITAIs is properly timing the delivery of interventions, to ensure that a user is receptive and ready to process and use the support provided. Some prior works have explored the association of context and some user-specific traits on receptivity, and have built post-study machine-learning models to detect receptivity. For effective intervention delivery, however, a JITAI system needs to make in-the-moment decisions about a user's receptivity. To this end, we conducted a study in which we deployed machine-learning models to detect receptivity in the natural environment, i.e., in free-living conditions. We leveraged prior work regarding receptivity to JITAIs and deployed a chatbot-based digital coach - Ally - that provided physical-activity interventions and motivated participants to achieve their step goals. We extended the original Ally app to include two types of machine-learning model that used contextual information about a person to predict when a person is receptive: a static model that was built before the study started and remained constant for all participants and an adaptive model that continuously learned the receptivity of individual participants and updated itself as the study progressed. For comparison, we included a control model that sent intervention messages at random times. The app randomly selected a delivery model for each intervention message. We observed that the machine-learning models led up to a 40% improvement in receptivity as compared to the control model. Further, we evaluated the temporal dynamics of the different models and observed that receptivity to messages from the adaptive model increased over the course of the study.
Mobile-based ecological-momentary-assessment (EMA) is an in-situ measurement methodology where an electronic device prompts a person to answer questions of research interest. EMA has a key limitation: interruption burden. Microinteraction-EMA(μEMA) may reduce burden without sacrificing high temporal density of measurement. In μEMA, all EMA prompts can be answered with 'at a glance' microinteractions. In a prior 4-week pilot study comparing standard EMA delivered on a phone (phone-EMA) vs. μEMA delivered on a smartwatch (watch-μEMA), watch-μEMA demonstrated higher response rates and lower perceived burden than phone-EMA, even when the watch-μEMA interruption rate was 8 times more than phone-EMA. A new 4-week dataset was gathered on smartwatch-based EMA (i.e., watch-EMA with 6 back-to-back, multiple-choice questions on a watch) to compare whether the high response rates of watch-μEMA previously observed were a result of using microinteractions, or due to the novelty and accessibility of the smartwatch. No statistically significant differences in compliance, completion, and first-prompt response rates were observed between phone-EMA and watch-EMA. However, watch-μEMA response rates were significantly higher than watch-EMA. This pilot suggests that (1) the high compliance and low burden previously observed in watch-μEMA is likely due to the microinteraction question technique, not simply the use of the watch versus the phone, and that (2) compliance with traditional EMA (with long surveys) may not improve simply by moving survey delivery from the phone to a smartwatch.
'Turn slightly to the left' the navigational system announces, with the aim of directing a blind user to merge into a corridor. Yet, due to long reaction time, the user turns too late and proceeds into the wrong hallway. Observations of such user behavior in real-world navigation settings motivate us to study the manner in which blind users react to the instructional feedback of a turn-by-turn guidance system. We found little previous work analyzing the extent of the variability among blind users in reaction to different instructional guidance during assisted navigation. To gain insight into how navigational interfaces can be better designed to accommodate the information needs of different users, we conduct a data-driven analysis of reaction variability as defined by motion and timing measures. Based on continuously tracked user motion during real-world navigation with a deployed system, we find significant variability between users in their reaction characteristics. Specifically, the statistical analysis reveals significant variability during the crucial elements of the navigation (e.g., turning and encountering obstacles). With the end-user experience in mind, we identify the need to not only adjust interface timing and content to each user's personal walking pace, but also their individual navigation skill and style. The design implications of our study inform the development of assistive systems which consider such user-specific behavior to ensure successful navigation.
In automated sleep monitoring systems, bed occupancy detection is the foundation or the first step before other downstream tasks, such as inferring sleep activities and vital signs. The existing methods do not generalize well to real-world environments due to single environment settings and rely on threshold-based approaches. Manually selecting thresholds requires observing a large amount of data and may not yield optimal results. In contrast, acquiring extensive labeled sensory data poses significant challenges regarding cost and time. Hence, developing models capable of generalizing across diverse environments with limited data is imperative. This paper introduces SeismoDot, which consists of a self-supervised learning module and a spectral-temporal feature fusion module for bed occupancy detection. Unlike conventional methods that require separate pre-training and fine-tuning, our self-supervised learning module is co-optimized with the primary target task, which directs learned representations toward a task-relevant embedding space while expanding the feature space. The proposed feature fusion module enables the simultaneous exploitation of temporal and spectral features, enhancing the diversity of information from both domains. By combining these techniques, SeismoDot expands the diversity of embedding space for both the temporal and spectral domains to enhance its generalizability across different environments. SeismoDot not only achieves high accuracy (98.49%) and F1 scores (98.08%) across 13 diverse environments, but it also maintains high performance (97.01% accuracy and 96.54% F1 score) even when trained with just 20% (4 days) of the total data. This demonstrates its exceptional ability to generalize across various environmental settings, even with limited data availability.
Many patients with neurological disorders, such as Ataxia, do not have easy access to neurologists, -especially those living in remote localities and developing/underdeveloped countries. Ataxia is a degenerative disease of the nervous system that surfaces as difficulty with motor control, such as walking imbalance. Previous studies have attempted automatic diagnosis of Ataxia with the help of wearable biomarkers, Kinect, and other sensors. These sensors, while accurate, do not scale efficiently well to naturalistic deployment settings. In this study, we propose a method for identifying ataxic symptoms by analyzing videos of participants walking down a hallway, captured with a standard monocular camera. In a collaboration with 11 medical sites located in 8 different states across the United States, we collected a dataset of 155 videos along with their severity rating from 89 participants (24 controls and 65 diagnosed with or are pre-manifest spinocerebellar ataxias). The participants performed the gait task of the Scale for the Assessment and Rating of Ataxia (SARA). We develop a computer vision pipeline to detect, track, and separate the participants from their surroundings and construct several features from their body pose coordinates to capture gait characteristics such as step width, step length, swing, stability, speed, etc. Our system is able to identify and track a patient in complex scenarios. For example, if there are multiple people present in the video or an interruption from a passerby. Our Ataxia risk-prediction model achieves 83.06% accuracy and an 80.23% F1 score. Similarly, our Ataxia severity-assessment model achieves a mean absolute error (MAE) score of 0.6225 and a Pearson's correlation coefficient score of 0.7268. Our model competitively performed when evaluated on data from medical sites not used during training. Through feature importance analysis, we found that our models associate wider steps, decreased walking speed, and increased instability with greater Ataxia severity, which is consistent with previously established clinical knowledge. Furthermore, we are releasing the models and the body-pose coordinate dataset to the research community - the largest dataset on ataxic gait (to our knowledge). Our models could contribute to improving health access by enabling remote Ataxia assessment in non-clinical settings without requiring any sensors or special cameras. Our dataset will help the computer science community to analyze different characteristics of Ataxia and to develop better algorithms for diagnosing other movement disorders.
Opioid use disorder is a medical condition with major social and economic consequences. While ubiquitous physiological sensing technologies have been widely adopted and extensively used to monitor day-to-day activities and deliver targeted interventions to improve human health, the use of these technologies to detect drug use in natural environments has been largely underexplored. The long-term goal of our work is to develop a mobile technology system that can identify high-risk opioid-related events (i.e., development of tolerance in the setting of prescription opioid use, return-to-use events in the setting of opioid use disorder) and deploy just-in-time interventions to mitigate the risk of overdose morbidity and mortality. In the current paper, we take an initial step by asking a crucial question: Can opioid use be detected using physiological signals obtained from a wrist-mounted sensor? Thirty-six individuals who were admitted to the hospital for an acute painful condition and received opioid analgesics as part of their clinical care were enrolled. Subjects wore a noninvasive wrist sensor during this time (1-14 days) that continuously measured physiological signals (heart rate, skin temperature, accelerometry, electrodermal activity, and interbeat interval). We collected a total of 2070 hours (≈ 86 days) of physiological data and observed a total of 339 opioid administrations. Our results are encouraging and show that using a Channel-Temporal Attention TCN (CTA-TCN) model, we can detect an opioid administration in a time-window with an F1-score of 0.80, a specificity of 0.77, sensitivity of 0.80, and an AUC of 0.77. We also predict the exact moment of administration in this time-window with a normalized mean absolute error of 8.6% and R 2 coefficient of 0.85.
Understanding the dynamics of mental health among undergraduate students across the college years is of critical importance, particularly during a global pandemic. In our study, we track two cohorts of first-year students at Dartmouth College for four years, both on and off campus, creating the longest longitudinal mobile sensing study to date. Using passive sensor data, surveys, and interviews, we capture changing behaviors before, during, and after the COVID-19 pandemic subsides. Our findings reveal the pandemic's impact on students' mental health, gender based behavioral differences, impact of changing living conditions and evidence of persistent behavioral patterns as the pandemic subsides. We observe that while some behaviors return to normal, others remain elevated. Tracking over 200 undergraduate students from high school to graduation, our study provides invaluable insights into changing behaviors, resilience and mental health in college life. Conducting a long-term study with frequent phone OS updates poses significant challenges for mobile sensing apps, data completeness and compliance. Our results offer new insights for Human-Computer Interaction researchers, educators and administrators regarding college life pressures. We also detail the public release of the de-identified College Experience Study dataset used in this paper and discuss a number of open research questions that could be studied using the public dataset.
Sleep behavior significantly impacts health and acts as an indicator of physical and mental well-being. Monitoring and predicting sleep behavior with ubiquitous sensors may therefore assist in both sleep management and tracking of related health conditions. While sleep behavior depends on, and is reflected in the physiology of a person, it is also impacted by external factors such as digital media usage, social network contagion, and the surrounding weather. In this work, we propose SleepNet, a system that exploits social contagion in sleep behavior through graph networks and integrates it with physiological and phone data extracted from ubiquitous mobile and wearable devices for predicting next-day sleep labels about sleep duration. Our architecture overcomes the limitations of large-scale graphs containing connections irrelevant to sleep behavior by devising an attention mechanism. The extensive experimental evaluation highlights the improvement provided by incorporating social networks in the model. Additionally, we conduct robustness analysis to demonstrate the system's performance in real-life conditions. The outcomes affirm the stability of SleepNet against perturbations in input data. Further analyses emphasize the significance of network topology in prediction performance revealing that users with higher eigenvalue centrality are more vulnerable to data perturbations.
Identifying and planning strategies that support a healthy lifestyle or manage a chronic disease often require patient-provider collaboration. For example, people with healthy eating goals often share everyday food, exercise, or sleep data with health coaches or nutritionists to find opportunities for change, and patients with irritable bowel syndrome (IBS) often gather food and symptom data as part of working with providers to diagnose and manage symptoms. However, a lack of effective support often prevents health experts from reviewing large amounts of data in time-constrained visits, prevents focusing on individual goals, and prevents generating correct, individualized, and actionable recommendations. To examine how to design photo-based diaries to help people and health experts exchange knowledge and focus on collaboration goals when reviewing the data together, we designed and developed Foodprint, a photo-based food diary. Foodprint includes three components: (1) A mobile app supporting lightweight data collection, (2) a web app with photo-based visualization and quantitative visualizations supporting collaborative reflection, and (3) a pre-visit note communicating an individual's expectations and questions to experts. We deployed Foodprint in two studies: (1) with 17 people with healthy eating goals and 7 health experts, and (2) with 16 IBS patients and 8 health experts. Building upon the lens of boundary negotiating artifacts and findings from two field studies, our research contributes design principles to (1) prepare individuals to collect data relevant to their health goals and for collaboration, (2) help health experts focus on an individual's eating context, experiences, and goals in collaborative review, and (3) support individuals and experts to develop individualized, actionable plans and strategies.