Deborah Estrin

From ELIZA to Alexa, Conversational Agents (CAs) have been deliberately designed to elicit or project empathy. Although empathy can help technology better serve human needs, it can also be deceptive and potentially exploitative. In this work, we characterize empathy in interactions with CAs, highlighting the importance of distinguishing evocations of empathy between two humans from ones between a human and a CA. To this end, we systematically prompt CAs backed by large language models (LLMs) to display empathy while conversing with, or about, 65 distinct human identities, and also compare how different LLMs display or model empathy. We find that CAs make value judgments about certain identities, and can be encouraging of identities related to harmful ideologies (eg, Nazism and xenophobia). Moreover, a computational approach to understanding empathy reveals that despite their ability to display empathy, CAs do poorly when interpreting and exploring a user’s experience, contrasting with their human counterparts.

Background: Digital health-tracking tools are changing mental health care by giving patients the ability to collect passively measured patient-generated health data (PGHD; ie, data collected from connected devices with little to no patient effort). Although there are existing clinical guidelines for how mental health clinicians should use more traditional, active forms of PGHD for clinical decision-making, there is less clarity on how passive PGHD can be used.Objective: We conducted a qualitative study to understand mental health clinicians’ perceptions and concerns regarding the use of technology-enabled, passively collected PGHD for clinical decision-making. Our interviews sought to understand participants’ current experiences with and visions for using passive PGHD.Methods: Mental health clinicians providing outpatient services were recruited to participate in semistructured interviews. Interview recordings were deidentified, transcribed, and qualitatively coded to identify overarching themes.Results: Overall, 12 mental health clinicians (n= 11, 92% psychiatrists and n= 1, 8% clinical psychologist) were interviewed. We identified 4 overarching themes. First, passive PGHD are patient driven—we found that current passive PGHD use was patient driven, not clinician driven; participating clinicians only considered passive PGHD for clinical decision-making when patients brought passive data to clinical encounters. The second theme was active versus passive data as subjective versus objective data—participants viewed the contrast between active and passive PGHD as a contrast between interpretive data on patients’ mental health and objective …

We focus on the stories of five older adults who became voice assistant users through our study, and with whom we speculated about future interfaces through two design probes, one for health data reporting and one for positive reminiscing. We delivered a voice-first ambient interface (VFAI) to each participant, and closely observed participants’ journeys through periodic themed interviews (16 hours, 21 minutes of transcribed recordings), usage log reviews (4,657 entries), and phone and text support. Participants’ lived experiences impacted their perceptions and interactions with their VFAI, fueling rich insights about how to design for diverse needs. For example, while one participant saw increased potential in the VFAI after interacting with the design probe for health data reporting, another was skeptical of using it to communicate with her doctor. We contribute an in-depth exploration of VFAIs to support aging in …

Background Over the last decade, augmented reality (AR) has emerged in health care as a tool for visualizing data and enhancing simulation learning. AR, which has largely been explored for communication and collaboration in nonhealth contexts, could play a role in shaping future remote medical services and training. This review summarized existing studies implementing AR in real-time telemedicine and telementoring to create a foundation for health care providers and technology developers to understand future opportunities in remote care and education. Objective This review described devices and platforms that use AR for real-time telemedicine and telementoring, the tasks for which AR was implemented, and the ways in which these implementations were evaluated to identify gaps in research that provide opportunities for further study. Methods We searched PubMed, Scopus, Embase, and MEDLINE to identify English-language studies published between January 1, 2012, and October 18, 2022, implementing AR technology in a real-time interaction related to telemedicine or telementoring. The search terms were “augmented reality” OR “AR” AND “remote” OR “telemedicine” OR “telehealth” OR “telementoring.” Systematic reviews, meta-analyses, and discussion-based articles were excluded from analysis. Results A total of 39 articles met the inclusion criteria and were categorized into themes of patient evaluation, medical intervention, and education. In total, 20 devices and platforms using AR were identified, with common features being the ability for remote …

Background Augmented reality (AR) and virtual reality (VR) have increasingly appeared in the medical literature in the past decade, with AR recently being studied for its potential role in remote health care delivery and communication. Recent literature describes AR’s implementation in real-time telemedicine contexts across multiple specialties and settings, with remote emergency services in particular using AR to enhance disaster support and simulation education. Despite the introduction of AR in the medical literature and its potential to shape the future of remote medical services, studies have yet to investigate the perspectives of telemedicine providers regarding this novel technology. Objective This study aimed to understand the applications and challenges of AR in telemedicine anticipated by emergency medicine providers with a range of experiences in using telemedicine and AR or VR technology. Methods Across 10 academic medical institutions, 21 emergency medicine providers with variable exposures to telemedicine and AR or VR technology were recruited for semistructured interviews via snowball sampling. The interview questions focused on various potential uses of AR, anticipated obstacles that prevent its implementation in the telemedicine area, and how providers and patients might respond to its introduction. We included video demonstrations of a prototype using AR during the interviews to elicit more informed and complete insights regarding AR’s potential in remote health care. Interviews were transcribed and analyzed via thematic coding. Results …

Digital biomarkers of mental health, created using data extracted from everyday technologies including smartphones, wearable devices, social media and computer interactions, have the opportunity to revolutionise mental health diagnosis and treatment by providing near-continuous unobtrusive and remote measures of behaviours associated with mental health symptoms. Machine learning models process data traces from these technologies to identify digital biomarkers. In this editorial, we caution clinicians against using digital biomarkers in practice until models are assessed for equitable predictions (‘model equity’) across demographically diverse patients at scale, behaviours over time, and data types extracted from different devices and platforms. We posit that it will be difficult for any individual clinic or large-scale study to assess and ensure model equity and alternatively call for the creation of a repository of …

Digital health advances have transformed many clinical areas including psychiatric and cardiovascular care. However, digital health innovation is relatively nascent in cancer care, which represents the fastest growing area of health-care spending. Opportunities for digital health innovation in oncology include patient-facing technologies that improve patient experience, safety, and patient-clinician interactions; clinician-facing technologies that improve their ability to diagnose pathology and predict adverse events; and quality of care and research infrastructure to improve clinical workflows, documentation, decision support, and clinical trial monitoring. The COVID-19 pandemic and associated shifts of care to the home and community dramatically accelerated the integration of digital health technologies into virtually every aspect of oncology care. However, the pandemic has also exposed potential flaws in the …

Older adults around the world lack access to a wide range of potentially life-changing digital applications, services, and information that could be provided by voice assistants (such as Amazon’s Alexa, Google’s Assistant, or Apple’s Siri). However, older adults’ needs are underrepresented in the design of voice assistants. Because of this, we are missing opportunities for digital inclusion, and increasing risks of excluding older adults as these devices permeate public settings. In this work, we video record older adults (n=26) interacting with a multi-modal voice assistants while waiting in line at food pantries, and use Interaction Analysis to draw insights from these recordings. We find that by being agnostic to body language, audio-prosodic features, and other contextual factors, voice assistants fail to capture and react to some important aspects of interactions. We discuss design (e.g, interpreting users’ posture as a …

In this short paper we explore the opportunities and challenges of designing XR technologies to support the collaborative work between family caregivers and clinicians as they attend to the physical care needs of patients in the home setting.

Background: Exercise has showed to reduce pain and improve function in patients with chronic lower back pain (CLBP), however, outcomes have been inconsistent. Limbr is a mobile application (app) developed to increase patient adherence to the exercise program, and enhance exercise’s technique for better results.Methods: Patients 18 to 65 years of age, with CLBP (more than 3 months) and evidence of lumbar disc pathology by magnetic resonance imaging (MRI) were enrolled. Patients’ symptomatology was prospectively evaluated at baseline and during 3 months of using the Limbr app with patient-reported outcome measures (PROMs) including: the Visual Analog Scale for pian (VAS), the Oswestry Disability Index (ODI), and the Your Activities of Daily Living (YADL) score. Patients’ compliance and satisfaction were evaluated at the end of the follow-up.Results: Seventy-five patients with CLBP were enrolled in the study. All patients had a significant improvement from baseline to final follow-up in all PROMs. Average VAS scores decreased from 5.17±2.1 at baseline to 3.8±2.6 at final follow-up (P= 0.016), and ODI scores from 25.7±13.8 to 18.2±13.6 (P= 0.001). Patients showed a significant decrease in the number of pain medications taken during a week (P= 0.001). Overall compliance with the app was greater than 52%, with almost 65% of the patients rating the overall experience as good or excellent.Conclusion: Limbr showed to be an effective tool in reducing low back pain symptomatology, as well as generating great compliance to an exercise program for patients with CLBP.Trial registration number: 2/2/2017 NCT03040310 …

DiscussionNon-completion of GA tools such as the eRFA may result in undiagnosed or undertreated frailty. Our preliminary data has shown that patients with higher degree of frailty required more assistance in completing the eRFA, which could lead to non-completion and have a negative impact on outcomes on patients who may need treatment for frailty the most [8](manuscript is under consideration for publication). To address this problem, we developed a novel way of delivering the GA, the v-RFA, to

Background Participation in ambulatory cardiac rehabilitation remains low, especially among older adults. Although mobile health cardiac rehabilitation (mHealth-CR) provides a novel opportunity to deliver care, age-specific impairments may limit older adults’ uptake, and efficacy data are currently lacking. Objective This study aims to describe the design of the rehabilitation using mobile health for older adults with ischemic heart disease in the home setting (RESILIENT) trial. Methods RESILIENT is a multicenter randomized clinical trial that is enrolling patients aged ≥65 years with ischemic heart disease in a 3:1 ratio to either an intervention (mHealth-CR) or control (usual care) arm, with a target sample size of 400 participants. mHealth-CR consists of a commercially available mobile health software platform coupled with weekly exercise therapist sessions to review progress and set new activity goals. The primary outcome is a change in functional mobility (6-minute walk distance), which is measured at baseline and 3 months. Secondary outcomes are health status, goal attainment, hospital readmission, and mortality. Among intervention participants, engagement with the mHealth-CR platform will be analyzed to understand the characteristics that determine different patterns of use (eg, persistent high engagement and declining engagement). Results As of December 2021, the RESILIENT trial had enrolled 116 participants. Enrollment is projected to continue until October 2023. The trial results are expected to be reported in 2024. Conclusions The …

BackgroundIntervertebral disc pathology is the most common identifiable cause of chronic lower back pain (CLBP). There are limited conservative alternatives to treat discogenic axial CLBP. Back Rx is a mobile application (app) developed to treat patients with this condition, following the Back Rx exercise program, assisted by a virtual coach.MethodsPatients 18 to 65 years of age, with axial CLBP (more than 3 months), and evidence of lumbar disc pathology by magnetic resonance imaging (MRI) were enrolled to the study. Patients’ symptomatology was prospectively evaluated at baseline and after 3 months of using the Back Rx app. The main outcome of the study was back pain evaluated using the visual analog scale (VAS) for pain. Secondary outcomes were the patient's functionality, the weekly pain medication intake, the patients’ adherence to the app, and the patients´ satisfaction rate.ResultsSeventy-five …

We design a scalable algorithm to privately generate location heatmaps over decentralized data from millions of user devices. It aims to ensure differential privacy before data becomes visible to a service provider while maintaining high data accuracy and minimizing resource consumption on users' devices. To achieve this, we revisit distributed differential privacy based on recent results in secure multiparty computation, and we design a scalable and adaptive distributed differential privacy approach for location analytics. Evaluation on public location datasets shows that this approach successfully generates metropolitan-scale heatmaps from millions of user samples with a worst-case client communication overhead that is significantly smaller than existing state-of-the-art private protocols of similar accuracy.

There is now a significant and growing functional gap between the public Internet, whose basic architecture has remained unchanged for several decades, and a new generation of more sophisticated private networks. To address this increasing divergence of functionality and overcome the Internet's architectural stagnation, we argue for the creation of an Extensible Internet (EI) that supports in-network services that go beyond best-effort packet delivery. To gain experience with this approach, we hope to soon deploy both an experimental version (for researchers) and a prototype version (for early adopters) of EI. In the longer term, making the Internet extensible will require a community to initiate and oversee the effort; this paper is the first step in creating such a community.

Digital medical records have enabled us to employ clinical data in many new and innovative ways. However, these advances have brought with them a complex set of demands for healthcare institutions regarding data sharing with topics such as data ownership, the loss of privacy, and the protection of the intellectual property. The lack of clear guidance from government entities often creates conflicting messages about data policy, leaving institutions to develop guidelines themselves. Through discussions with multiple stakeholders at various institutions, we have generated a set of guidelines with 10 key principles to guide the responsible and appropriate use and sharing of clinical data for the purposes of care and discovery. Industry, universities, and healthcare institutions can build upon these guidelines toward creating a responsible, ethical, and practical response to data sharing.

Background: Mobile health technology has demonstrated the ability of smartphone apps and sensors to collect data pertaining to patient activity, behavior, and cognition. It also offers the opportunity to understand how everyday passive mobile metrics such as battery life and screen time relate to mental health outcomes through continuous sensing. Impulsivity is an underlying factor in numerous physical and mental health problems. However, few studies have been designed to help us understand how mobile sensors and self-report data can improve our understanding of impulsive behavior.Objective: The objective of this study was to explore the feasibility of using mobile sensor data to detect and monitor self-reported state impulsivity and impulsive behavior passively via a cross-platform mobile sensing application.Methods: We enrolled 26 participants who were part of a larger study of impulsivity to take part in a real-world, continuous mobile sensing study over 21 days on both Apple operating system (iOS) and Android platforms. The mobile sensing system (mPulse) collected data from call logs, battery charging, and screen checking. To validate the model, we used mobile sensing features to predict common self-reported impulsivity traits, objective mobile behavioral and cognitive measures, and ecological momentary assessment (EMA) of state impulsivity and constructs related to impulsive behavior (ie, risk-taking, attention, and affect).Results: Overall, the findings suggested that passive measures of mobile phone use such as call logs, battery charging, and screen checking can predict different facets of trait and state impulsivity and impulsive …

Background: The classic Marshmallow Test, where children were offered a choice between one small but immediate reward (eg, one marshmallow) or a larger reward (eg, two marshmallows) if they waited for a period of time, instigated a wealth of research on the relationships among impulsive responding, self-regulation, and clinical and life outcomes. Impulsivity is a hallmark feature of self-regulation failures that lead to poor health decisions and outcomes, making understanding and treating impulsivity one of the most important constructs to tackle in building a culture of health. Despite a large literature base, impulsivity measurement remains difficult due to the multidimensional nature of the construct and limited methods of assessment in daily life. Mobile devices and the rise of mobile health (mHealth) have changed our ability to assess and intervene with individuals remotely, providing an avenue for ambulatory diagnostic testing and interventions. Longitudinal studies with mobile devices can further help to understand impulsive behaviors and variation in state impulsivity in daily life.Objective: The aim of this study was to develop and validate an impulsivity mHealth diagnostics and monitoring app called Digital Marshmallow Test (DMT) using both the Apple and Android platforms for widespread dissemination to researchers, clinicians, and the general public.Methods: The DMT app was developed using Apple’s ResearchKit (iOS) and Android’s ResearchStack open source frameworks for developing health research study apps. The DMT app consists of three main modules: self-report, ecological momentary assessment, and active behavioral …

Background Lifelong learning is embedded in the culture of medicine, but there are limited tools currently available for many clinicians, including hospitalists, to help improve their own practice. Although there are requirements for continuing medical education, resources for learning new clinical guidelines, and developing fields aimed at facilitating peer-to-peer feedback, there is a gap in the availability of tools that enable clinicians to learn based on their own patients and clinical decisions. Objective The aim of this study was to explore the technologies or modifications to existing systems that could be used to benefit hospitalist physicians in pursuing self-assessment and improvement by understanding physicians’ current practices and their reactions to proposed possibilities. Methods Semistructured interviews were conducted in two separate stages with analysis performed after each stage. In the first stage, interviews (N=12) were conducted to understand the ways in which hospitalist physicians are currently gathering feedback and assessing their practice. A thematic analysis of these interviews informed the prototype used to elicit responses in the second stage. Results Clinicians actively look for feedback that they can apply to their practice, with the majority of the feedback obtained through self-assessment. The following three themes surrounding this aspect were identified in the first round of semistructured interviews: collaboration, self-reliance, and uncertainty, each with three related subthemes. Using a wireframe, the second round of interviews led to identifying the …

Mental illness often emerges during the formative years of adolescence and young adult development and interferes with the establishment of healthy educational, vocational, and social foundations. Despite the severity of symptoms and decline in functioning, the time between illness onset and receiving appropriate care can be lengthy. A method by which to objectively identify early signs of emerging psychiatric symptoms could improve early intervention strategies. We analyzed a total of 405,523 search queries from 105 individuals with schizophrenia spectrum disorders (SSD, N = 36), non-psychotic mood disorders (MD, N = 38) and healthy volunteers (HV, N = 31) utilizing one year’s worth of data prior to the first psychiatric hospitalization. Across 52 weeks, we found significant differences in the timing (p<0.05) and frequency (p<0.001) of searches between individuals with SSD and MD compared to HV up to a year in advance of the first psychiatric hospitalization. We additionally identified significant linguistic differences in search content among the three groups including use of words related to sadness and perception, use of first and second person pronouns, and use of punctuation (all p<0.05). In the weeks before hospitalization, both participants with SSD and MD displayed significant shifts in search timing (p<0.05), and participants with SSD displayed significant shifts in search content (p<0.05). Our findings demonstrate promise for utilizing personal patterns of online search activity to inform clinical care.

Federated models are created by aggregating model updates submittedby participants. To protect confidentiality of the training data, the aggregator by design has no visibility into how these updates aregenerated. We show that this makes federated learning vulnerable to amodel-poisoning attack that is significantly more powerful than poisoningattacks that target only the training data. A single or multiple malicious participants can use modelreplacement to introduce backdoor functionality into the joint model, eg, modify an image classifier so that it assigns an attacker-chosenlabel to images with certain features, or force a word predictor tocomplete certain sentences with an attacker-chosen word. We evaluatemodel replacement under different assumptions for the standardfederated-learning tasks and show that it greatly outperformstraining-data poisoning. Federated learning employs secure aggregation to protect confidentialityof participants’ local models and thus cannot detect anomalies inparticipants’ contributions to the joint model. To demonstrate thatanomaly detection would not have been effective in any case, we alsodevelop and evaluate a generic constrain-and-scale technique thatincorporates the evasion of defenses into the attacker’s loss functionduring training.

In this paper we present PoliFL, a decentralized, edge-based framework that supports heterogeneous privacy policies for federated learning. We evaluate our system on three use cases that train models with sensitive user data collected by mobile phones - predictive text, image classification, and notification engagement prediction - on a Raspberry Pi edge device. We find that PoliFL is able to perform accurate model training and inference within reasonable resource and time budgets while also enforcing heterogeneous privacy policies.