There are 50 million family caregivers caring for their loved ones in the United States. Caregivers experience significant stress and burnout, and need on-demand support with minimal resources investment. With the burgeoning development of artificial intelligence, conversational agents (chatbots) emerged as a solution for symptom self-management and self-care. Our research team developed Caring for Caregivers Online (COCO) - an AI-enhanced platform providing on-demand, empathetic, and tailored caregiving support. One of the key challenges in designing a chatbot is to ensure effective communication. To address this problem, our research team aims to determine the best practice in conveying health symptoms and solutions in a conversational agent to improve users' understanding of their health data and increase their trust in the technology. User testing and surveys are our main tools to address this question. After the initial session with the chatbot, users were asked to fill out a post-conversation survey. Based on users' ratings on their symptom intensity and solution effectiveness in the post-conversation survey, the chatbot generates personalized health recommendations. Next, we randomly assigned users to one of the two surveys that present the same recommendations in different ways—one with text, the other with a visualization. Users then provided ratings on how well they understood the visualization, their trust toward the chatbot, and some additional feedback. The results of the rating will be analyzed with an average score comparison between the two groups, and qualitative analysis will be employed to evaluate users' feedback. We expect the group presented with visualizations to give higher ratings in areas of both trust and data comprehension. Effective data visualizations inform caregivers of their health progress which motivates them to continue monitoring their health conditions and further practice proposed solutions. In addition, visualization increases caregivers' level of trust in chatbots and helps improve their chronic health conditions.

Bias in healthcare is often hidden and expressed through unintentionally prejudiced communication between providers and patients. These “implicit biases'' often relate to a patient’s race, gender, or sexual orientation. Implicit biases are automatic attitudes and stereotypes that can operate outside personal awareness and lead to unequal treatment, health disparities, and a lack of patient support. Although much research focuses on implicit bias, the perspectives of those who experience it and the impact it has on these individuals is less explored. Of particular importance are voices of Black, Indigenous, and People of color (BIPOC) and those with marginalized gender identities or sexual orientations (LGBTQ+). These groups have historically suffered from health inequities. For example, research indicates that BIPOC people may be undertreated for pain and LGBTQ+ people may be refused care. The UnBIASED project at the University of Washington and the University of California, San Diego addresses implicit bias in patient-provider communication through computational sensing tools to provide communication feedback. Through 25 interviews with people who identified as BIPOC, LGBTQ+, or both, we explored patients' perspectives on experiencing implicit bias when communicating with healthcare providers. We analyzed interviews through an inductive qualitative approach to understand negative and positive experiences, and identify participants' ideal solutions for improving patient-provider communication. For example, participants suggested having a patient advocate, providing feedback to the provider, and improving providers’ cultural competence. We report on these findings with the goal of describing common pain points and specific sources of dissatisfaction among patients who experience implicit bias. These findings help raise awareness of clinical implicit bias from the perspectives of patients, encourage further research, and suggest patient-driven, patient-centered solutions for how implicit bias can be overcome at personal and institutional levels.

The coronavirus pandemic has had significant global effects since the beginning of 2020. To slow the progression of a pandemic it is critical that the public cooperates with and follows health guidelines. In the United States, public health officials’ recommendations have been heavily politicized and disputed, resulting in confusion and uncertainty. Furthermore, the United States has had varied policies and messaging both over time and in different jurisdictions, augmenting contention. Twitter has emerged as a viable resource to explore public sentiment related to the pandemic, such as anxiety and fear. Tweets can also be leveraged to gauge the general understanding of the population, and to assess the effectiveness of public policy and communication from public health officials. In our work, we seek to understand how the public discourse varies in different areas of the United States over time. We first collect tweets about Covid-19 over a four-month period from March through July, 2020 from users in the United States. Next, we use Latent Dirichlet Allocation (LDA), an unsupervised algorithm that automatically constructs topics that appear in documents based on word co-occurrences. Using LDA, we present a topic-model of ten themes around Covid-19 that encapsulate various types of discourse throughout the pandemic. Furthermore, we examine trends in topic frequencies over time in each state, identify common patterns shared by states, and discuss these implications. Themes of anxiety are common, although their prevalence increases in some states over time, and decrease in others. Finally, topics of social distancing and wearing masks vary widely in sentiment and in accordance with local policy.