Health & Sensing at CHI'25
Highlights and Reflections from the 2025
1. Introduction to CHI conferences
The ACM CHI Conference on Human Factors in Computing Systems (CHI) is the premier international venue for research in Human-Computer Interaction (HCI). It brings together a global community of researchers, designers, and practitioners to explore how interactive technologies shape and are shaped by people and society.
At the core of CHI is its technical program, featuring peer-reviewed papers that present advances across a wide range of HCI topics, such as interaction techniques, accessibility, artificial intelligence, and digital health. In addition to paper sessions, CHI includes Workshops, Courses, Posters, Panels, Case Studies, and Doctoral Consortium, which support interdisciplinary exchange and practical exploration.
CHI also serves as a forum for addressing broader concerns in HCI, including ethical and inclusive design, responsible AI, sustainability, and the social impact of emerging technologies. Each year, the conference is held in a different location worldwide.
This year, CHI 2025 took place in Yokohama, Japan, from April 26 to May 1. The conference opened with a striking calligraphy performance during the Opening Plenary, where the artist painted the CHI 2025 logo alongside the characters “超界”, meaning transcending boundaries.
For the full conference program, visit: https://programs.sigchi.org/chi/2025/program/all
2. Keynote Speech & Discussion: AI for the People, Computing for a Better World
Attending the CHI 2025 opening keynote, “AI for the People, Computing for a Better World” by Mutale Nkonde, was a powerful and eye-opening experience that challenged conventional approaches to technology by centering justice, inclusion, and lived experience. Rather than focusing on technological innovation alone, Nkonde invited the audience to reframe their thinking—asking not “What can this technology do?” but “Who is affected, and how?” Drawing from Black Queer theory and her work with AI for the People, a Black female-led nonprofit, she demonstrated how marginalized perspectives can illuminate biases and harms embedded in algorithmic systems, while offering critical tools for building more equitable and human-centered AI. Through concrete examples from policy advocacy and inclusive design practices, Nkonde highlighted how values-driven development can actively counter surveillance and discrimination, ultimately positioning AI as a tool for collective liberation. This keynote not only broadened my understanding of the societal dimensions of computing but also set a compelling tone for the rest of CHI 2025—reminding us that computing must be grounded in responsibility, not just innovation.
3. Main Sessions: Paper Presentation
3.1. Lifetime Digital Health
This session examined the intersection of digital technologies and long-term health, highlighting how mobile sensing, LLM-based agents, and AI-driven systems can support mental and physical well-being throughout someone’s life. It showcased methods for designing AI interventions that evolve with users over time; insights directly relevant to digital health researchers building sustainable, user-centered health solutions.
JournalAIde: Empowering Older Adults in Digital Journal Writing
https://dl.acm.org/doi/10.1145/3706598.3713339
Digital journaling can offer older adults cognitive, emotional, and social benefits, yet many face challenges such as low confidence, difficulty organizing ideas, and sustaining attention. To address this, the authors developed JournalAIde, an interactive system powered by large language models designed to support older adults in digital journaling through idea organization via chatbot interactions, LLM-assisted text generation, and visual cues for editing. Through a lab-based between-subjects study and a 10-day field deployment, the study found that JournalAIde significantly improved users’ writing confidence, ability, and engagement, while also raising important considerations about authorship, reliance, and transparency in human-AI writing collaboration. The findings highlight design opportunities for AI-assisted creativity tools that are tailored to older adults’ cognitive and emotional needs.
Although this study focused on supporting older adults in digital journaling, I found many of its ideas broadly applicable to the design of AI-assisted tools in other domains. What stood out to me was how the system was carefully designed to empower users rather than take over the task, striking a balance between AI assistance and user agency. The use of peer writing examples to build confidence, and the integration of LLM-powered features like chatbot-guided idea generation and visual editing cues, all reflect a strong user-centered design mindset. These elements could easily inspire similar applications in areas like mental health support, education, or lifelong learning. The paper also provides insights into how older adults perceive authorship, collaboration, and reliance when working with AI, which I think are important questions in HCI domain. If you're interested in human-AI collaboration, I highly recommend reading this paper!
Towards Personalized Physiotherapy through Interactive Machine Learning: A Conceptual Infrastructure Design for In-Clinic and Out-of-Clinic Support
https://dl.acm.org/doi/10.1145/3706598.3713823
The paper explores how interactive machine learning can be integrated into physiotherapy not just as a technical solution, but as a sociotechnical partner that supports the therapist–patient relationship. The authors propose a distributed structure that spans both in-clinic and out-of-clinic settings, identifying key roles for ML across this loop, including personalization, reflection, encouragement, and feedback. Through interviews, workshops, and speculative design, they conceptualize an ML infrastructure that fosters care practices by enabling therapists and patients to collaboratively engage with ML systems. Rather than focusing solely on automation or efficiency, the work emphasizes designing for relational care and long-term therapeutic engagement.
What I found most compelling about this paper was that they considered machine learning not just as a system for improving efficiency but as an active participant in the practice of care. By embedding ML into the dynamic between therapists and patients and attending to their emotional as well as situational needs, the authors introduce a thoughtful and human-centered design perspective. The concept of a feedback loop connecting in-clinic and at-home experiences was particularly insightful for me, as it mirrors how physiotherapy actually unfolds in everyday life. I also appreciated the focus on interactive ML that adapts over time and offers not only practical guidance but also emotional support. This approach feels especially relevant for domains like mental health or chronic disease care, where sustained engagement and trust are essential. For those exploring human-AI partnerships grounded in empathy and co-experience, this paper is worth reading.
3.2. Autonomous Vehicles
This session focused on autonomous vehicles from HCI perspectives, specifically how people perceive autonomous vehicles and how they interact, or should interact, with people. Not only did this session provide insights into our lab’s current research on Human-Vehicle Interaction (HVI), focusing on predicting drivers’ behavioral and cognitive states in-car, it also introduced a couple of interesting methodologies applicable to other HCI domains, such as digital health.
OptiCarVis: Improving Automated Vehicle Functionality Visualizations Using Bayesian Optimization to Enhance User Experience
https://dl.acm.org/doi/10.1145/3706598.3713514
Researchers have studied how to improve visualizations for autonomous vehicles to ensure drivers’ trust, perceived safety, and acceptance. As a result, we now have an abundance of design choices, such as the transparency and size of visual elements. Building on prior studies that underscored individual differences in the perceptions of those visual elements, this study proposed a personalization strategy using human-in-the-loop, multi-objective Bayesian optimization. The authors first defined visual elements that can be parameterized for personalization (design parameters), and criteria for optimization (objective functions), including safety, trust, predictability, acceptance, and aesthetics. Their human-in-the-loop optimization algorithm then optimized these design parameters based on users’ subjective ratings. This personalization algorithm was compared in six different variations, including a non-personalized one, through an online study. The authors shared findings, including that their personalization algorithm significantly outperformed others in terms of trust, acceptance, perceived safety, and predictability without increasing cognitive load.
Even though this study focused on visualizing driving-related explanations, I found their approach to personalization can be applied to other domains, such as personalizing digital health interventions, by considering different health objectives or personal beliefs and values. Moreover, exploring the final parameter sets after optimization can also provide practical insights, as the study’s authors demonstrated. If you’re interested in personalization, I would recommend reading the paper. (This paper also received an Honorable Mention!)
People Attribute Purpose to Autonomous Vehicles When Explaining Their Behavior: Insights from Cognitive Science for Explainable AI
https://dl.acm.org/doi/10.1145/3706598.3713509
Prior work on explainable AI has mainly focused on how to generate explanations from data, with most of them providing “maybe-causes” explanations. This study delved into different types of explanations by understanding how humans generate explanations for driving contexts. It builds on the Framework of Explanatory Modes, which consists of:
Teleological explanation (e.g., Traffic laws are there to coordinate driver behavior and prevent accidents; in the form of ‘x was done to bring about y’)
Mechanistic explanation (e.g., The car stopped because it ran out of gas; in the form of ‘y happened because x happened’)
Counterfactual explanation (e.g., If the car hadn’t braked, it would have hit the pedestrian; in the form of ‘if x had been done then y would have happened’)
Descriptive explanation (e.g., The car stopped because it ground to a halt; in the form of ‘x because x’)
The authors first collected different types of explanations on Prolific using the aforementioned framework, and then gathered evaluations of these explanations using various criteria. Through their study, the authors observed that teleological and mechanistic explanations are preferred. Furthermore, they noted that changing the subject of the vehicle action from human to machine (i.e., autonomous vehicle) did not change the result, indicating that people also ascribe teleological concepts to autonomous vehicles.
I found the introduction of various explanatory modes from humans’ perspectives very useful. The preference for explanatory modes may differ in other domains like digital health; Alternatively, people might prefer teleological explanations in digital health as well over data-driven, personalized explanations. It would be interesting future work to see which types of explanations people prefer when trying to understand their current health behaviors and make decisions, and how researchers should design explanations to engage users for behavior change.
4. LBW Sessions: Poster Presentation
4.1 LLM-Based Clinical Reasoning for Health Prediction
CLONE: Synthetic Guideline-based Clinical Reasoning with Large Language Models for Early Diagnosis of Mild Cognitive Impairment
https://doi.org/10.1145/3706599.3720111
Recent advancements in large language models (LLMs) have opened new avenues for interpretable, scalable, and expert-like reasoning in health prediction and user modeling. One particularly interesting development is the CLONE framework, which was designed for the early diagnosis of Mild Cognitive Impairment (MCI) and demonstrates how LLMs can be harnessed for guideline-based clinical reasoning.
The early detection of MCI is essential for timely intervention, yet it remains a persistent challenge in clinical practice. Traditional expert-driven diagnostics, while thorough, tend to be labor-intensive and often lack transparency, making it difficult for clinicians and patients to fully trust the outcomes. On the other hand, standard machine learning models, despite their accuracy, usually fall short in providing interpretable rationales, which limits their adoption in real-world healthcare settings.
CLONE addresses these issues by introducing a thoughtfully structured, three-stage pipeline. First, the LLM is prompted to role-play as a neuropsychologist, interpreting patient data with a clinician’s mindset. Next, the model synthesizes its reasoning into step-by-step diagnostic guidelines, making the thought process transparent and traceable. Finally, a second LLM stage takes these synthesized guidelines and uses them to make final diagnostic decisions, ensuring a clear separation between reasoning and decision-making. This two-pass design is key to enhancing both transparency and reproducibility in clinical AI.
The effectiveness of CLONE was evaluated using a dataset of 65 real-world patient cases for MCI screening. The results were impressive: CLONE achieved an accuracy of 89.23%, outperforming the few-shot Chain-of-Thought (CoT) baseline by 6.15% and improving specificity by 10.71%. What’s more, expert reviewers found the explanations produced by CLONE to be more consistent, specific, helpful, and human-like than those generated by baseline models, highlighting the system’s potential for delivering trustworthy clinical reasoning.
By demonstrating that LLMs can effectively emulate expert reasoning and produce human-readable, guideline-based rationales for clinical decisions, CLONE not only improves diagnostic accuracy but also tackles the long-standing challenge of interpretability in AI-driven health prediction. Looking to the future, this framework could be scaled up to larger and more diverse clinical datasets. It also opens exciting opportunities for integrating retrieval-augmented or multi-agent LLM reasoning frameworks and extending guideline-based modeling to other domains, such as mood, energy expenditure, and neurological disorders like Parkinson’s Disease or sleep stage analysis.
Ultimately, CLONE reflects a broader paradigm shift in user modeling and health prediction—a move away from black-box models and toward transparent, guideline-driven LLM architectures. This shift promises to bring greater trust, adoption, and generalizability to the next generation of clinical AI systems.
4.2 LLM-Based Mental Health Assessment
Niclas Rosteck, Julian Striegl, and Claudia Loitsch. 2025. Bridging the Treatment Gap: A Novel LLM-Driven System for Scalable Initial Patient Assessments in Mental Healthcare.
https://doi.org/10.1145/3706599.3720043
Mental health care is facing a major bottleneck: growing demand versus limited clinician resources. Traditional intake interviews are both time-consuming and costly, leaving many patients underserved. In response, Rosteck, Striegl, and Loitsch (CHI 2025) introduce a hybrid LLM-driven system designed to automate initial mental health assessments—an important step toward closing the treatment gap.
The core design blends rule-based dialogue management with the semantic intelligence of LLMs. Rule-based modules ensure coverage of essential clinical domains, making sure the system asks all the right questions. Meanwhile, the LLM component adapts dynamically to each patient’s responses—asking follow-up questions, adjusting tone, and navigating conversational flow as a clinician would. This hybrid design achieves clinical comprehensiveness while maintaining conversation fluidity, avoiding the rigidity of scripts or the unpredictability of open-ended chatbots.
Experts evaluated the system via clinician walkthroughs, comparing its performance to in-person intakes. The findings were encouraging: the hybrid agent matched human interviewers in both topic coverage and conversational coherence . It effectively balances rigorous, structured data collection with patient-centered, adaptive dialogue, which are vital for building trust in mental health settings.
This paper marks a significant milestone in LLM applications for mental health triage. By coupling structured guidance with LLM adaptability, it demonstrates a scalable, semantically sophisticated, and clinically reliable tool for initial patient intake. As mental health systems continue to strain under demand, such approaches may significantly improve access and efficiency in preliminary assessments.
Looking forward, future work should involve real-world user trials, assessing patient perceptions of trust, comfort, and engagement. Comparative studies of purely LLM-generated, hybrid, and traditional rule-based intake systems would help clarify performance trade-offs. Moreover, integrating ethical safeguards, privacy protections, and clinical risk detection mechanisms will be essential before these systems enter real-world deployment.
5. Learning
We would like to conclude this article with brief remarks from the authors who contributed to this article.
Yugyeong
Attending CHI 2025 was a meaningful experience for me. Preparing for the presentation gave me the chance to revisit the core motivations behind my work - why I started, what I hoped to learn, and who I wanted it to impact. During the conference, I had so many valuable conversations with researchers, many of whom brought different perspectives to similar problems. CHI gave me a chance to step back and think about the bigger picture of HCI, and what kind of researcher I want to be moving forward.
Sueun
I not only had the chance to talk to lots of brilliant, like-minded people, but I also learned through repeated trials and errors how to introduce my work to those who might not be familiar with my topic. I don’t think I’ve found the perfect introduction yet, but hopefully, I’ll be more prepared for my next work and my next trip to CHI. It was a great opportunity to attend CHI and expose myself to the trends and people of HCI.
Gyuna
Before attending CHI, I looked forward to engaging with research through live presentations and conversations with diverse researchers. Throughout the conference, I found myself not only deepening my interest in familiar areas but also exploring new topics that broadened my perspective. Presentations and poster sessions helped reveal the motivations behind each study and made the research feel more engaging. Most importantly, it prompted me to reflect on what makes research meaningful to me and the kind of researcher I aspire to be.
Jeonghyun
Attending CHI 2025 was my first academic conference, and it was a truly unforgettable experience. Seeing the authors, whose work I had only read as text, come to life as they passionately shared and discussed their research was incredibly inspiring. It was hard to believe that by gathering the courage to speak up, I was having conversations with these researchers. Listening to the presentations helped me distinguish between research topics that captured my interest and those that didn’t, allowing me to deeply reflect on and explore the type of research I truly want to pursue.
Panyu
This was my first time attending CHI, and I found the experience extremely valuable. Engaging with people from diverse backgrounds allowed me to receive meaningful feedback and gain new perspectives. I was especially impressed by the number of attendees interested in causal analysis using mobile sensor data. The feedback I received on how to apply these techniques in human-centered research was particularly helpful.
Bio
Yugyeong
Yugyeong Jung is a fourth-year Ph.D. student in Computer Science at KAIST. Her research focuses on designing LLM-based mental health agents that support emotional reflection and intervention, as well as building systems for managing data quality in mobile sensing environments. She is interested in how AI systems can better understand and respond to human experiences in the wild, especially in emotionally and contextually rich settings.
Sueun
Sueun Jang is a 2nd year PhD student in the School of Computing at KAIST. Her research interest lies in digital health. She is currently working on personalized behavior change interventions for digital wellbeing and sensor data-based understanding of human behaviors and cognitive states.
Gyuna
Gyuna Kim is a Master’s student in the Graduate School of Data Science at KAIST. Her research interests include AI/ML in healthcare, digital phenotyping, multimodal data analysis, and interactive systems. She aims to develop intelligent and interpretable technologies that support mental health and wellbeing in everyday scenarios.
Jeonghyun
Jeonghyun Kim is a Master’s student in School of Computing at KAIST. Her research focuses on designing and evaluating interactive systems that empower people in their daily lives, with a special interest in healthcare and personalized computing. Currently, she is working on stress management systems supporting self-awareness using diverse everyday data such as wearable and mobile sensing.
Panyu
Panyu Zhang is a third-year doctoral student at the KAIST Graduate School of Data Science, where he focuses on mobile affective sensing. His primary research interests include the reproducibility and generalizability of mobile affective sensing methods. By exploring these areas, he aims to develop adaptation and generalization algorithms that enable models to function effectively across different users. Ultimately, his work is dedicated to making mobile affective sensing technologies more robust and applicable to a wide range of real-world, human-centered scenarios.
