Florian von Wangenheim’s research while affiliated with ETH Zurich and other places

Trust is widely regarded as a critical component to building artificial intelligence (AI) systems that people will use and safely rely upon. As research in this area continues to evolve, it becomes imperative that the research community synchronizes its empirical efforts and aligns on the path toward effective knowledge creation. To lay the groundwork toward achieving this objective, we performed a comprehensive bibliometric analysis, supplemented with a qualitative content analysis of over two decades of empirical research measuring trust in AI, comprising 1’156 core articles and 36’306 cited articles across multiple disciplines. Our analysis reveals several “elephants in the room” pertaining to missing perspectives in global discussions on trust in AI, a lack of contextualized theoretical models and a reliance on exploratory methodologies. We highlight strategies for the empirical research community that are aimed at fostering an in-depth understanding of trust in AI.

Background: Blended mobile health (mHealth) interventions – combining self-guided and human support components – could play a major role in preventing non-communicable diseases (NCDs) and common mental disorders (CMDs). This protocol paper describes a sequential, multiple assignment, randomised trial aimed at (i) evaluating the effectiveness and cost-effectiveness of LvL UP, an mHealth lifestyle intervention for the prevention of NCDs and CMDs, and (ii) establishing the optimal blended approach in LvL UP that balances effective personalised lifestyle support with scalability.Methods: LvL UP is a 6-month mHealth holistic intervention targeting physical activity, diet, and emotional regulation. In this trial, young and middle-aged Singaporean adults at risk of developing NCDs or CMDs will be randomly allocated to one of two initial conditions (‘LvL UP’ or ‘comparison’). After 4 weeks, participants categorised as non-responders from the LvL UP group will be re-randomised into second-stage conditions: (i) continuing with the initial intervention (LvL UP) or (ii) additional motivational interviewing (MI) support sessions by trained health coaches (LvL UP + adaptive MI). The primary outcome is mental well-being (via the Warwick-Edinburgh Mental Wellbeing Scale). Secondary outcomes include anthropometric measurements, resting blood pressure, blood metabolic profile, health status, health behaviours (physical activity, diet), work productivity, and healthcare utilisation. Outcomes will be measured at baseline, 6 months (post-intervention), and 12 months (follow-up).Discussion: In addition to evaluating the effectiveness of LvL UP, the proposed study design will contribute to increasing evidence on how to introduce human support in mHealth interventions to maximise their effectiveness while remaining scalable.Trial registration: The LvL UP Pilot trial was prospectively registered with ClinicalTrials.gov (NCT06360029) on 7 April 2024.

Large language models (LLMs) present an enormous evolution in the strategic potential of conversational recommender systems (CRS). Yet to date, research has predominantly focused upon technical frameworks to implement LLM-driven CRS, rather than end-user evaluations or strategic implications for firms, particularly from the perspective of a small to medium enterprises (SME) that makeup the bedrock of the global economy. In the current paper, we detail the design of an LLM-driven CRS in an SME setting, and its subsequent performance in the field using both objective system metrics and subjective user evaluations. While doing so, we additionally outline a short-form revised ResQue model for evaluating LLM-driven CRS, enabling replicability in a rapidly evolving field. Our results reveal good system performance from a user experience perspective (85.5% recommendation accuracy) but underscore latency, cost, and quality issues challenging business viability. Notably, with a median cost of $0.04 per interaction and a latency of 5.7s, cost-effectiveness and response time emerge as crucial areas for achieving a more user-friendly and economically viable LLM-driven CRS for SME settings. One major driver of these costs is the use of an advanced LLM as a ranker within the retrieval-augmented generation (RAG) technique. Our results additionally indicate that relying solely on approaches such as Prompt-based learning with ChatGPT as the underlying LLM makes it challenging to achieve satisfying quality in a production environment. Strategic considerations for SMEs deploying an LLM-driven CRS are outlined, particularly considering trade-offs in the current technical landscape.

Decades of research have created a vast archive of information on human behavior, with relevant new studies being published daily. Despite these advances, knowledge generated by behavioral science – the social and biological sciences concerned with the study of human behavior – is not efficiently translated for those who will apply it to benefit individuals and society. The gap between what is known and the capacity to act on that knowledge continues to widen as current evidence synthesis methods struggle to process a large, ever-growing body of evidence characterized by its complexity and lack of shared terminologies. The purpose of the present position paper is twofold: (i) to highlight the pitfalls of traditional evidence synthesis methods in supporting effective knowledge translation to applied settings, and (ii) to sketch a potential alternative evidence synthesis approach which leverages on the use of ontologies – formal systems for organizing knowledge – to enable a more effec tive, artificial intelligence-driven accumulation and implementation of knowledge. The paper concludes with future research directions across behavioral, computer, and information sciences to help realize such innovative approach to evidence synthesis, allowing behavioral science to advance at a faster pace.

Background Heart rate variability biofeedback (HRV-BF) can be used for stress management. Recent feasibility studies suggest that delivering HRV-BF in virtual reality (VR) is associated with better user experience (UX) and might yield more beneficial changes in HRV than two-dimensional screens. The effectiveness of a VR-supported HRV-BF intervention program has, however, not been investigated yet. Methods In this study, 87 healthy women and men were assigned to a VR-supported HRV-BF intervention (INT; 𝑛=44 ) or a wait-list control (WLC; 𝑛=43 ) group. The INT came to the lab for four weekly HRV-BF sessions in VR using a head-mounted display. Between lab sessions, participants were asked to perform breathing exercises without biofeedback supported by a mobile application. Stress-related psychological and psychophysiological outcomes were assessed pre- and post-intervention and at a follow-up four weeks after the intervention in both groups. A psychosocial stress test was conducted post-intervention to investigate changes in stress reactivity. UX was assessed after each HRV-BF session in the INT. Results Analysis revealed that LF increased significantly from pre- to post-, whereas pNN50 increased and chronic stress decreased significantly from pre-intervention to follow-up in the INT compared to the WLC. Anxiety and mental fatigue decreased significantly, while mindfulness and health-related quality of life increased significantly from pre- to post- and from pre-intervention to follow-up in the INT compared to the WLC (all small effects). The two groups did not differ in their stress reactivity post-intervention. As for UX in the INT, the degree of feeling autonomous concerning technology adoption significantly decreased over time. Competence, involvement, and immersion, however, increased significantly from the first to the last HRV-BF session, while hedonic motivation significantly peaked in the second session and then gradually returned to first-session levels. Conclusions This HRV-BF intervention program, supported by VR and mobile technology, was able to significantly improve stress indicators and stress-related symptoms and achieved good to very good UX. Future studies should control for potential placebo effects and emphasize higher degrees of personalization and adaptability to increase autonomy and, thereby, long-term health and well-being. These findings may serve as a first step towards future HRV-BF applications of cutting-edge, increasingly accessible technologies, such as wearables, VR, and smartphones, in the service of mental health and healthcare. Trial registration The study was registered retrospectively as a clinical trial on ISRCTN registry (ISRCTN11331226, 26 May 2023).

Background The current paper details findings from Elena+: Care for COVID-19, an app developed to tackle the collateral damage of lockdowns and social distancing, by offering pandemic lifestyle coaching across seven health areas: anxiety, loneliness, mental resources, sleep, diet and nutrition, physical activity, and COVID-19 information. Methods The Elena+ app functions as a single-arm interventional study, with participants recruited predominantly via social media. We used paired samples T-tests and within subjects ANOVA to examine changes in health outcome assessments and user experience evaluations over time. To investigate the mediating role of behavioral activation (i.e., users setting behavioral intentions and reporting actual behaviors) we use mixed-effect regression models. Free-text entries were analyzed qualitatively. Results Results show strong demand for publicly available lifestyle coaching during the pandemic, with total downloads (N = 7′135) and 55.8% of downloaders opening the app (n = 3,928) with 9.8% completing at least one subtopic (n = 698). Greatest areas of health vulnerability as assessed with screening measures were physical activity with 62% (n = 1,000) and anxiety with 46.5% (n = 760). The app was effective in the treatment of mental health; with a significant decrease in depression between first (14 days), second (28 days), and third (42 days) assessments: F2,38 = 7.01, p = 0.003, with a large effect size (η2G = 0.14), and anxiety between first and second assessments: t54 = 3.7, p = <0.001 with a medium effect size (Cohen d = 0.499). Those that followed the coaching program increased in net promoter score between the first and second assessment: t36 = 2.08, p = 0.045 with a small to medium effect size (Cohen d = 0.342). Mediation analyses showed that while increasing number of subtopics completed increased behavioral activation (i.e., match between behavioral intentions and self-reported actual behaviors), behavioral activation did not mediate the relationship to improvements in health outcome assessments. Conclusions Findings show that: (i) there is public demand for chatbot led digital coaching, (ii) such tools can be effective in delivering treatment success, and (iii) they are highly valued by their long-term user base. As the current intervention was developed at rapid speed to meet the emergency pandemic context, the future looks bright for other public health focused chatbot-led digital health interventions.