Research
Research
Everhart A, Sen S, Stern A, Zhu Y, Karaca-Mandic P (2023) Association Between Regulatory Submission Characteristics and Recalls of Medical Devices Receiving 510(k) Clearance. Journal of the American Medical Association (JAMA) 329(2):144-156.
Zhu Y, Carroll C, Vu K, Sen S, Georgiou A, Karaca-Mandic P (2022) COVID-19 Hospitalization Trends in Rural Versus Urban Areas in the United States. Medical Care Research and Review 80(2):236-244.
Karaca-Mandic P, Sen S, Georgiou A, Zhu Y, Basu A (2020) Association of COVID-19-Related Hospital Use and Overall COVID-19 Mortality in the USA. Journal of General Internal Medicine.
Zhu Y, Sen S, Everhart A, Karaca-Mandic P. A Deep Learning Approach for Predicting FDA 510(k) Medical Device Recalls Using Device Citation Relationships.
Abstract: More than 90% of medical devices in the US enter the market through the FDA’s 510(k) clearance pathway, which is primarily based on demonstrating the equivalence of new devices (known as applicant devices) to previously cleared devices (known as predicate devices). However, healthcare professionals have raised concerns that applicant devices cleared this way may be more prone to recalls, which can result in substantial patient harm (if patients are exposed to such recalled devices) and financial strain on the healthcare system. In response, this work introduces a data-driven IT artifact for predicting medical device recalls, aiming to alleviate these safety concerns. In addition to the characteristics of applicant devices themselves, our predictive model leverages the characteristics of the network formed by predicate device citation relationships (predicate network). It incorporates various deep learning techniques to tackle three predictive model design challenges, including learning the predicate network structure, capturing the temporal patterns of predicate network characteristics, and accounting for the dependencies across the predicate citation history. We show with 45,398 medical devices cleared between 2003 to 2020 that our approach substantially improves the recall prediction accuracy and timeliness relative to existing state-of-the-art approaches. The improved recall prediction performance and the insights into the performance variations across device categories contribute to the literature in health information systems for societal good, as they provide opportunities for preemptive actions to potential recalls and for improving the safety of devices cleared through the 510(k) pathway.
Zhu Y, Sen S, Wu H. Hope or Hazard? The Impact of AI on Medical Device Safety.
Abstract: The Food and Drug Administration (FDA) cleared an increasing number of artificial intelligence and machine learning-enabled medical devices (referred to as AI devices), which utilize various AI technologies (e.g., image recognition) to enhance healthcare delivery. Despite the rapid growth of AI usage, no consensus has been reached among scholars and practitioners on whether AI is a hope or a hazard to medical device safety. We aim to investigate the impact of incorporating AI in medical devices on medical device safety, using device recall rates (a critical indicator for device safety) as the measurement. Relying on public data from the FDA, we employ a difference-in-differences approach, coupled with propensity score matching, to empirically analyze such an impact. Specifically, we conduct two comparisons: a cross-sectional comparison between AI and non-AI devices and a temporal comparison between AI devices and their previous non-AI device generations. Our results show that AI reduces the device recall rate by 4.4 to 10.7 percentage points for recall windows one to five years, respectively. Moreover, the lower recall rates can be partially explained by the longer review time for AI devices’ clearance and the fact that AI provides more benefits for radiology and cardiovascular devices. These findings highlight the importance of exploring additional AI opportunities in medical devices, conducting more comprehensive reviews for device clearance, and targeting specific medical specialties for AI usage to enhance the safety of medical devices.
Zhu Y, Bi X, Adomavicious G, Curley S. Predicting Goal-Directed Exercise Outcomes Leveraging Temporal and Multivariate Data Dependencies
Abstract: One-third of individuals worldwide lack regular exercise, a major contributor to serious chronic diseases. This research introduces an advanced approach for improving the prediction of individuals’ exercise outcomes. The improved prediction can facilitate personalized recommendations for exercise goal-setting and goal-striving – two critical strategies for attaining exercise goals and adhering to regular exercise according to goal theories. For example, one can (1) adjust individuals’ exercise goals to be challenging but attainable based on the prediction insights, facilitating better exercise outcomes; and (2) set up regulation plans to constantly motivate predicted inactive individuals to strive for their goals. The proposed approach is driven by design considerations that leverage the temporal and multivariate data dependencies. It uses multi-head self-attention to capture temporal dependencies of historical exercise data and uses convolution to learn multivariate interdependencies across different exercise types and contextual events. Relying on simulated data and real-world exercise records of 18,501 users on a major digital exercise platform, we show that the proposed approach substantially outperforms state-of-the-art methods, particularly under challenging conditions such as noisy input. The proposed approach advances research in predictive health information systems and goal theories with its structured designs and advantageous performance. It also provides practical value for setting and striving for exercise goals and supports sustainable user engagement on profit-driven wellness platforms that host goal-directed exercise programs.
Zhu Y, Mani A. Complex Contagion of Churn in Wellness Programs: Evidence from an Online Exercise Platform
Abstract: We study the social contagion of negative wellness activities, focusing on the churn in digital wellness programs. Leveraging a popular digital running program, running field data, and a large social network on a global online fitness platform, we (i) explore the diffusion pattern of running program churn, (ii) investigate the social contagion of churn, and (iii) analyze the heterogeneity of churn contagion. We show that churn diffuses from the network’s peripheral individuals with fewer peers to central individuals with more peers, and such churns are socially contagious. We find that the contagion of churn is a complex contagion, which explains the diffusion pattern. In addition, churn contagion varies by individual characteristics and is confined mainly to sparse network communities. These results suggest that interdependent intervention strategies based on peer connections, network structures, and individual characteristics can effectively prevent wellness program churns, improve individuals’ adherence to wellness activities, and promote individuals’ wellness.
Zhu Y, Chan J, Bi X, Guo Y, Wu J. More Is Not Always Better: The Operational Risks of Puffery in High-Intensity Advertising Environments
Abstract: We investigate how the intensity of puffery (a popular advertising tactic that attracts consumers with vague, subjective, comparative, and/or exaggerated claims) affects consumer responses to the advertised product and provide three practical intensity-based recommendations for deploying puffery. Understanding the intensity-based puffery strategies (i.e., the number of puffery advertisements shown to a unit number of consumers) is important, as research shows that excessive advertising intensity can lead to fatigue or irritation; however, how similar effects generalize to puffery remains unknown. Our study leverages an advertising campaign for a cellular network’s generational upgrade, utilizing real-world datasets of consumer service calls about the cellular network and outdoor image advertisements containing puffery claims that promote the upgrade. We investigate how puffery affects consumer sentiments in service calls using a continuous difference-in-differences model, which leverages the varied puffery advertising intensities across different geographic locations and the comparison of consumer sentiments during the pre- and post-upgrade periods. Our results reveal that higher puffery intensity can lead to more negative consumer sentiments, and this is not due to overexposure to the advertisement. Drawn on this insights, we conduct additional analyses to recommend that, in order to mitigate negative consumer response in practice, firms should keep their puffery deployment intensity at or below 4.43 advertisements per 100,000 people, apply moderate intensity when deploying puffery involving claims that are more likely for consumers to perceive as credible and compare with reality, and be cautious on increasing the puffery intensity when targeting consumers with higher income and education levels. These findings highlight a crucial yet often overlooked aspect of puffery advertising strategy: beyond advertising content, it is essential to effectively leverage advertising intensity and consider its balance with consumer perception of advertising claims to ensure a successful puffery campaign.