Presentation
DH6 - Bridging the Gap: The Impact of Multimodal AI on Clinician Trust and Acceptance in Ophthalmic Care
SessionPoster Session 1
DescriptionIn this study, we explored the integration of artificial intelligence (AI) in healthcare, specifically in ophthalmology, with a focus on the differences between unimodal and multimodal AI systems and their impact on clinician trust and user acceptance. AI has shown potential in the diagnosis and management of complex eye diseases such as glaucoma, a leading cause of irreversible blindness globally (Sun et al., 2022). These AI systems rely on either a single type of diagnostic data, termed as unimodal (Fu et al., 2018), or combine multiple data sources, such as fundus imaging, visual field tests, and patient histories, which is called a multimodal system (Lim et al., 2022). These AI systems aim to improve diagnostic accuracy and support clinical decision-making. However, the effectiveness of these systems is not determined solely by their technical accuracy, but also by how well they align with human mental models and how much trust they inspire in the users.
Trust and user acceptance are critical factors in the successful adoption of AI systems, especially in healthcare where errors can have severe consequences. Trust in AI systems is influenced by factors such as input data, information and system quality, accuracy and transparency (Cai et al., 2019; Lukyanenko et al., 2022; Miller, 2019). It is also necessary to keep an appropriate balance of trust; insufficient trust leads to underutilization, while over-reliance can result in misuse or overconfidence in AI recommendations (Parasuraman & Riley, 1997). User acceptance refers to the degree to which users are willing to adopt and use a new system.
We developed and evaluated both unimodal and multimodal AI mockup systems designed for glaucoma diagnosis and interviewed 20 optometrists to assess their trust in each system across five key dimensions: input data modality, information quality, technical accuracy, output explainability, and system quality. We further used a modified version of a unified theory of acceptance and use of technology model to evaluate user acceptance and predict the likelihood of successful adoption of unimodal and multimodal systems by examining factors such as modality preference, performance expectancy, effort expectancy, social influence, and facilitating conditions,
The results showed a clear preference for multimodal systems, which received an average trust rating of 2.0 compared to 2.9 for unimodal systems, with a statistically significant difference (p < .001) on the Wilcoxon signed-rank test. Across all five dimensions, the multimodal system consistently outperformed the unimodal system, demonstrating that optometrists trust the multimodal system more than the unimodal system. Similarly, the user acceptance prefers a multimodal system with an average rating of 1.8 compared to 2.5 with a statistically significant difference of P<0.001 on the Wilcoxon signed-rank test. Across all dimensions of the UTAUT model, the multimodal has higher acceptance than the unimodal, which shows that users would accept the multimodal system more than the unimodal system for glaucoma diagnosis.
In terms of diagnostic performance, optometrists achieved higher accuracy metrics with the multimodal system. The multimodal AI system demonstrated superior precision (71.43%), recall (88.71%), and F1 score (79.14%), while the unimodal system produced lower results, with precision at 60%, recall at 67.92%, and F1 score at 63.72%. These findings suggest that the increased trust in multimodal systems directly correlates with improved diagnostic outcomes, making these systems more effective in clinical practice.
Key takeaway points:
▪Aligning AI systems with clinicians' mental models and workflows is critical for successful collaboration and decision-making in healthcare environments.
▪ Trust in AI systems is essential for their effective implementation in healthcare. Multimodal AI systems inspire greater user trust and have higher user acceptance due to their use of diverse data sources and their better alignment with clinicians' decision-making processes.
▪The study shows a statistically significant higher trust and higher acceptance in multimodal systems compared to unimodal systems), and this trust and user acceptance is strongly associated with better diagnostic performance.
▪The multimodal system had higher diagnostic accuracy, precision, recall, and F1 score compared to the unimodal system, indicating that combining multiple data modalities improves not only trust but also clinical outcomes.
These findings provide valuable insights for the design and deployment of AI systems in healthcare, especially for conditions like glaucoma, where precision is crucial for preventing irreversible vision loss. The results emphasize the importance of considering trust and system design when implementing AI technologies in clinical practice.
References
Cai, C. J., Winter, S., Steiner, D., Wilcox, L., & Terry, M. (2019). “Hello AI”: Uncovering the Onboarding Needs of Medical Practitioners for Human-AI Collaborative Decision-Making. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–24. https://doi.org/10.1145/3359206
Fu, H., Cheng, J., Xu, Y., Zhang, C., Wong, D. W. K., Liu, J., & Cao, X. (2018). Disc-Aware Ensemble Network for Glaucoma Screening From Fundus Image. IEEE Transactions on Medical Imaging, 37(11), 2493–2501. https://doi.org/10.1109/TMI.2018.2837012
Lim, W. S., Ho, H.-Y., Ho, H.-C., Chen, Y.-W., Lee, C.-K., Chen, P.-J., Lai, F., Jang, J.-S. R., & Ko, M.-L. (2022). Use of multimodal dataset in AI for detecting glaucoma based on fundus photographs assessed with OCT: focus group study on high prevalence of myopia. BMC Medical Imaging, 22(1), 206. https://doi.org/10.1186/s12880-022-00933-z
Lukyanenko, R., Maass, W., & Storey, V. C. (2022). Trust in artificial intelligence: From a Foundational Trust Framework to emerging research opportunities. Electronic Markets, 32(4), 1993–2020. https://doi.org/10.1007/s12525-022-00605-4
Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267, 1–38. https://doi.org/10.1016/j.artint.2018.07.007
Parasuraman, R., & Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors: The Journal of the Human Factors and Ergonomics Society, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
Sun, Y., Chen, A., Zou, M., Zhang, Y., Jin, L., Li, Y., Zheng, D., Jin, G., & Congdon, N. (2022). Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the Global Burden of Disease Study 2017. BMJ Open, 12(1), e053805. https://doi.org/10.1136/bmjopen-2021-053805
Trust and user acceptance are critical factors in the successful adoption of AI systems, especially in healthcare where errors can have severe consequences. Trust in AI systems is influenced by factors such as input data, information and system quality, accuracy and transparency (Cai et al., 2019; Lukyanenko et al., 2022; Miller, 2019). It is also necessary to keep an appropriate balance of trust; insufficient trust leads to underutilization, while over-reliance can result in misuse or overconfidence in AI recommendations (Parasuraman & Riley, 1997). User acceptance refers to the degree to which users are willing to adopt and use a new system.
We developed and evaluated both unimodal and multimodal AI mockup systems designed for glaucoma diagnosis and interviewed 20 optometrists to assess their trust in each system across five key dimensions: input data modality, information quality, technical accuracy, output explainability, and system quality. We further used a modified version of a unified theory of acceptance and use of technology model to evaluate user acceptance and predict the likelihood of successful adoption of unimodal and multimodal systems by examining factors such as modality preference, performance expectancy, effort expectancy, social influence, and facilitating conditions,
The results showed a clear preference for multimodal systems, which received an average trust rating of 2.0 compared to 2.9 for unimodal systems, with a statistically significant difference (p < .001) on the Wilcoxon signed-rank test. Across all five dimensions, the multimodal system consistently outperformed the unimodal system, demonstrating that optometrists trust the multimodal system more than the unimodal system. Similarly, the user acceptance prefers a multimodal system with an average rating of 1.8 compared to 2.5 with a statistically significant difference of P<0.001 on the Wilcoxon signed-rank test. Across all dimensions of the UTAUT model, the multimodal has higher acceptance than the unimodal, which shows that users would accept the multimodal system more than the unimodal system for glaucoma diagnosis.
In terms of diagnostic performance, optometrists achieved higher accuracy metrics with the multimodal system. The multimodal AI system demonstrated superior precision (71.43%), recall (88.71%), and F1 score (79.14%), while the unimodal system produced lower results, with precision at 60%, recall at 67.92%, and F1 score at 63.72%. These findings suggest that the increased trust in multimodal systems directly correlates with improved diagnostic outcomes, making these systems more effective in clinical practice.
Key takeaway points:
▪Aligning AI systems with clinicians' mental models and workflows is critical for successful collaboration and decision-making in healthcare environments.
▪ Trust in AI systems is essential for their effective implementation in healthcare. Multimodal AI systems inspire greater user trust and have higher user acceptance due to their use of diverse data sources and their better alignment with clinicians' decision-making processes.
▪The study shows a statistically significant higher trust and higher acceptance in multimodal systems compared to unimodal systems), and this trust and user acceptance is strongly associated with better diagnostic performance.
▪The multimodal system had higher diagnostic accuracy, precision, recall, and F1 score compared to the unimodal system, indicating that combining multiple data modalities improves not only trust but also clinical outcomes.
These findings provide valuable insights for the design and deployment of AI systems in healthcare, especially for conditions like glaucoma, where precision is crucial for preventing irreversible vision loss. The results emphasize the importance of considering trust and system design when implementing AI technologies in clinical practice.
References
Cai, C. J., Winter, S., Steiner, D., Wilcox, L., & Terry, M. (2019). “Hello AI”: Uncovering the Onboarding Needs of Medical Practitioners for Human-AI Collaborative Decision-Making. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–24. https://doi.org/10.1145/3359206
Fu, H., Cheng, J., Xu, Y., Zhang, C., Wong, D. W. K., Liu, J., & Cao, X. (2018). Disc-Aware Ensemble Network for Glaucoma Screening From Fundus Image. IEEE Transactions on Medical Imaging, 37(11), 2493–2501. https://doi.org/10.1109/TMI.2018.2837012
Lim, W. S., Ho, H.-Y., Ho, H.-C., Chen, Y.-W., Lee, C.-K., Chen, P.-J., Lai, F., Jang, J.-S. R., & Ko, M.-L. (2022). Use of multimodal dataset in AI for detecting glaucoma based on fundus photographs assessed with OCT: focus group study on high prevalence of myopia. BMC Medical Imaging, 22(1), 206. https://doi.org/10.1186/s12880-022-00933-z
Lukyanenko, R., Maass, W., & Storey, V. C. (2022). Trust in artificial intelligence: From a Foundational Trust Framework to emerging research opportunities. Electronic Markets, 32(4), 1993–2020. https://doi.org/10.1007/s12525-022-00605-4
Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267, 1–38. https://doi.org/10.1016/j.artint.2018.07.007
Parasuraman, R., & Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors: The Journal of the Human Factors and Ergonomics Society, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
Sun, Y., Chen, A., Zou, M., Zhang, Y., Jin, L., Li, Y., Zheng, D., Jin, G., & Congdon, N. (2022). Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the Global Burden of Disease Study 2017. BMJ Open, 12(1), e053805. https://doi.org/10.1136/bmjopen-2021-053805
Event Type
Poster Presentation
TimeMonday, March 314:45pm - 6:15pm EDT
LocationFrontenac Foyer
Digital Health (DH)
Simulation and Education (SE)
Hospital Environments (HE)
Medical and Drug Delivery Devices (MDD)
Patient Safety and Research Initiatives (PS)