Visualizing uncertainties of drug response predictions and their explanations in a clinical context
Requirements
-
Required: Experience in Web development (html/css/javascript, python)
- Preferred: Completion of the lectures on "Human-Computer Interaction" or "Data Visualization", the seminar on “Interactive Intelligent Systems” and the lecture on "Wissenschaftliches Arbeiten in der Informatik"
Contents
Description
Uncertainty of predictions are a key factor to assess the reasons for drug response prediction models not working. In a clinical context, interactive visualization tools including GUIs are a key element to communicate results to medical doctors and enhance the acceptance of results of computational methods.
The goal of this project is to program an visualization platform that interactively represents drug response predictions of multiple drugs and entities (cell lines) along with their uncertainty and explanations of them. Meta-data should be taken into account, and the similarity of a new patient or drug with existing data points should be determined, visualized and considered for highlighting relevant subsets. Potentially, tools such as biclustering could be used to map input data and output results into the same space.
There are many different ways to define, compute, and represent uncertainty, ranging from icon arrays and error bars to more complex visualizations (Bhatt et al., 2021; Yu et al., 2022). Although uncertainty representations have been and are often used, for example, to communicate risk in the medical field, users (i.e., physicians) do not necessarily perform better when provided with uncertainty representations. This may be due to how and in what format uncertainty is represented and communicated (Cao et al. 2024). Therefore, when designing an application with additional information such as uncertainty and other aspects, we need to consider what the user needs and how the user may interpret specific visualizations, i.e., human-centered design.
References
- Park, A., Lee, Y., & Nam, S. (2023). A performance evaluation of drug response prediction models for individual drugs. Scientific Reports, 13, Article 11911. https://doi.org/10.1038/s41598-023-39179-2 - drug response prediction example paper
- Iversen, P., Witzke, S., Baum, K., & Renard, B. Y. (2024). Identifying drivers of predictive aleatoric uncertainty. arXiv preprint. https://doi.org/10.48550/arXiv.2312.07252 - Baum's research group paper on explaining uncertainties
- Bishop, C. M. (1994). Mixture density networks. NCRG Technical Report, Aston University. https://publications.aston.ac.uk/373/1/NCRG_94_004.pdf - mixture density networks as a way to predict uncertainty
- Covert, I., & Lee, K. (2021). Improving KernelSHAP: Practical Shapley value estimation using linear regression. In Proceedings of the 38th International Conference on Machine Learning (Vol. 130). https://proceedings.mlr.press/v130/covert21a.html - kernelSHAP as feature attribution method
- Bhatt, U., Antorán, J., Zhang, Y., Liao, Q. V., Sattigeri, P., Fogliato, R., ... & Xiang, A. (2021, July). Uncertainty as a form of transparency: Measuring, communicating, and using uncertainty. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (pp. 401-413).
- Yu, J., Wang, D., & Zheng, M. (2022). Uncertainty quantification: Can we trust artificial intelligence in drug discovery? iScience, 25(8).
- Cao, S., Liu, A., & Huang, C. M. (2024). Designing for Appropriate Reliance: The Roles of AI Uncertainty Presentation, Initial User Decision, and User Demographics in AI-Assisted Decision-Making. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1), 1-32.