Please note: this event has passed
Title: Quantifying Accuracy - Interpretability Trade Off for Prediction Models in Health Research
Abstract: In this talk, Diana discusses the balance between model interpretability and its accuracy and flexibility. Undoubtfully, transparency is important: for researchers, transparency will help to get an insight into the predictors-outcomes relationships, while for clinicians and patients, the interpretability is critical while incorporating model predictions into medical decision making. However, transparency is often an attribute of the simpler models, and opting out from more complex models may compromise their prediction performance.
Here, a practical approach to quantifying such a trade off will be presented, developed in collaboration with Prof. Daniel Stahl (KCL), Dr Angus Roberts (KCL), and Dr Daniel Stamate (Goldsmith University). We consider linear models such as regularised regressions or Cox Proportionate Hazards model as intrinsically transparent, and our baseline; and a flexible "black box" machine learning (ML) algorithm such as XGBoost or Survival Random Forest, serves as an alternative. We then test models' relative performance using a repeated nested cross-validation. The difference in the performance will quantify the prediction accuracy one could gain by opting to a less transparent model, or the accuracy - interpretability trade off.
We will present the results of such analysis for several health datasets, and demonstrate how that was performed in our recently developed R package 'survcompare' (a pilot version is already available at CRAN, https://cran.rstudio.com/web/packages/survcompare/index.html, https://github.com/dianashams/survcompare).
Biography: Diana Shamsutdinova is a Researcher at Biostatistics and Health Informatics Department at IoPPN, King's College London. She completed her PhD in 2023, and continues working at the department focusing on Prediction Modelling, Survival Analysis and Explainable AI, as well as on applications of such theories to Mental and Physical Health outcomes. Prior to that, she studied Mathematics at Moscow State University, and Neuroscience and Psychology of Mental Health at King's College London.