Lena Stempfle: Interpretable machine learning for prediction with missing values at test time

Missing values are inevitable in real-world healthcare data, yet common strategies like imputation or missingness indicators often undermine interpretability. In this talk, I present methods for building interpretable models that remain reliable under test-time missingness. I will introduce Sharing Pattern Submodels, a way to leverage missingness patterns while retaining generalization; MINTY, a rule-based model that avoids imputation through logical substitutions; and a broader framework that adapts trees, sparse linear models, and ensembles to minimize reliance on missing features. Supported by insights from a clinician survey, these approaches aim to make machine learning both robust to missingness and transparent enough for clinical decision support.

Lena Stempfle is a postdoctoral associated with MIT healthy ML group led by Marzyeh Ghassemi with WASP International Postdoctoral Scholarship. She completed her PhD in Computer Science at Chalmers University of Technology in Sweden. Her research focuses on the intersection of machine learning and healthcare, with a particular interest in predictions with missing values at test time, time series, and causality. Lena’s aims to develop interpretable and accurate models to support clinical decision-making.