Making Sense of Temporal Black-box Models in Biomedicine
Temporal concepts in routinely collected clinical data are critical to better understand acute pathophysiological processes. Machine learning models that leverage time-series clinical data are challenged with complex sociotechnical issues such as pre-existing biases, inconsistent sampling, and sparsity, making the already opaque model behaviors harder to interpret. In this talk, I will attempt to (1) describe guiding principles and conceptual frameworks for making sense of (i.e., explaining) decisions and behaviors of machine learning models for clinical applications, (2) identify and compare model-agnostic explanation methods, and (3) apply those explanation methods to a clinical case study in the context of time-series data.
Zoom: https://arizona.zoom.us/j/89712326534 Password: math