a comparison and analysis of explainable clinical decision making using white box and black box models
abstract
explainability is a crucial element of machine learning-based making in high stake
scenarios such as risk assessment in criminal justice [80], climate modeling [79], disaster
response [82], education [81] and critical care. there currently exists a performance
tradeoff between low-complexity machine learning models capable of making predictions
that are inherently interpretable (white box) to a human, and cutting-edge high
complexity (black box) models are not readily interpretable.
in this thesis we first aim to assess the reliability of the predictions made by black box
models. we train a series of machine learning models on an icu (intensive care unit)
outcome prediction task on the mimic iii dataset. we perform a comparison of the
predictions made by white box models and their black box counterparts by contrasting
explainable model feature coefficients/importances to feature importance values generated
by a post-hoc shap (shapley additive explanation) values. we then validate our results
with a panel of clinical experts. the first study shows that both black box and white box
models prioritize clinically relevant variables when making outcome predictions. higher
performing models showed prioritizations to more clinically relevant variables than lower
performing models. the black box models show better overall performance than the white
box models. [...]