Early identification of ICU patients at risk of complications: Regularization based on robustness and stability of explanations

The aim of this study is to build machine learning models to predict severe complications using administrative and clinical elements that are collected immediately after patient admission to the intensive care unit (ICU). Risk models are of increasing importance in the ICU setting. However, they generally present the black-box issue because they do not provide meaningful information about the logic involved in patient-specific predictions. Fortunately, effective algorithms exist for explaining black-box models, and in practice, they offer valuable explanations for model predictions. These explanations are becoming essential to engender trust and accreditation to the model. However, once the model is implemented, a major issue is whether it will continue to employ the same prediction logic as originally intended to. To build our models, features are obtained from patient administrative data, laboratory results and vital signs available within the first hour after ICU admission. This enables our models to provide great anticipation because complications can occur at any moment during ICU stay. To build models that continue to work as originally designed we first propose to measure (i) how the provided explanations vary for different inputs (that is, robustness), and (ii) how the provided explanations change with models built from different patient sub-populations (that is, stability). Second, we employ these measures as regularization terms that are coupled with a feature selection procedure such that the final model provides predictions with more robust and stable explanations. Experiments were conducted on a dataset containing 6000 ICU admissions of 5474 patients. Results obtained on an external validation cohort of 1069 patients with 1086 ICU admissions showed that selecting features based on robustness led to gains in terms of predictive power that varied from 6.8% to 9.4%, whereas selecting features based on stability led to gains that varied from 7.2% to 11.5%, depending on the target complication. Our results are of practical importance as our models predict complications with great anticipation, thus facilitating timely and protective interventions.