Modernizing Pressure Injury Risk Assessment in the ICU in the COVID Era: Ensemble Super-Learning and Explainable AI
Author(s)
Zhou Y1, Reis L2, Alderden J3, Zhang Y4, Wilson A5
1Berkeley, Berkeley, CA, USA, 2Parexel, Stockholm, Sweden, 3University of Utah, Salt Lake City, UT, USA, 4Parexel, Halethorpe, MD, USA, 5Parexel International, Waltham, MA, USA
Presentation Documents
OBJECTIVES: Hospital-acquired pressure injuries (HAPrI) are areas of injury to the skin and/or underlying tissues. Risk stratification is essential for guiding prevention in the ICU, but current risk assessment tools require labor-intensive input. This motivates a tactical, parsimonious, and automatic risk profiling algorithm, that can be based on readily available clinical measures (e.g., COVID status, race, Medicare/Medicaid status). Additionally, International Pressure Injury Prevention guidelines call for the development of machine learning-based risk assessment algorithms that are clinician-interpretable and context-informed.
METHODS:
Adult patients admitted to one of two ICUs between April 2020, and April 2021 were eligible for inclusion. Discrete and ensemble super-learning models, adjusting for class imbalance, were created from a rich library of candidate base learners. For explainability, SHAP (SHapley Additive exPlanations) global and local values were derived to help explain variable average marginal contributions (across all permutations) to the model. An iteration of clinical expert review was performed with the SHAP values, and simulations of patient profiles and results were used to reformat and re-weight predictor variables. All analysis was run in open Python (version 3.7), and code/results will be made available via a GitHub page.RESULTS:
The final sample consisted of 1,911 patients (removing 9 with missing pressure injury status). Hospital-acquired pressure injuries (defined as stage 2, or worse) occurred in 18.5% of the sample (n=354). We achieved the best overall performance on the testing data with a stacked ensemble using three base models: random forest (rf), gradient boosted machine (gbm), and neural network (NN) (Performance on 20% holdout: Accuracy: 81%; AUC: 0.77; AUCPR: 0.53).CONCLUSIONS: Prediction engineering should be done in collaboration with c
linical experts to optimize tactical implementation to both optimize performance, with minimal interruption to workflow. XAI enhanced adoption of the experts’ advice based on the selected model features.Conference/Value in Health Info
Value in Health, Volume 25, Issue 12S (December 2022)
Code
MSR17
Topic
Methodological & Statistical Research, Study Approaches
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Confounding, Selection Bias Correction, Causal Inference, Electronic Medical & Health Records
Disease
No Additional Disease & Conditions/Specialized Treatment Areas