Clinically Informed, Cost-Sensitive Machine Learning for Predicting Hospital-Acquired Pressure Injuries: A Practical Approach Using Routinely Collected Data
Author(s)
Alexandra Pasi, PhD1, Andy Wilson, MS, PhD2, Meghann Gregg, PhD3, Jenny Alderden, PhD, APRN4, Dirk Tolson III, B.Sc5, Alan Mullenix, PhD1.
1Lucidity Sciences, Salt Lake City, UT, USA, 2PAREXEL, Waltham, MA, USA, 3Parexel, Austin, TX, USA, 4University of Utah, Boise, ID, USA, 5UC Berkeley, Berkeley, CA, USA.
1Lucidity Sciences, Salt Lake City, UT, USA, 2PAREXEL, Waltham, MA, USA, 3Parexel, Austin, TX, USA, 4University of Utah, Boise, ID, USA, 5UC Berkeley, Berkeley, CA, USA.
OBJECTIVES: Hospital-acquired pressure injuries (HAPIs) have a major impact on patient outcomes in intensive care units (ICUs), and effective prevention relies on early and accurate risk-prediction. Existing tools for prediction fail to provide adequate sensitivity and precision to actionably inform triage and care in settings where the patient population is generally high-risk. We seek to bridge this gap using a novel machine learning approach.
METHODS: We present a translational machine-learning-powered approach for predicting HAPIs that provides a breakthrough improvement in predictive precision and sensitivity using real-world clinical data from the MIMIC IV data set (version 2.2), specifically variables that are easily and frequently captured within clinical settings. Furthermore, we analyze the generalizability of the models to new clinical scenarios via assessing performance on post-COVID data using version 3.1 of MIMIC IV. Lastly, we use the resulting ML models in conjunction with ML explainability and simulation techniques to investigate possible interventions for preventing HAPIs in-hospital.
RESULTS: We evaluated the performance of a novel kernel-based machine learning approach and compared it to the performance of the prior state-of-the-art ensemble-based model (h20.ai AutoML) using the same training and hold-out validation sets. The kernel-based model offered significant improvements in both balanced accuracy (70.3% for kernel-based vs. 63.9% for AutoML) as well as in the particularly clinically relevant metrics of sensitivity (54.8% for kernel-based vs. 36.6% for AutoML) and precision (19.2% for kernel-based vs. 17.5% for AutoML). With alternate parameterizations of the kernel-approach, sensitivity can be preserved (relative to the AutoML model), while significantly increasing precision.
CONCLUSIONS: This investigation into novel machine learning approaches to clinical prediction, driven by a framework of clinical need and constraint, lays out a roadmap from data to practice that centers questions of usability, explainability, intervenability, and generalizability. These advances represent significant progress in clinical machine learning with substantial potential for improving patient outcomes.
METHODS: We present a translational machine-learning-powered approach for predicting HAPIs that provides a breakthrough improvement in predictive precision and sensitivity using real-world clinical data from the MIMIC IV data set (version 2.2), specifically variables that are easily and frequently captured within clinical settings. Furthermore, we analyze the generalizability of the models to new clinical scenarios via assessing performance on post-COVID data using version 3.1 of MIMIC IV. Lastly, we use the resulting ML models in conjunction with ML explainability and simulation techniques to investigate possible interventions for preventing HAPIs in-hospital.
RESULTS: We evaluated the performance of a novel kernel-based machine learning approach and compared it to the performance of the prior state-of-the-art ensemble-based model (h20.ai AutoML) using the same training and hold-out validation sets. The kernel-based model offered significant improvements in both balanced accuracy (70.3% for kernel-based vs. 63.9% for AutoML) as well as in the particularly clinically relevant metrics of sensitivity (54.8% for kernel-based vs. 36.6% for AutoML) and precision (19.2% for kernel-based vs. 17.5% for AutoML). With alternate parameterizations of the kernel-approach, sensitivity can be preserved (relative to the AutoML model), while significantly increasing precision.
CONCLUSIONS: This investigation into novel machine learning approaches to clinical prediction, driven by a framework of clinical need and constraint, lays out a roadmap from data to practice that centers questions of usability, explainability, intervenability, and generalizability. These advances represent significant progress in clinical machine learning with substantial potential for improving patient outcomes.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR55
Topic
Health Service Delivery & Process of Care, Medical Technologies, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Injury & Trauma, No Additional Disease & Conditions/Specialized Treatment Areas