OBJECTIVES: Traditional causal methods typically rely on parametric statistical models that impose restrictive assumptions about underlying data structure. Recent advancements in targeted machine learning (ML) and super learning enable the identification of causal estimates in real-world data, irrespective of data complexity or structure, offering a flexible and comprehensive approach to causal inference. In our study, we leverage the MIMIC IV database and ML models to ascertain the causes of hospital-acquired pressure injuries (HAPrI) and develop a risk prediction algorithm.

METHODS: Utilizing the MIMIC IV dataset – a deidentified electronic health records dataset from Beth Israel Deaconess Medical Center, capturing admissions from 2008-2019 for nearly 300,000 patients – we used cost-sensitive ensemble super learning to predict HAPrI in the ICU. We then estimated the potential causal effect of albumin on HAPrI development via clinically-informed debiasing methods, including directed acyclic graphs and targeted maximum likelihood estimation (TMLE).

RESULTS: Of 28,395 eligible cases, 1,395 developed a pressure injury (4.9%). The ensemble super learner had a cross-validated AUC of 0.8, with 45.6% sensitivity and 88.8% specificity. The crude odds ratio of low (below 3.0) albumin on pressure injury was significant: OR = 2.86, p<0.0001. The TMLE-adjusted estimate was significant but attenuated: OR = 2.22, p<0.0001.

CONCLUSIONS: Previous models predicting pressure injuries often favour overall accuracy and have a clinically uninformative sensitivity, whereas traditional Braden scales classify almost all patients as high-risk. Our results suggest that ML methods can be used to develop accurate risk prediction algorithms for HAPrI. We also identified a significant (causal) effect of low albumin levels on the development of pressure injuries. These findings demonstrate how ML methods generate valuable causal and predictive models and improve our understanding of data interdependencies. The future of clinical prediction lies at the intersection of accuracy, clinical wisdom, and machine learning's amplified use of real-world data.