Background: Most critical care patients are considered high risk for hospital-acquired pressure injury (HAPrI) according to industry-standard assessment tools, such as the Braden Scale. Yet, accurately assessing HAPrI risk is critical for allocating preventive resources. Machine learning (ML) techniques have been used to create algorithms that have the potential to better identify high-risk critical care patients. This improved discrimination could result in a more efficient use of limited resources. In response to this need, the National Pressure Injury Advisory Panel's 2019 guidelines called for developing ML algorithms for HAPrI risk prediction that are easily interpretable by healthcare professionals.
Objectives: This research addresses the need for accurate, efficient, and clinically understandable HAPrI risk prediction algorithms using explainable artificial intelligence (XAI) to provide clinician-friendly HAPrI risk predictions.
Methods: For this study, we used Electronic Health Record (EHR) data from adult surgical and cardiovascular surgical Intensive Care Unit (ICU) patients to develop and test ML based algorithms to predict HAPrI. Along with evaluating the effectiveness of the models, we also used model-independent interpretation techniques to present the models in a manner that is easily understandable to humans.
Results: Our final sample consisted of 5,101 patients (split into 80:20 train:test sets). The sample was predominantly male (n = 3302, 65%) and White (n = 4256, 83%). The mean age was 58 years (SD = 17), and HAPrI developed among 333 patients (6.5%). The predictions made by three ML models (extreme gradient boosted machines, random forests, and logistic regression) were similar in performance when evaluated on the test data (AUC range: 0.80-0.81; F1 range: 0.28-0.33). However, using model-agnostic interpretation methods, we observed differences in the importance of variables, individual conditional expectations, and local effect fluctuations. Thus, while the models had similar performance, they used different variables and could predict different outcomes for clinically similar patients.
Conclusions: Leveraging existing EHR data to create risk algorithms enables near-real-time risk assessment and reduces the need for redundant charting. Using XAI techniques, we gained insight into how specific features within the model affected HAPrI formation. The increased transparency will allow clinicians to gain insight into the algorithm’s decision-making at the patient level, which is essential for enhancing trust. In the future, XAi may also facilitate collaboration between clinicians and data scientists to improve algorithms through expert-augmented ML.
