Using nursing data for machine learning-based prediction modeling in intensive care units: A scoping review

Abstract

Background: Nursing data can help detect patient deterioration early and predict patient outcomes. Moreover, rapid advancements in machine learning have highlighted the need for clinical prognosis prediction models for intensive care unit patients. Although prediction models that incorporate nursing data generated during the care of critically ill patients are increasing, a comprehensive understanding of the specific types of nursing data utilized and these models to predict health outcomes has not yet been achieved.

Objective: This scoping review aimed to identify the current state of research on machine learning-based models that utilize nursing data to predict health outcomes of intensive care unit patients, focusing on the types of nursing data in these models.

Methods: This scoping review was conducted with a systematic literature search until December 2023 across seven databases. Literature that utilized machine learning using nursing data to predict the prognosis of adult patients hospitalized in the intensive care unit was included. Data were organized into the study, model-related, and nursing data characteristics.

Results: A total of 151 studies were included, which were published between 2004 and 2023, with an upward trend since 2018. More than half of the studies developed prediction models using open access data, with Medical Information Mart for Intensive Care data being the most frequently used. Most studies employed supervised learning, followed by deep learning and neural networks, while other methods were rarely used. Among supervised learning techniques, regression was the most commonly used, followed by boosting and random forests. Nursing-sensitive outcomes (13.0 %) were chosen less frequently than clinical ones (87.0 %) in prediction models. In this review, nursing data were classified into nursing scales (n = 150), nursing assessment records (n = 83), nursing activity records (n = 13), and nursing notes (n = 23), with nursing scales being the most frequent. Nursing scales and notes exhibited an increasing trend recently.

Conclusions: This scoping review identified the various utilization of nursing data in models to predict the prognoses of critically ill patients. Overall, nursing scales, structured data that objectively show specific health conditions of patients, were the most utilized. As other types of nursing data also have the potential to predict patients' clinical prognoses, future research should explore the development of prediction models incorporating various nursing data. These findings may contribute to providing insights into the use of nursing data and could aid healthcare providers and researchers aiming to develop prediction models related to clinical prognoses in the intensive care unit setting. Social media abstract: This scoping review identified the various utilization of nursing data in models to predict the prognoses of critically ill patients.