Prediction of pain intensity in the intensive care unit: Performance evaluation of an Artificial Intelligence model

Abstract

In practice, the rate of uncontrolled pain remains very high in the Intensive Care Unit. This is due to the fact that pain management remains difficult in the Intensive Care Unit. The multitude of pain etiologies makes the assessment of its intensity a real challenge. Thus, proper pain management requires, above all, meticulous analysis of these variables. Unfortunately, processing, exploring and interpreting these gigantic datasets requires not only time, but also patience. Fortunately for us, technology is here to make life easier.We conducted a prospective observational study in the central Intensive Care Unit at The Main Military Hospital of Tunis. We included 37 patients, from whom 1012 samples were collected. Data was collected on demographics, clinical data and vital signs. This data was used to train a Machine Learning model capable of predicting pain intensity in a personalized, autonomous and continuous manner.Our Machine Learning model achieved an Area Under Receiver Operating Characteristic Curve score of 0.87, an Area Under Precision-Recall Curve score of 0.63. More specifically, the classifier dedicated to the 1st pain level remained the best performer. Subsequently, the 3rd pain level classifier performed fairly well. Finally, the worst-performing classifier was the 2nd tier. Among the features, Richmond Agitation Sedation Scale score with a Mean Decrease in Impurity at 0.2, blood pressure at the time of pain and age were the most important in predicting pain intensity. Length of hospital stay, vital signs before 1 min of assessment and at rest, weight and tobacco consumption were fairly important in pain assessment.We can conclude that vital signs are indeed objective in pain monitoring. This is supported by the performance of our Machine Learning model, despite the scarcity of samples. Further development of this Artificial Intelligence model will assist practitioners in their decision-making and provide a framework for the creation of intelligent medical devices.