Automatic pain assessment from facial action units in ICU patients via various machine learning models.

Pain represents a critical vital sign monitored in intensive care unit (ICU) patients. The facial action coding system (FACS) defines facial action units (AUs) and provides a structured framework for pain recognition. Currently, the most broadly used pain assessment method is the pain intensity scale developed by Prkachin and Solomon (PSPI), which relies on predefined AUs to quantify facial expressions. However, due to the influence of underlying diseases and facial texture variations in ICU patients, AUs can fail to transfer to clinical settings accurately. To address this problem, this study uses video sequences of pain states collected from 61 ICU patients under resting, daily, and procedural conditions by using an advanced AU detection system. By evaluating the AUs with statistical features and various classification models, this study identifies six key AUs that outperform the PSPI's predefined AUs in terms of accuracy, precision, recall, and F1-score metrics. Further, this study explores the performance of various temporal self-learning networks in the pain assessment task, thus further validating the effectiveness of the identified AU combination. The results presented in this study demonstrate that using AU dynamic learning in combination with deep temporal analysis can improve the reliability of clinical pain assessment. Finally, this study offers a promising approach for automated pain monitoring systems in ICU settings.