Evaluation of pain in intensive care patients with a pain diagnostic system that measuring facial movement

Standard scales used in pain assessment of patients who are conscious and cannot communicate verbally in intensive care are limited. Automatic, standardized, continuous, unbiased and scalable pain scales should be developed for these patients. The study was conducted to evaluate pain in intensive care patients who are conscious and unable to communicate verbally, using a pain diagnosis system that measures facial movements. The sample of the study consisted of patients who were followed up in Çukurova University Faculty of Medicine Balcalı Hospital General Surgery Intensive Care Unit, Neurology Intensive Care Unit, Internal Medicine Intensive Care Unit and Adana City Training and Research Hospital General Surgery Intensive Care Unit and Brain Surgery Intensive Care Unit between April 2023 and October 2024 and who met the inclusion criteria. The sample size of the study was calculated as 46 people to be reached in the two-tailed hypothesis with 95% power, type 1 error, 0.05 reference, and 0.50 correlation coefficient. Patient Identification Form, Numerical Pain Scale (NPS), Wong Baker Faces Pain Scale (WFPS), Patient Follow-up Form, Facial Action Coding System (FACS), Video Camera, Camera Tripod were used to collect data. Pain assessment was made by the patient with FACS, by the nurses with FACS, and by the computer with WFPS. Data were collected twice a day, in the morning and evening, for 2 days, immediately before, during, and 20 minutes after painful interventions. Pain was assessed by four nurses using photographs taken from video recordings of the patients. Statistical analyses of the data were performed using the SPSS (IBM SPSS Statistics 27) package program. In the analysis of the patient's Facial Action Coding System, machine learning algorithms; EfficientNet_B0, ResNet50 and ConvNeXt-tiny deep learning models were used. In the comparison of patient and machine measurements, among the models; in Class 4, ConvNeXt-tiny achieved 99% success on the training set and 57.9% on the test data, and in Class 6, ResNet50 achieved 98% success on the training data set and 52.9% on the test data. A positive and statistically significant relationship was found between the mean pain levels assessed by the patient and four nurses (p<0.05). It was determined that the patients' own pain report and the machine measurements were more consistent than the nurses' estimates, and this consistency was significant (p<0.001). Our study was conducted to improve the pain assessment of conscious patients who cannot express themselves verbally in intensive care. Objective pain assessment scales to be developed by utilizing the accuracy of machine learning can reduce subjectivity and decrease nurses' observational differences in pain assessment. In this context, the findings show that machine learning-based pain assessment systems are more compatible with the patient's own pain report and nurse assessment, and can produce more objective and consistent results.