Action Analysis And Early Warning System For Unplanned Extubation Of ICU Patients Based On Human Action Recognition

【Objective】To observe,describe and analyse video images of the whole process of unplanned extubation of ICU patients,to clarify the characteristic behavioural actions of the UEX process,and to construct an action-based UEX prediction model and an intelligent early warning system.To facilitate early identification of unplanned extubation in ICU patients by healthcare professionals and to take targeted preventive and control measures to prevent the occurrence of unplanned extubation.【Method】The expert meeting method was used to delineate and name the action phases of unplanned extraction.Descriptive analysis of 41 ICU patients in various stages of UEX using the video observation method.Construction of an action-based decision tree and logistic regression prediction model for UEX using a case-control study design.Intelligent early warning system for UEX using Yolo model,convolutional neural network and STformer.【Results】1.The expert positivity factor was 100% and the expert authority factor was 0.835.Through the expert meeting,the UEX was divided into three stages,namely the pre-intentional stage,the intentional extubation stage and the extubation completion stage.2.The duration of the pre-intent was 193.0s,the duration of the intentional extubation was 8.0s and the duration of the completed extubation was 6.0s,with a statistically significant difference in the duration of the three phases compared.The pre-intent phase was richer in behaviours,mainly upper limb movements(95.1%),head and neck movements(31.7%),lower limb movements(29.3%)and whole body movements(31.7%).The upper limb movements were the most diverse in terms of behavioural types.For UEX,the index finger and thumb participation are both 100%,and the middle finger participation is 92.9%.3.The action-based logistic regression model for UEX of ICU patients incorporated 9 variables as shaking the leg,touching the tube,struggling to get up,grasping the bedrail,unclamping the oxygen saturation,wiping the mouth,touching the chin,groping,and tapping the bed surface.The decision tree model had 5 levels and 18 nodes,11 classification rules were extracted and 8 categories of high risk patients were screened.The area under the ROC curve for the logistic regression model was 0.83 and the area under the ROC curve for the decision tree model was 0.78.4.The action-based intelligent early warning system for UEX of ICU patients has an accuracy of 80.28% ± 1.44%,runs in 0.00371 s and takes up 16.71 M of memory.It outperforms other deep learning models.【Conclusion】The key phase of UEX control is the pre-intent phase;the key control actions in the pre-intent phase are "groping" and "lifting";the key control finger are the index finger and thumb.The logistic regression has a better area under the ROC curve than the decision tree.Because both models have their advantages and disadvantages,we should be used in clinical practice in conjunction with both models for early identification of UEX from patients’ behaviour.The UEX warning system combining convolutional neural network and Transformer has good recognition efficiency and recognition rate.Further research will explore its practical application in ICU and provide a basis for intelligent and dynamic recognition of UEX warning.