This study investigates whether artificial intelligence (AI)-assisted video monitoring can identify early behavioral changes that precede accidental or harmful events in Intensive Care Unit (ICU) patients. ICU patients are vulnerable to a series of sudden and potentially dangerous events-such as agitation, delirium, accidental device removal, and significant sleep disruption-many of which develop gradually and are difficult to detect solely from routine physiological monitoring. This project aims to determine whether AI analysis of continuous bedside video recordings, combined with noise-level information and vital-sign data already collected during standard ICU care, can provide clinicians with timely warnings before these events occur.
Rationale Traditional ICU monitoring systems focus on physiological parameters such as heart rate, blood pressure, and oxygen saturation. While essential, these measurements do not fully represent patient behavior. Many high-risk events are preceded by subtle motor patterns or behavioral cues-for example, repeated reaching toward tubes, rising restlessness, or disturbed sleep cycles. Such cues are often intermittent, brief, or masked by sedation or other treatments, making them difficult for staff to detect in busy clinical environments.
Computer vision and AI technologies offer an opportunity to objectively observe and interpret patient movements and behavioral trends continuously, without adding clinical workload. By integrating video information with physiologic data and environmental noise levels, the AI system may identify patterns that indicate emerging delirium, increased agitation, or imminent attempts to remove medical devices. Early identification may support timely preventive interventions and reduce the rates of adverse events.
Study Overview The study will prospectively enroll ICU patients who consent to video monitoring and data use. A small camera will be installed above each bed to continuously capture patient movement and posture. The camera view is restricted to the patient zone, excluding unnecessary areas such as the nursing station. All recordings follow strict privacy-protection procedures, including automated face masking, background blurring, and removal of identifying information from objects in the frame.
Environmental noise is recorded through a decibel meter, and routine vital-sign data are synchronized with the video timeline. These combined multimodal data will serve as input for AI model development.
The study is divided into three components:
Data collection phase - real-world continuous recording of behavioral and physiological data.
Data processing and annotation - cleaning, de-identification, and labeling of key behavioral events by trained researchers.
Model development and evaluation - training AI models to identify behavioral patterns associated with clinically meaningful events, and evaluating their predictive performance.
Data Integration and Processing
All raw videos remain stored securely inside the hospital's protected data environment and are not transferred outside. A standardized de-identification pipeline is applied before any analytical use. This includes:
Masking or replacing patient faces. Removing identifying elements such as bed numbers and equipment labels. Blurring all background areas outside the patient zone. Excluding frames containing staff faces or unrelated activities. After de-identification, videos are aligned with vital-sign and noise-level timelines to create multimodal time-series datasets. Human annotators, trained with a unified labeling guideline, identify episodes of agitation, possible delirium-related behavior, attempts at device removal, and sleep-wake transitions. These labels serve as ground truth for AI training.
AI Model Development Multiple AI architectures will be explored, particularly those suited for temporal video analysis. Potential approaches include convolutional neural networks (CNNs), 3D CNNs, long short-term memory networks (LSTM), or transformer-based models capable of learning long-range dependencies in behavior sequences. Additional feature extraction methods will be evaluated to integrate physiologic and environmental signals.
To avoid model overfitting and ensure generalizability, the dataset will be split into training, validation, and independent test sets. Cross-validation will be used during parameter tuning. Model output will include risk scores or prediction probabilities indicating the likelihood of an impending accidental event.
Performance will be evaluated using accuracy, sensitivity, specificity, F1 score, and lead time (the time interval between system alert and actual event). The lead-time metric is particularly important because practical utility in clinical care depends on whether alerts occur early enough for staff to intervene.
Outcome Interpretation This study does not impose any medical intervention on participants. All adverse events are part of routine clinical care; the study merely investigates whether AI can anticipate them. Through continuous monitoring and analytical modeling, the research aims to quantify how much predictive information is contained in patient behavior, movement patterns, and environmental context captured by video.
The findings will help determine the feasibility and clinical value of AI-assisted behavioral monitoring in real-world ICUs. If successful, such systems may provide early warnings of delirium, accidental device removal, or other behavior-linked risks. This may reduce emergency interventions, shorten ICU stays, and improve overall patient safety.
Follow-Up To understand the longer-term relevance of the AI predictions, patients will undergo follow-up assessments after discharge at 1 month and 6 months. Follow-up evaluates general health recovery, sleep status, cognition, and whether any delayed complications occurred. Patient and family feedback regarding video monitoring-including comfort level, perceived benefit, or privacy concerns-will also be collected to guide system refinement.
Ethical and Privacy Considerations The study emphasizes privacy protection and informed consent. Cameras are positioned to minimize exposure of unnecessary areas. De-identification is applied before analysis, and all data are managed within controlled hospital systems. Participants may withdraw at any point without affecting their care. The study involves no experimental treatment or additional medical procedures beyond standard ICU monitoring.
Scientific and Clinical Significance This research addresses a critical gap in ICU safety: behavior-based early warning. By combining AI, video analysis, physiology, and environmental data, the study explores an approach that could complement routine monitoring. Beyond predicting specific events, the project may contribute to broader understanding of ICU patient behavioral trajectories and the role of environmental factors such as noise.
The long-term vision is to create a clinically deployable system that supports early intervention, reduces preventable harm, and enhances the efficiency of ICU care.
