Early delirium detection and prediction in hospitalised patients through analysis of clinical signs and behaviours.

This research will collect data on the clinical and behavioural patterns linked to delirium in hospitalized patients. All consecutive individuals with and without a current diagnosis of delirium admitted as inpatients in the acute and sub-acute services of St Vincent’s Hospital Melbourne will be considered for recruitment to this prospective, longitudinal observational study. First Nations peoples over 50 years of age, and other individuals over 65 years of age, will be eligible for inclusion from the following SVHM study sites: St Vincent’s Hospital Melbourne Inpatient Services (IPS), Bolte Wing (SVHM), and St George’s Hospital, Kew. Once recruited, patients will continue to receive standard medical care from the treating team in addition to their participation in the study. To effectively monitor participants, re-identified data will be collected from both wearable devices and the hospital database system. The wearable device used in this study is developed by Verisense Health and is a wrist-worn sensor designed to capture continuous physiological and activity-related data. Participants will be asked to wear the device continuously, including while sleeping, for uninterrupted data collection. The device may be temporarily removed for activities such as showering, bathing, or during medical procedures—particularly radiation-based procedures or any intervention where the attending technician advises removal. In all such cases, participants will be reminded to reapply the device as soon as the procedure is completed to maintain consistent data capture. Unique patient IDs will be created for each participant, and their wearable information and clinical information will be linked to these IDs to facilitate continuous monitoring and recording of relevant data. The wearable device will track physiological indicators including skin temperature, heart rate variability, sleep stages from movement, blood volume changes, sympathetic arousal metrics, and levels of physical activity. From the hospital database, re-identified demographic information including age, gender, past clinical history (including previous admissions and comorbidities), and current clinical information relevant to delirium risk (e.g., blood test results, fluid and food intake, bowel movements, medication usage, other laboratory test results, and vital signs) will be collected. Observed signs and indications of delirium, such as agitation and any positive results from the 4AT assessment, will also be documented. The 4AT will be administered by trained clinical staff at least every 72 hours and also upon any observed behavioural change suggestive of delirium. The presence or absence of delirium will be determined based on clinical diagnoses documented in the patient’s medical records by the treating clinicians. These diagnoses will generate timestamped ground truth labels, marking the onset and resolution of delirium episodes. The timestamped labels will be aligned with the wearable sensor data and associated clinical information to create a comprehensive dataset for training and evaluating AI models developed for delirium prediction. Participants will be observed for the duration of their hospital admission or up to a maximum of 30 days from admission, whichever is shorter. Machine learning and deep learning models, including Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Decision Tree, Random Forest, Gradient Boosting, Artificial Neural Networks (ANNs), Long Short-Term Memory (LSTM), and Transformer models will be developed to analyse the data. These models will be compared using standard performance measures such as accuracy, precision, recall, and F1 score to determine the most effective approach for predicting delirium.