Predicting Hospital Readmission for Surgical Patients Using Deep Learning Models With Smart Watch and Smart Ring Sensors Data

Overview

Hospital readmissions are an important measure of healthcare quality and safety. These events create a substantial burden for patients, families, and health systems because they may increase costs, extend recovery time, and lead to more serious postoperative complications. Predicting which patients are at higher risk of readmission remains difficult, as many complications begin silently and are not easily identified in routine clinical evaluations. This study aims to evaluate whether artificial intelligence (AI) can help predict hospital readmissions in surgical patients by analyzing physiological and behavioral data collected before and after surgery. To achieve this, participants will use wearable devices-specifically a smartwatch and a smart ring-capable of continuously monitoring health biomarkers such as heart rate, electrocardiogram (ECG), oxygen saturation, sleep patterns, blood pressure trends, body composition through bioimpedance, and stress indicators. These devices are provided through a technology partnership and sponsorship from Samsung, which supports the study with advanced health technologies. This is a prospective, single-center cohort study conducted at the main tertiary hospital in the state of Amazonas. Approximately 225 to 300 adults undergoing medium- or large-scale elective surgeries will be invited to participate over a 25-month period. All participants will provide informed consent. After enrollment, the study will collect demographic information, preoperative assessments, validated sleep questionnaires, comorbidity indexes such as the Charlson Comorbidity Index, laboratory exams, pulmonary function tests, intraoperative and postoperative data, and hospital discharge information. Participants will be continuously monitored using wearable devices during their hospital stay-including the first 48 hours in the intensive care unit when applicable-and for 30 days after hospital discharge. These physiological data will be integrated with clinical and laboratory information to create a comprehensive dataset. The primary objective is to develop and test artificial intelligence models capable of predicting 30-day hospital readmission following elective surgery. Both deep learning approaches and classical machine-learning techniques will be evaluated. By analyzing large volumes of continuous physiologic data, these models may identify early signs of postoperative deterioration that would otherwise go unnoticed. If successful, this study may improve postoperative care, support earlier clinical intervention, reduce complications, and help healthcare teams provide safer recovery pathways for surgical patients.

Hospital readmissions are well established as key indicators of healthcare quality, particularly in the postoperative setting. Unplanned readmissions often reflect complications that were not fully predicted or adequately managed, generating substantial financial and clinical burden for health systems. In the United States, 3.8 million adult readmissions occurred within 30 days in 2018, with an average cost of US$15,200 per event. In Brazil, similarly high readmission rates have been reported, with 13.4% of more than 22,000 hospitalizations resulting in return visits, largely due to hospital infection, obstructive sleep apnea, and cardiac arrhythmias. Sleep disturbances-including obstructive sleep apnea-are among the most significant contributors to postoperative complications and readmissions. They delay recovery, increase pain sensitivity, impair cognition, and elevate cardiovascular risk. Vascular comorbidities such as hypertension, diabetes, and hyperlipidemia further worsen outcomes by increasing the likelihood of ischemic events, impaired wound healing, and respiratory instability. Understanding how physiologic, behavioral, and sleep-related variables interact to influence postoperative recovery is essential for designing early detection and prevention strategies. Artificial intelligence (AI) has shown increasing promise in predicting clinical deterioration and identifying high-risk patients across multiple surgical fields. A recent systematic review of 26 studies demonstrated that AI-based models-including natural language processing, classical machine-learning algorithms, and deep learning methods-can accurately predict hospital readmissions following major surgery. However, most available datasets do not combine preoperative, intraoperative, postoperative, and continuous physiologic signals collected in real time from wearable devices, representing a major gap in the literature. This prospective cohort study aims to address this gap by integrating high-resolution physiologic and behavioral data captured from wearable devices-specifically smartwatches and smart rings-into a unified perioperative database, with the overarching goal of developing AI models capable of predicting unplanned hospital readmissions within 30 days after surgery. The study seeks to build a comprehensive database of patients undergoing medium- to large-scale elective surgeries, capturing physiologic and behavioral data from wearables and multiparametric monitors during the pre- and post-operative periods, including continuous monitoring in the first 48 hours of intensive care when indicated. Data collection will include continuous wearable monitoring via a web-based interface developed by CETELI/UFAM, pre- and post-operative clinical assessments, laboratory tests, imaging studies, pulmonary function evaluations, polysomnography, and validated sleep questionnaires. The study will evaluate correlations between physiologic and sleep-related variables derived from smartwatches and smart rings and the risk of hospital readmission. Additionally, the study will develop applications for smartwatch and smart ring use in perioperative data capture and employ machine learning approaches-including deep architectures such as causal convolutional networks and recurrent neural networks, as well as classical methods such as support vector machines and decision trees-to predict hospital readmissions. Associations between wearable-derived sleep parameters and 30-day readmission risk will be analyzed, pre- and post-operative sleep data will be correlated with polysomnography and questionnaire results, and sleep quality will be compared between preoperative and post-operative periods. This research is conducted through a multidisciplinary collaboration between the Hospital Universitário Getúlio Vargas (HUGV-UFAM/Ebserh), and the Center for Research and Development in Electronic and Information Technology (CETELI/UFAM). CETELI will be responsible for developing the data-collection software, storage architecture, and predictive algorithms, while clinical oversight, recruitment, and protocol execution will be led by HUGV. Beyond scientific relevance, this project has significant practical importance. In geographically complex regions such as the Amazon, continuous remote monitoring using wearable devices can enhance follow-up, reduce disparities in access to postoperative care, and enable earlier identification of complications. Integrating AI-based prediction models with hospital workflows may ultimately reduce costs, optimize patient flow, and improve the safety and quality of surgical care.