LLM-Based Intelligent Health Management Assistant in Life-Cycle Health Management of Cardiac Surgery Patients

Overview

This is a single-center, prospective, randomized, open-label, parallel-controlled clinical study to evaluate the effectiveness of a large language model (LLM)-based intelligent health management assistant in the life-cycle health management of patients after cardiac surgery. A total of 500 adult patients who undergo cardiac surgery (including coronary artery bypass grafting, heart valve surgery, great vessel surgery, congenital heart disease correction, and other cardiac procedures) at Beijing Anzhen Hospital will be randomly assigned in a 1:1 ratio to an intervention group or a control group, stratified by age (\<65 vs ≥65 years) and surgery type. The intervention group will use the LLM-based mobile health management application in addition to standard postoperative care, while the control group will receive standard postoperative care alone. The application integrates multimodal clinical data into a personalized health profile and provides surgery-type-specific postoperative management recommendations, medication adherence reminders, complication early warning, and cardiac rehabilitation guidance. The primary outcome is the composite endpoint of major adverse cardiac and cerebrovascular events (MACCE), defined as all-cause death, non-fatal myocardial infarction, non-fatal stroke, or unplanned cardiovascular reoperation/reintervention, within 12 months after randomization. Secondary outcomes include health-related quality of life (EQ-5D-5L), cardiovascular rehospitalization rate, medication adherence (MMAS-8), postoperative complication rate, and cardiac rehabilitation achievement rate. Follow-up visits are scheduled at 1, 3, 6, 9, and 12 months post-randomization.

Cardiac surgery, including coronary artery bypass grafting (CABG), heart valve replacement/repair, great vessel surgery, and congenital heart disease correction, is a critical treatment for severe cardiovascular diseases. Despite significant improvements in surgical safety and perioperative care, long-term postoperative health management remains challenging. Patients face complex needs including anticoagulation therapy management, cardiac rehabilitation, complication prevention, comorbidity control, and psychological well-being. Current postoperative follow-up models are limited by poor individualization across different surgery types, suboptimal anticoagulation control (particularly after mechanical valve replacement), low cardiac rehabilitation participation, high loss-to-follow-up rates, and insufficient attention to postoperative psychological distress. Large language models (LLMs) offer new opportunities for personalized health management through their capabilities in natural language understanding, knowledge reasoning, and multimodal data integration. This study evaluates an LLM-based intelligent health management assistant, co-developed by Beijing Anzhen Hospital, Beijing Zhilan Medical Technology Co., Ltd., and Beijing Haitian Ruisheng Technology Co., Ltd., deployed as a mobile application. This is a single-center, prospective, randomized, open-label, parallel-controlled study enrolling 500 patients discharged after cardiac surgery from Beijing Anzhen Hospital. Eligible participants are adults (≥18 years) who have undergone cardiac surgery, are clinically stable for discharge, have smartphone access, and provide informed consent. Key exclusion criteria include prolonged ICU stay (\>30 days), life expectancy less than 12 months, severe cognitive impairment without caregiver assistance, heart transplantation, concurrent interventional trial participation, or inability to complete 12-month follow-up. Participants are randomized 1:1 to intervention or control using computer-generated block randomization (block sizes of 4 or 6) via an interactive web response system (IWRS), stratified by age (\<65 vs ≥65 years) and surgery type (CABG vs valve surgery vs great vessel surgery vs other). The intervention group receives, in addition to standard postoperative care, the LLM-based mobile health management application with the following core functions: (1) construction of a surgery-centered structured personal health profile by uploading operative records, discharge summaries, laboratory results, imaging data, and prescriptions; (2) intelligent postoperative health assessment informed by continuously updated clinical practice guidelines covering post-cardiac surgery rehabilitation, anticoagulation management, valve surgery follow-up, and great vessel postoperative monitoring; (3) surgery-type-specific management recommendations, including secondary prevention and lifestyle interventions for CABG, anticoagulation and International Normalized Ratio (INR) monitoring reminders for valve surgery, blood pressure control and imaging follow-up reminders for great vessel surgery, and activity guidance for congenital heart disease repair; (4) medication adherence management with reminders, drug interaction alerts, and adverse reaction monitoring for anticoagulants, antiplatelets, beta-blockers, ACEI/ARBs, and statins; (5) early warning for common postoperative complications (wound infection, arrhythmia, pericardial effusion, pleural effusion) and individualized cardiac rehabilitation exercise prescriptions. The control group receives standard postoperative clinical care and follow-up, including routine outpatient visits, standard pharmacotherapy, and conventional discharge education, without access to the application. Follow-up assessments are conducted at 1, 3, 6, 9, and 12 months post-randomization via clinic visits or remote follow-up (telephone/video). Comprehensive questionnaire assessments are performed at 6 and 12 months. For patients requiring anticoagulation, INR values and anticoagulant dosages are recorded at each visit. The primary outcome is the composite MACCE endpoint within 12 months, comprising all-cause death, non-fatal myocardial infarction (per the Fourth Universal Definition), non-fatal stroke (imaging-confirmed), and unplanned cardiovascular reoperation or reintervention. Secondary outcomes include change in EQ-5D-5L index and VAS scores, cardiovascular rehospitalization rate, medication adherence by MMAS-8 (with time in therapeutic INR range for anticoagulated patients), postoperative complication rates, and cardiac rehabilitation achievement rate. Exploratory outcomes include application usage patterns, user satisfaction, MACCE differences across surgery-type subgroups, and postoperative anxiety/depression scores (PHQ-9, GAD-7). Sample size was calculated assuming a 12-month MACCE rate of 13% in the control group and 7% in the intervention group (two-sided alpha=0.05, power=80%), requiring 210 per group, inflated to 250 per group (500 total) to account for 15-20% attrition. The primary analysis uses the intention-to-treat population with Kaplan-Meier estimation and Cox proportional hazards regression adjusting for stratification factors and key baseline covariates. An interim analysis is planned after 50% enrollment completes 6-month follow-up, using the O'Brien-Fleming alpha-spending boundary. The study is self-funded, with technology partners involved only in application development and maintenance, not in study design, data analysis, or result interpretation.