Nurse-Led Large Language Model Chatbot for Predicting and Preventing Complications After Coronary Artery Bypass...

Thirty-day unplanned readmission following coronary artery bypass grafting (CABG) affects 10%-20% of patients and is a key quality indicator, particularly in low- and middle-income countries (LMICs) where access to cardiac rehabilitation is limited. Existing risk models are static, lack real-time engagement, and no validated large language model (LLM)-based clinical decision support (CDS) system exists for post-CABG readmission prevention.

This protocol describes the development, validation, and evaluation of Smart CABGuard, a nurse-led, LLM-based CDS chatbot with continuous remote electrocardiographic (ECG) monitoring, to predict and prevent postoperative complications, and 30-day unplanned readmission after isolated CABG.

This multiphase translational study is conducted at Amrita Institute of Medical Sciences, Kochi, India, following TRIPOD-LLM (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis-Large Language Model) and CONSORT (Consolidated Standards of Reporting Trials) AI reporting guidelines. Phase I uses an ambispective design (retrospective: January 2020 to December 2024; prospective needs survey: June 2025 to January 2026) to develop a logistic regression-based Complication Risk Index (CRI), assessed via receiver operating characteristic area under the curve, Brier score, and decision curve analysis. Phase II refines Smart CABGuard by integrating a frozen Mistral-7B LLM (low-rank adaptation fine-tuned), the CRI engine, explainable AI, and single-lead ECG telemetry (Amrita Spandanam device), with a usability threshold of ≥80%. Phase III is a prospective, parallel-group, open-label randomized controlled trial (RCT). Adults aged ≥18 years undergoing isolated CABG with smartphone access are randomized 1:1 via permuted block randomization, with allocation concealment. The intervention arm receives Smart CABGuard-assisted care (daily chatbot check-ins, CRI-based risk stratification, ECG monitoring, and nurse-led triage) for approximately 24 days postdischarge plus standard care; the control arm receives standard care with structured telephone follow-up for outcome ascertainment only. Outcome assessors and statisticians are blinded; analyses follow the intention-to-treat principle. The primary outcome is 30-day all-cause unplanned readmission. Based on a baseline rate of 10.71%, a 50% relative reduction, α=.10, and 80% power, the target sample size is 800 patients (400 per arm, including 10% attrition).

Ethics approval was granted by the Institutional Ethics Committee of Amrita Institute of Medical Sciences on April 16, 2025. Funding was awarded by Sigma Theta Tau International Honor Society of Nursing, Small Grants Program (grant 21650) in June 2025. Phase I data extraction commenced in May 2025 and is projected to be completed by April 2026; the patient needs survey (June 2025 to January 2026) is ongoing. Phase II usability evaluation is projected from May to July 2026. Phase III recruitment is anticipated from August 2026 to September 2027. Results are expected to be published in early 2028.

Smart CABGuard is the first RCT protocol of an LLM-based nurse-led CDS system for post-CABG readmission prevention in an LMIC setting, aiming to establish the usefulness, safety, and feasibility of AI-augmented postoperative surveillance.