Phase II focuses on refining the Smart CABGuard platform and assessing its feasibility, usability, and acceptability. The intervention integrates a frozen, medically instruction-tuned LLM, a deterministic CRI engine derived from phase I, explainable AI modules, continuous remote ECG telemetry, and a nurse-led triage dashboard. Adult post-CABG patients will be enrolled before discharge and will undergo structured onboarding, including chatbot training and setup of a wearable ECG device, which enables continuous postoperative rhythm surveillance for early detection of arrhythmias such as atrial fibrillation and ventricular tachycardia—common causes of early post-CABG readmission [14], thereby supporting timely risk alerts and clinical escalation. Usability and satisfaction will be measured using the chatBot Usability Scale (BUS) [42] and the Client Satisfaction Questionnaire (8-item version) (CSQ-8) [43].

Phase II will enroll a minimum of 20 post-CABG patients for usability and satisfaction evaluation using the BUS and CSQ-8. This sample size follows established guidelines recommending at least 20 participants for quantitative usability studies to generate statistically meaningful scores [44] and is consistent with published chatbot usability and feasibility studies that typically include 15‐30 participants for pilot validation of a CBT-based mental health chatbot pilot RCT-enrolled 18 participants for feasibility and usability assessment [45], and health information chatbot evaluations commonly report samples of 20‐50 users [46]. Descriptive statistics and thematic analysis will inform refinements, with a usability threshold of ≥80% required before starting the RCT phase. The key components of phase II are illustrated in Figure 1 (see phase II panel).

Phase III (Figure 1, phase III panel) will be a prospective, parallel group, open-label randomized controlled trial that compares Smart CABGuard–assisted care with standard postoperative management. Eligible participants will be adults aged 18 years or older undergoing off-pump CABG, who can give informed consent and have access to a smartphone with internet. Patients who are having other major cardiac procedures (such as valve repair) at the same time, or who have severe cognitive impairment, a terminal illness, or who cannot complete follow-up, will be excluded. Participants will be randomized in a 1:1 ratio to Smart CABGuard plus standard care or standard care alone using a computer-generated permuted block sequence with variable block sizes prepared by an independent statistician. Sequentially numbered, opaque, and sealed envelopes will be assembled by a research assistant not involved in enrollment; after confirming eligibility and obtaining consent, the enrolling clinician will open the next envelope in sequence, thereby maintaining allocation concealment. Because of the nature of the intervention, participants and treating clinicians will not be blinded, but outcome adjudicators and statisticians will remain blinded to allocation. Randomization will not be stratified; any clinically important covariates that show baseline imbalance, defined as a standardized mean difference of >0.10 will be adjusted for in prespecified multivariable analyses. All randomized participants will be analyzed according to the intention-to-treat principle, and any deviations from the randomization procedure will be documented and reported. The intervention group will receive LLM-based chatbot-assisted postoperative monitoring, automated symptom surveillance, CRI-based risk stratification, continuous remote ECG monitoring, nurse-led triage and escalation, and structured digital education with daily chatbot check-ins, which will continue for approximately 24 days postdischarge, with automated alerts notifying nurses of high-risk symptom patterns or ECG-detected arrhythmias.

The research nurse will serve as the primary personnel providing the intervention, including triage and escalation based on chatbot alerts. In the absence of the research nurse (out-of-hours alert management), 2 trained research assistants will support the intervention under the researcher’s supervision to ensure continuity of care. The research assistants will undergo standardized training similar to that of the research nurse to maintain consistency in alert response, escalation thresholds, and documentation practices. There may also be variability in nurse triage and escalation despite standardized training and protocols, which could influence responsiveness to alerts; adherence to the triage algorithm will therefore be monitored as a process measure. The integrated post-CABG triage plan, including Green, Amber, and Red zone criteria for symptom domains, ECG alerts, wound care, and escalation pathways, is provided in Multimedia Appendix 1.

The control group will receive standard discharge instructions and routine outpatient follow-up according to institutional practice. In addition, brief structured telephone calls will be conducted on approximately postoperative days 7, 14, 21, and 30, primarily to document symptoms and capture readmission or emergency visits for outcome ascertainment and safety monitoring. These calls will follow a standardized script and will not provide individualized counseling, risk-stratified advice, or ECG-based monitoring, thereby avoiding replication of the intensity or decision support features of the Smart CABGuard intervention.

The estimated number of participants needed to achieve study objectives is based on the observed 30-day readmission rate of 10.71% following isolated CABG procedures at our institution. For the intervention arm, we assume a 50% relative reduction in readmission (intervention rate: 5.36%) supported by prior digital and telemonitoring interventions: the Perfect Care telemedicine program, which demonstrated a 75% relative reduction in 30-day readmissions after CABG (4% vs 16%) [47], and by remote patient monitoring after CABG showing a 44% reduction in 30-day readmission and mortality [48]. Smart CABGuard integrates continuous ECG monitoring, daily LLM-based symptom surveillance, and nurse-led triage, making a 50% reduction a conservative estimate relative to these comparators.

As this is an exploratory first RCT with no prior randomized evidence for LLM-based post-CABG monitoring, we use a priori power analysis with a 2-sided α=.10 for the primary outcome to reduce the risk of a false-negative result. With 80% power (1−β=.80), the required sample is 361 participants per arm. After allowing for 10% attrition, the final sample size is 400 participants per arm (total 800). Sample size was calculated using G*Power. A sensitivity analysis showing power under effect sizes of 35%, 50%, and 70% relative reduction is provided in Multimedia Appendix 2.

The system uses a Clinical Prediction with Large Language Models framework, in which structured and unstructured patient data are converted into standardized natural-language inputs for prediction tasks. A pretrained 7B-parameter model (Mistral-7B-v0.1) is adapted using parameter-efficient fine-tuning (low-rank adaptation). A fixed few-shot prompting strategy defines the patient context, the prediction objective (30-day readmission), and structured outputs (a risk score with an explanation). To ensure deterministic behavior, decoding parameters are fixed (temperature 0.2, top-p 0.9, and maximum tokens 512). Inputs are provided in structured templates, and outputs are constrained to predefined formats. The LLM is used solely for clinical decision support, and all outputs are checked by qualified clinicians before use.

The AI framework pairs a frozen, medically tuned LLM with a deterministic CRI derived from logistic regression. The LLM will remain unchanged throughout the trial to avoid model drift. All prompts, coefficients, and decision thresholds will be version-controlled and finalized before trial initiation. Any modification will require ethical approval and formal documentation.

Smart CABGuard is intended as a nurse-led decision support without autonomous treatment changes, and all outputs are reviewed by clinicians. It is designed to provide continuous symptom surveillance, rhythm monitoring, and risk-stratified recommendations by a trained cardiac nurse during the first 30 days after CABG surgery, with the explicit aim of reducing unplanned hospital readmissions.

On-site, Smart CABGuard requires secure institutional network access, EHR connectivity, and nurse workstations or tablets with the triage dashboard. Off-site, participants need an Android smartphone with internet, the Smart CABGuard app, and a Bluetooth-enabled single-lead ECG device. All components connect via a secure Trusted Research Environment that manages data exchange between the ECG device, app, AI engine, risk model, and nurse dashboard.

Smart CABGuard is designed as a nurse-led decision support tool. Only registered nurses with at least 1 year of experience in cardiac or cardiothoracic surgery units, and attending physicians in the Department of Cardiovascular and Thoracic Surgery, are authorized to review AI outputs and act on triage recommendations. All authorized users will complete a structured training module (approximately 2 hours) covering interpretation of CRI scores and risk categories, review of LLM-generated explanations, ECG alert management, and escalation algorithms. The research nurse and research assistants responsible for day-to-day triage and escalation in the Smart CABGuard arm will complete this training module and pass competency assessments before independently handling alerts. The training will include case-based simulations and competency checks to ensure that nurses and physicians understand the system’s capabilities and limitations, including the requirement to override or disregard AI recommendations when they conflict with clinical judgment.

Structured EHR variables and deidentified clinical summaries will be processed at discharge within a secure Trusted Research Environment. High-impact predictors (eg, left ventricular ejection fraction and prior stroke) will not undergo imputation. If critical data are missing, the system will issue an “incomplete data advisory” and downgrade prediction confidence, prompting clinician review. The model output includes CRI score, risk category (low, moderate, and high), recommended triage intensity, and clinician-oriented natural-language explanation. The system functions strictly as CDS. Final treatment decisions remain under the clinician’s authority.

The AI intervention applies explicit input data eligibility criteria and continuous quality checks in addition to participant-level inclusion and exclusion. During follow-up, the chatbot delivers daily structured symptom check-ins, and the wearable ECG device records single-lead rhythm strips at scheduled times and on demand. Only perioperative EHR records containing predefined core variables and passing basic consistency checks are used for CRI-based predictions; the full list of input data variables extracted from the EHR is provided in Multimedia Appendix 3. Incomplete or inconsistent records are excluded, flagged for clinician review, and generate an incomplete data advisory with downgraded prediction confidence.

During follow-up, only fully completed check-ins reporting symptoms, medication adherence, and wound status in supported languages are considered eligible for analysis. ECG segments that meet predefined signal-quality thresholds are processed by the LLM and ECG pipeline, while poor-quality or unsupported inputs trigger repeat recording or charging prompts. If daily check-ins are missed for more than 48 hours, automated reminders are sent, followed by nurse phone contact after 72 hours of persistent nonadherence, with reasons documented. AI-generated risk estimates and triage recommendations are used only when these input data eligibility and quality criteria are satisfied.

The intervention incorporates a single-lead wearable ECG device (Amrita Spandanam, Amrita Vishwa Vidyapeetham, India), which records clinical-grade ECG via skin contact electrodes and streams data via Bluetooth to the patient’s Android smartphone for secure transmission to the Smart CABGuard nurse dashboard. The device and remote-monitoring framework, developed within the Amrita-Spandanam/AIM-Vitals platform, are compliant with ANSIAAMI EC13-2002 and IEC 60601-2-25/60601-2-27 standards for diagnostic ECG systems and are protected under US Patent 10,542,889 (Systems and Methods for Remote Health Monitoring and Management). Prior studies have shown satisfactory ECG signal quality and clinical usability across in-hospital treadmill testing, ambulatory monitoring, and remote rural deployments, with performance comparable with standard clinical ECG devices. In an Internet of Things–based smart-edge remote-monitoring study using this platform, downstream alerting models achieved high diagnostic performance for cardiac conditions (precision 0.87, recall 0.83, and F₁-score 0.85), and deep-learning models trained on Amrita Spandanam ECG data for arrhythmia detection have demonstrated acceptable multiclass classification performance (F₁-scores around 0.71 on benchmark datasets), supporting its suitability for continuous, safety-focused rhythm surveillance in resource-limited settings [49-51].

The wearable device can be worn on the wrist or carried in a pocket and connects via Bluetooth to a mobile app on the patient’s phone, with location services enabled for continuous monitoring. It should be removed only during bathing. Each patient will receive 2 devices: 1 active unit with a 24-hour battery life and 1 spare, as charging each device takes approximately 5‐6 hours. The troubleshooting steps for the Amrita Spandanam ECG device are provided in Multimedia Appendix 4. In accordance with the signed informed consent, all enrolled participants will return the remote ECG-monitoring devices at the first preplanned follow-up visit within 30 days after surgery.

For each enrolled participant, the AI system produces (1) a numerical CRI score, (2) a categorical risk label (low, moderate, or high), (3) a natural-language explanation summarizing key drivers of risk, (4) triage recommendations specifying suggested monitoring intensity and actions, and (5) ECG-based alert flags for arrhythmias of interest. These outputs are displayed to the nursing team on the triage dashboard and are used to prioritize follow-up calls, schedule expedited clinic reviews, and prompt urgent physician assessment or emergency referral when indicated. Final decisions regarding diagnostic workup, hospital readmission, and treatment modifications remain the responsibility of the treating clinicians, who may override or disregard AI recommendations whenever they conflict with clinical judgment.

The primary outcome is the 30-day all-cause unplanned hospital readmission rate following the index CABG admission. The 30 days will be measured from the date of surgery. Readmission is defined as any unplanned inpatient admission within 30 days of surgery; planned admissions are excluded. Emergency department or outpatient visits without admission will not be considered readmissions but will be captured as secondary outcomes. Secondary outcomes include time to detection of postoperative complications and postoperative morbidity, assessed using the validated 13-domain Cardiac Postoperative Morbidity Score (C-POMS) [52]. Given its limited psychometric validation in Indian post-CABG cohorts, it will be used as a secondary outcome, and its association with clinical end points (eg, length of stay and readmission) will be explored.

All potential 30-day readmissions and emergency visits will be identified from institutional EHRs; structured follow-up telephone calls on postoperative days 7, 14, 21, and 30; and patient or caregiver self-reports. Two clinicians (a cardiac nurse and a cardiothoracic surgeon or cardiologist), blinded to trial allocation, will independently review discharge summaries, admission records, and other source documents using a predefined adjudication form to determine whether events meet the protocol definition of 30-day all-cause unplanned readmission. Discrepancies will be resolved by consensus, with a third adjudicator consulted if required. For patients admitted to nonindex hospitals, records will be requested with participant consent. If documentation cannot be obtained, events will be classified using the best available information and flagged as “probable” readmissions for sensitivity analyses. Events will be categorized as “readmission,” “planned admission,” “ED or outpatient visit without admission,” or “no event,” without further formal cause-of-readmission adjudication beyond descriptive reporting.

The CRI will be derived from a multivariable logistic regression model, restricted to a maximum of 7‐8 predictors based on the ≥10 events per variable criterion (≈75 available outcome events). Internal validation will use 1000 iteration bootstrap resampling; temporal validation will use the 2023‐2024 holdout cohort. Model performance will be reported using ROC-AUC, PR-AUC, calibration plots, Hosmer-Lemeshow test, and Brier score. Clinical usefulness will be assessed by decision curve analysis. The risk classification threshold (low/moderate/high) will be selected based on net benefit across a probability range of 0.05‐0.30 and frozen before phase II. The LLM-based model will be formally compared with the logistic regression CRI using DeLong’s AUC comparison, net reclassification improvement, and integrated discrimination improvement on the same temporal validation set. Shapley Additive Explanations values will identify key predictors. Interrater agreement for PEMAT-A/V reviews will be assessed using the intraclass correlation coefficient.

The BUS total and subscale scores will be reported as means and SDs or medians and IQRs, depending on the normality of data distribution. The proportion meeting the prespecified usability threshold (≥80%) will be reported. CSQ-8 total scores will be summarized with mean and SD; the proportion scoring ≥26 or in intermediate or high satisfaction categories will be reported against a target of ≥75%. Associations between usability or satisfaction scores and engagement measures will be examined using appropriate parametric or nonparametric tests.

All randomized participants will be analyzed according to the intention-to-treat principle. Baseline comparability will be assessed using chi-square or Fisher exact test and 2-tailed t test or Mann-Whitney U test.

Process indicators, such as alert frequency, response time, and chatbot engagement, will be explored as potential mediators of outcomes. For the randomized controlled trial, all randomized participants with available 30-day readmission data will be included in the primary analysis. Missing primary outcome data will be minimized through structured follow-up and EHR verification.