Effectiveness of an AI-enhanced management system for coronary heart disease (AIM-CHD): rationale and design of a single-centre, open-label, randomised, parallel-controlled trial

Abstract

Introduction

Effective secondary prevention of coronary heart disease (CHD) is often hindered by limited healthcare resources and poor patient adherence. We therefore developed an artificial intelligence (AI)-enhanced CHD management platform (AIM-CHD) that (i) automatically captures follow-up data through AI-driven voice calls, optical character recognition of laboratory reports and wearable sensor streams; (ii) enables closed-loop, automated risk factor management; and (iii) dynamically personalises follow-up intensity via continuously updated risk stratification and achievement of treatment targets. This trial aims to evaluate whether AIM-CHD improves risk factor control and reduces cardiovascular events compared with usual care.

Methods and analysis

In this prospective, single-centre, open-label, randomised controlled trial, 1100 CHD patients aged 18–85 years will be enrolled at Fuwai Hospital and randomised 1:1 to either the AIM-CHD group (n=550) or the usual care group (n=550) for a 3 month post-discharge intervention. The primary outcome is low-density lipoprotein cholesterol (LDL-C) level at 3 months. Secondary outcomes include target achievement for LDL-C and blood pressure, as well as glycosylated haemoglobin level, nonsmoking status, body mass index, composite cardiovascular endpoint and medication adherence.

Ethics and dissemination

Ethical approval was approved by the Ethics Committee of Fuwai Hospital on 4 November 2024 (2024-2422). The findings will be disseminated in peer-reviewed publications. An anonymised template of the written informed-consent form (Chinese and English versions) is available as Supplementary Material 1.

Trial registration number

ClinicalTrial, NCT06686056.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This prospective, open-label, randomised controlled trial will enrol 1100 post-discharge patients with coronary heart disease, providing adequate power for the primary lipid outcome.

• Randomisation stratified by baseline low-density lipoprotein cholesterol (<1.8 vs ≥1.8 mmol/L) minimises allocation and baseline imbalances.

• Single-centre recruitment may limit external validity and warrants confirmation in multicentre settings.

Introduction

Coronary heart disease (CHD) remains a leading cause of morbidity and mortality worldwide. In China, an estimated 11.39 million people were living with CHD in 2022, and the burden is expected to rise.¹ Secondary prevention—emphasising risk factor control and lifestyle modification—is a cornerstone of long-term CHD management, as endorsed by global guidelines.²,⁶

However, adherence to secondary prevention remains suboptimal, particularly in low- and middle-income settings.⁷ Many patients discontinue healthy behaviours and medications soon after discharge, mistakenly perceiving recovery as cure. Conventional follow-up strategies, often limited to verbal advice or printed materials, lack continuity and engagement. Studies show that approximately 26–65.8 % of patients misinterpret the chronic nature of CHD, undermining the long-term benefits of guideline-based care.⁸,¹⁰ Mobile health (mHealth) technologies offer new opportunities for scalable, patient-centred interventions through phone calls, SMS and apps.¹¹ Evidence suggests mHealth can improve blood pressure control, increase physical activity and reduce psychological distress in cardiac patients.^{12 13} However, its effect on low-density lipoprotein cholesterol (LDL-C) control remains uncertain, and randomised trials show mixed results.¹⁴,¹⁶ Furthermore, existing interventions often face key limitations: fragmented management of CHD risk factors,¹⁷ generic messaging with poor personalisation,^{14 15} high staffing demands^{18 19} and poor integration of hospital and community data systems.¹⁷,¹⁹

To address these challenges, we developed an AI-enhanced mobile platform for CHD management (AIM-CHD), designed to support long-term patient self-care and real-time monitoring. We are conducting a randomised controlled trial (RCT) to evaluate the effectiveness of AIM-CHD in optimising cardiovascular risk control and reducing adverse events (AEs). The study design adheres to the Mobile Health Evidence Reporting and Assessment and Standard Protocol Items: Recommendations for Interventional Trials guidelines,^{20 21} ensuring methodological rigour and transparency.

Methods and analysis

Intervention platform overview

The AIM-CHD platform was collaboratively developed by clinicians, nurses, patients and software engineers at Fuwai Hospital. The platform underwent rigorous co-creation, rapid prototyping and preliminary usability and feasibility testing before deployment. As illustrated in figure 1, the system integrates diverse multimodal data streams including synchronised hospital electronic health records (EHR), in-app questionnaires, biometric data from wearable devices, intelligent voice calls and artificial intelligence (AI) medical image recognition. This integrated data environment facilitates comprehensive patient risk stratification, identifies poorly controlled cardiovascular risk factors and dynamically customises follow-up intervals and therapeutic strategies through explicit algorithmic logic described below. Detailed software interfaces and practical demonstrations of the AIM-CHD system are provided in the online supplemental figures S1-S6.

Figure 1. Operational logic and intervention indicators of the AIM-CHD App AI, Artificial Intelligence; App, Application; LDL-C, Low-Density Lipoprotein Cholesterol; AIM-CHD, artificial intelligence-enhanced management system for coronary artery disease.

Data sources and integration

Hospital EHRs

De-identified structured EHR data including demographics, clinical history, family history, vital signs, diagnoses, treatments, auxiliary examinations, coronary angiography and intervention outcomes are continuously synchronised to maintain an up-to-date clinical profile.

In-app questionnaires and automated laboratory data extraction (optical character recognition (OCR) + large language model (LLM))

The platform employs structured digital questionnaires collecting real-time patient-reported outcomes across four critical domains: risk factor modification, pharmacotherapeutic adherence, laboratory parameters and symptomatic assessments. Patients complete follow-ups via Proactive Follow-Up Questionnaires or scheduled Planned Follow-Up Questionnaires. Laboratory data from uploaded medical reports are extracted automatically via a combined OCR and LLM technology. This advanced natural language processing (NLP)-based approach accurately recognises and integrates diverse formats of laboratory indices directly into the patient’s EHR, reducing manual entry errors and enhancing data completeness.

Wearable devices

Real-time biometric data from wearable devices (eg, Huawei smartwatches) are integrated through standardised Application Programming Interface. Collected parameters include continuous ambulatory blood pressure, heart rate variability, step counts and physical activity duration, enabling real-time assessment of lifestyle modifications and cardiovascular status. Huawei devices used in this trial have been clinically validated and shown to deliver reliable accuracy for all listed parameters.²²,²⁴ Participation in the wearable-monitoring component is optional and serves solely as a supplementary data source, with core trial procedures remaining unaffected if a participant declines or discontinues device use.

Intelligent voice calls

Intelligent voice calls powered by NLP and semantic intent recognition are used to engage patients in structured conversations. Scheduled around each study visit (1 and 3 months post-discharge), these calls remind and encourage patients to complete follow-up visits and questionnaires while simultaneously collecting patient-reported symptoms, medication adherence and risk factor control status.

Risk stratification and dynamic follow-up customisation

Risk stratification logic

Baseline risk is quantified with the validated 6 month post-discharge Global Registry of Acute Coronary Events (GRACE) model, which predicts medium-term mortality across the full spectrum of acute coronary syndrome.²⁵ Specific thresholds used are non-ST-elevation acute coronary syndrome (NSTE-ACS): low-intermediate (<3%), intermediate-high (3–8%), high (>8%); ST-elevation acute coronary syndrome (STE-ACS): low-intermediate (<4.4%), intermediate-high (4.4–11%), high (>11%); chronic coronary syndrome cases use thresholds from NSTE-ACS provisionally, due to lack of specific validated standards. All variables required for the GRACE calculation are collected routinely at discharge, facilitating seamless integration into the AIM-CHD system. Individual risk profiles are automatically recalculated every month as new clinical data become available.

Dynamic follow-up interval adjustment

In the AIM-CHD system, follow-up intervals are dynamically adjusted by algorithm based on patients’ updated data. In real-world deployment, the AIM-CHD recommends personalised schedules: monthly for high-risk individuals, quarterly for intermediate-risk and biannually for low-risk, based on dynamic risk stratification and target achievement status. For the current 3 month randomised trial, however, follow-up assessments are protocol-specified and uniformly conducted at 1 and 3 months post-discharge. Additionally, suboptimal control (blood pressure, heart rate or fasting glucose) triggers intensified monitoring intervals, with reassessments every 3 days until control targets are achieved (see table 1).

Table 1. Management targets and adverse scenario criteria in artificial intelligence-enhanced management system for coronary artery disease.

Intervention item	Indicator	Target goal	Threshold for medical adjustment	Adverse scenario identification
Lipids	LDL-C	<1.8 mmol/L	>1.8 mmol/L	_
Blood pressure	Systolic blood pressure	<130 mmHg	>130 mmHg	Serum creatinine increase >30% from discharge; (2) systolic BP <90 or >160 mmHg; (3) diastolic BP <60 or >120 mmHg
	Diastolic blood pressure	<80 mmHg	>80 mmHg
Heart rate	Resting heart rate	55–60 bpm	>70 bpm	Heart rate <50 bpm or >100 bpm
Blood glucose	Haemoglobin A1c	<7.0%	>7.0%	Fasting blood glucose ≤4 mmol/L or ≥14 mmol/L; symptoms of hypoglycaemia
	Fasting blood glucose	4.4–7.2 mmol/L	>7.2 mmol/L
	2 hour postprandial glucose	<10 mmol/L	>10 mmol/L
CHD symptoms	Symptoms	No symptoms	_	Chest tightness, chest pain, palpitations or shortness of breath occurring over the past month, characterised by prolonged, intermittent or spontaneous episodes
Medication adherence	Antiplatelet agents	≥80% adherence	_	Bleeding (eg, eye, nasal or oral bleeding, black/red stool); 2) Hb decrease >30 g/L or <100 g/L
	Statins	≥80% adherence	_	Muscle pain, weakness or discomfort; (2) ALT >120 U/L; (3) CK >1000 U/L
Smoking	Smoking	Complete cessation	_	_
Alcohol consumption	Intake	Abstain if no history; limit to <100 g/week if prior use	_	_
Physical activity	Exercise	30–60 min of moderate-intensity, 5 days/week	_	_
Dietary	Diet	Rich in fruits, vegetables, legumes, fibre, polyunsaturated fats, nuts, fish	_	_

ALT, alanine aminotransferase; CK, creatine kinase; Hb, haemoglobin; LDL-C, low-density lipoprotein-cholesterol.

Intervention methods

Real-time clinical surveillance and emergency response

Critical clinical deterioration indicators (see table 1) trigger automated alerts. Patients immediately receive in-app notifications, facilitating rapid online telemedicine consultations and direct emergency referrals to Fuwai Hospital, thus minimising therapeutic latency.

Precision management of cardiovascular risk factors

In clinically stable patients, the AIM-CHD system conducts comprehensive assessments of cardiovascular risk control, integrating both lifestyle factors (diet, physical activity, smoking, alcohol use, body weight and medication adherence) and clinical indicators (LDL-C, blood pressure, heart rate and blood glucose). Management strategies include (1) quantitative feedback: real-time personalised comparison of biomarker trajectories vs evidence-based targets (see table 1), supporting patient self-awareness and provider decision-making; (2) behavioural reinforcement: automated motivational prompts are delivered to reinforce treatment adherence and promote healthy behaviours; (3) therapeutic escalation recommendations: for parameters failing to meet targets, automated alerts suggest therapy intensification. For example, if LDL-C exceeds 1.8 mmol/L, a message such as “Your lipid control is currently suboptimal. If you are already taking medications regularly, please consult your physician about adding ezetimibe or other agents” will be triggered; (4) medication adherence optimisation: the platform synchronises multimodal reminder tools (eg, in-app notifications and AI voice calls) with individualised chronotherapy schedules. Patients can customise reminder frequency and preferred time windows to improve adherence.

Adaptive patient education module

The platform provides personalised educational content—such as instructional videos and evidence-based articles—tailored to patients’ evolving clinical profiles. The education library is organised into four modules—disease understanding, medication adherence, lifestyle modification and cardiac rehabilitation—comprising a total of 8–12 multimedia lessons per module (≈40 lessons overall). Materials are updated every 3 months to reflect the latest guideline recommendations, while preserving the four-module structure during the study period.

Technical architecture

AIM-CHD adopts a modular architecture comprising a WeChat-based front-end interface and a secure back-end infrastructure built on the .NET framework. All data transmission and storage comply with China’s Level three cybersecurity standard (GB/T 22 239–2019), incorporating dual-layer encryption based on national cryptographic protocols (SM4 for symmetric encryption, SM9 for asymmetric encryption), with TLS 1.3 for secure data transmission and SHA-256 for data integrity verification.

Clinical integration and workflow automation

The AIM-CHD framework enables bidirectional interoperability across patients, medical centre and community health services, establishing a cohesive care continuum to support efficient resource allocation and seamless multidisciplinary integration (figure 2). Tertiary medical centres use the platform to deliver guideline-based telemedicine interventions, coordinate outpatient and inpatient workflows, and ensure system stability via a dedicated informatics infrastructure.

Figure 2. Components of the artificial intelligence-enhanced CHD management platform (AIM-CHD) Healthcare Model AIM-CHD efficiently integrates patients, the medical centre and community medical resources, enhancing the overall coordination and utilisation of healthcare services. CHD, coronary artery disease.

Meanwhile, AIM-CHD empowers patients to participate actively in the secondary prevention of CHD through real-time biometric data transmission and collaborative follow-up with community healthcare providers.

Study design

This single-centre, open-label, RCT employs a parallel-group design to assess the efficacy of the AIM-CHD platform in optimising CHD risk factor control and reducing cardiovascular event incidence (study flowchart: figure 3). Recruitment began with the first participant randomised on 23 November 2024. As of 24 May 2025 (manuscript-submission date), 1100 participants—the full planned sample—had been enrolled. The follow-up for the last participant is concluded on 30 June 2025. At the time of revision, data cleaning is still in progress, and no statistical analysis has yet been performed.

Figure 3. AIM-CHD study flow diagram AIM-CHD, artificial intelligence-enhanced management system for coronary artery disease; LDL-C, low-density lipoprotein cholesterol, BMI, body mass index.

Participant recruitment and eligibility

Participants are being consecutively recruited at the Coronary Heart Disease Centre of Fuwai Hospital prior to discharge. Inclusion criteria include (1) age 18–85 years with clinically confirmed CHD, (2) patient or caregiver proficient in smartphone usage and mobile applications and (3) willingness to provide written informed consent. Exclusion criteria are (1) severe cognitive impairment, (2) active malignancy or life expectancy <3 months, (3) multisystem organ failure and (4) declined participation online supplemental material 2.

Randomisation

After routine discharge prescriptions are finalised, an independent study coordinator randomises each eligible patient in a 1:1 ratio through a secure, central web-based system that automatically retrieves the most recent in-hospital LDL-C result and applies permuted-block allocation stratified by LDL-C (<1.8 mmol/L vs ≥1.8 mmol/L).

Intervention group

Participants in the intervention group will receive training to use AIM-CHD alongside usual post-discharge care. Following installation of the application on personal mobile devices and automated synchronisation with inpatient data, participants will complete a structured orientation session (10 min) featuring core functionality demonstrations and operational guidelines. Competency verification will be ensured through supervised interactive simulations with clinical staff. To reinforce longitudinal engagement, multimodal educational resources (print-ready manuals and in-app tutorials) are accessible. A dedicated technical support hotline will provide ongoing assistance throughout the trial duration.

Control group

Participants in the control group will receive usual discharge instructions, including verbal and printed materials regarding medication schedules, follow-up timelines and lifestyle optimisation, in line with pre-existing institutional practices. To ensure comparable data capture without exposure to the AIM-CHD intervention, control participants will use a restricted software interface limited strictly to data collection at the 3 month endpoint without additional risk stratification, educational content or reminders.

Outcomes

Primary outcome

The primary outcome is the plasma LDL-C level (mmol/L) obtained 3 months (±14 days) after hospital discharge; when multiple values are available within the window, the most recent result is used. LDL-C is the guideline-endorsed primary lipid target in secondary prevention. Early reductions show a strong association with long-term cardiovascular risk, making it a sensitive, quantifiable surrogate that can capture the intervention’s effect within the study’s 3 month follow-up window.²⁶

Secondary outcomes

1. Achievement of LDL-C target: proportion of participants with LDL-C <1.8 mmol/L at 3 months.

2. Achievement of blood-pressure target: proportion with self-reported resting blood pressure <130/80 mmHg at 3 months.

3. Glycaemic control: glycated haemoglobin (HbA1c, %) at 3 months.

4. Smoking cessation: proportion reporting zero cigarettes in the preceding month.

5. Body mass index (BMI): BMI (kg/m²) calculated from self-reported weight and height at 3 months.

6. Composite cardiovascular endpoint: first occurrence of all-cause death, non-fatal myocardial infarction, non-fatal stroke or cardiovascular-related rehospitalisation within 3 months post-discharge.

7. Medication adherence: proportion of days covered ≥80% for both an antiplatelet agent and a statin during the 30 days before the 3 month visit.

Data collection and follow-up

Baseline clinical parameters will be automatically extracted from the institutional Hospital Information System (HIS). Outcome data for both groups will be collected at 3 months post-discharge using structured in-app questionnaires and intelligent voice calls.

Post-discharge LDL-C and HbA1c levels will be measured at local hospitals. Patients are instructed to upload photographs of their laboratory reports. Extracted values are automatically synchronised with the EHR. Self-monitored blood pressure and glucose values may be submitted either through manual input into the app or reported verbally during AI-driven voice calls. To ensure timely data collection, automated voice reminders are issued at −5,–3, 0 and +3 days relative to each scheduled follow-up point. If data remain incomplete, WeChat push notifications are triggered to prompt patient response. For patients who do not respond to automated prompts, trained study personnel will conduct telephone interviews to complete outcome ascertainment.

Adverse event reporting and safety monitoring

AEs will be documented and graded based on the Common Terminology Criteria for Adverse Events version 5.0. All safety assessments will be conducted under the oversight of the Ethics Committee of Fuwai Hospital. Participants may withdraw from the trial at any time if they are unable to tolerate the intervention.

Sample size determination

The primary efficacy analysis will assess superiority of the intervention by comparing mean LDL-C levels between groups at 3 month follow-up. Contemporary Chinese registry studies²⁷,³⁰ involving patients with CHD undergoing percutaneous coronary intervention report baseline LDL-C levels ranging from 2.1 to 2.7 mmol/L, with SD between 0.80 and 0.95 mmol/L. Assuming a conservative SD of 0.95 mmol/L and a clinically meaningful between-group difference of 0.2 mmol/L (approximately 10% relative reduction), we estimated that a sample size of 950 participants would provide 90% statistical power at a two-sided significance level (α=0.05). Allowing for a 10% dropout rate, the target enrolment is set at 1100 participants.

Statistical methods

All analyses will adhere strictly to the intention-to-treat principle. Continuous variables will be presented as means (SD) for normally distributed data or medians (IQRs) for skewed data, while categorical variables will be summarised as counts (percentages). Baseline characteristics between groups will be compared using t-tests or Wilcoxon rank-sum tests for continuous variables, and X² tests or Fisher’s exact tests for categorical variables, as appropriate. Primary and secondary continuous outcomes will be analysed using analysis of covariance, adjusting for corresponding baseline values. Binary secondary outcomes will use multivariable logistic regression adjusted for baseline covariates. Cardiovascular composite events will undergo Kaplan–Meier analysis with log-rank comparisons, supplemented by Cox proportional hazards models. Predefined subgroup analyses will include age (<65 vs. ≥65), sex, baseline LDL-C (<1.8 vs. ≥1.8 mmol/L) and baseline GRACE risk stratification (low, intermediate, high). Additional exploratory subgroup analyses will assess socio-economic factors and patient engagement metrics. Sensitivity analyses for missing data will be conducted using multiple imputation by chained equations. No interim analysis is planned, and a two-sided p value <0.05 denotes statistical significance. Statistical analyses will use the latest R software.

Data de-identification and analyst blinding. A study-independent data manager exports a fully de-identified dataset in which treatment allocation is replaced by non-informative codes (Group A vs Group B). The final statistical analysis plan is approved and archived before the masked dataset is released to an independent statistician. All primary and secondary analyses are conducted on this blinded dataset; the allocation code is disclosed only after the analysis scripts are locked and a blinded interpretation meeting is held with the steering committee.

Process evaluation

A mixed-methods process evaluation will run alongside the effectiveness analysis. Implementation fidelity will be documented with the TIDieR checklist and summarised across the Reach, Effectiveness, Adoption, Implementation and Maintenance domains. Data sources include app-use analytics, voice-call logs and the 10-item System Usability Scale (SUS; range 0–100) administered at the 3 month visit.

Patient and public involvement

After the AIM-CHD software was initially developed, preliminary usability testing was conducted with patients and individuals without medical backgrounds. Their feedback on the design and functionality was collected to inform early-stage refinements. Following the start of the trial, ongoing user feedback is being gathered, including completion of the SUS and verbal feedback provided via telephone follow-up. The software will be further optimised where appropriate based on this input.

Discussion

Given China’s dual challenges of rising cardiovascular disease burden and limited healthcare capacity, mHealth platforms represent a scalable strategy to enhance care delivery and efficiency. The combination of near-universal smartphone adoption, a 97.33% literacy rate and robust internet infrastructure provide fertile ground for deploying mHealth tools—especially among socio-economically disadvantaged or geographically remote populations.^{31 32}

This trial highlights several methodological innovations of the AIM-CHD platform in secondary CHD prevention:

1. AI-enhanced automation: AIM-CHD integrates automated processes across data synchronisation, guideline-based management and closed-loop follow-up monitoring. It is one of the first management platforms to incorporate AI–driven voice follow-up and OCR, substantially reducing administrative burden while enhancing patient engagement.

2. Peri-care data integration: by linking inpatient, outpatient and remote monitoring data, the system enables real-time risk stratification and dynamic protocol adjustment. This continuity fosters precision care while reducing unnecessary medical utilisation.

3. Comprehensive risk management: unlike earlier mHealth solutions that target limited risk factors, AIM-CHD addresses a broader spectrum of modifiable risks. This multifactorial approach increases the likelihood of meaningful clinical benefit.

4. Translational scalability: its modular design and structured intervention logic offer a replicable blueprint for chronic disease management beyond CHD. Automated surveillance, early warnings and decentralised delivery make it especially promising for under-resourced healthcare settings.

Despite these strengths, several limitations must be acknowledged:

1. Single-centre design: the single-centre setting limits generalisability. To address this, we are actively preparing for a multicentre expansion. To overcome HIS interoperability issues, we plan to deploy indirect data entry mechanisms (eg, OCR-based capture of discharge summaries, manual upload of follow-up reports) to enable standardised data acquisition across heterogeneous institutions.

2. Multicomponent intervention attribution: The intervention combines three elements—comprehensive risk factor management, risk-stratified personalised follow-up and AI-assisted adherence enhancement. The synergistic design limits attribution of outcomes to individual components. Future work will focus on component-level analyses to better understand each module contribution. Among these, we hypothesise that comprehensive risk factor management may drive the greatest clinical impact, while risk-based follow-up scheduling improves efficiency, and AI modules may boost short-term follow-up and adherence metrics.

3. Short follow-up duration: the 3 month follow-up may not capture late events such as major adverse cardiovascular events. Longer follow-up is ongoing to evaluate the durability of benefit.

4. Iterative system evolution: ongoing updates to the app interface and AI modules may introduce time-related heterogeneity. We have implemented version tracking and time-stratified usage monitoring to assess potential confounding.

5. Health equity considerations: elderly users and those from lower socio-economic backgrounds may experience lower usability. Targeted solutions such as simplified user interfaces, caregiver support and digital literacy programmes are being piloted.