An artificial intelligence model integrating culprit lesion diagnosis and risk assessment for acute coronary syndrome.

Background: An integrated machine learning (ML) approach capable of both diagnosing acute coronary syndrome (ACS, encompassing myocardial infarction and unstable angina) and predicting future ACS risk within defined optimal timeframes remains an unmet need in cardiovascular risk stratification.

Methods: We conducted a multicentre cohort study in China between January 2012 and December 2021. The derivation cohort (cohort 1+ cohort 2), consisting of a training, validation, and independent external test set, retrospectively included 1854 patients from four hospitals who underwent coronary computed tomography angiography (CCTA) and received a definitive diagnosis of ACS or chronic coronary syndromes (CCS) within 7 days. The diagnostic performance of five commonly used ML algorithms for identifying ACS culprit lesions was developed and compared within this cohort. After selecting the optimal ML model, its performance was further validated in a single-center prospective cohort (cohort 3) and a nested case-control cohort (cohort 4) from two multicenter prospective cohorts to assess its ability to predict future ACS risk stratification (≥30 days).

Findings: The derivation cohort comprised 1854 participants, among whom 281 experienced new-onset ACS within 7 day. The single-centre prospective cohort included 563 participants, of whom 23 developed ACS during follow-up (median 60.0 months). The nested case-control cohort included 202 participants, of whom 43 developed ACS during follow-up (median 28.0 months). In the derivation cohort, the random forest (RF) model demonstrated the best diagnostic performance and robustness among the five ML algorithms for detecting ACS culprit lesions, achieving an area under the receiver operating characteristic curve (AUC) of 0.85 across the training (95% confidence interval [Cl]: 0.82-0.89), validation (95% Cl: 0.80-0.90), and test sets (95% Cl: 0.78-0.92). In two prospective cohorts, Kaplan-Meier curves showed that the RF model successfully stratified the risk of future ACS events (both log-rank < 0.001). The time-dependent AUC curve further indicated that the optimal validity period for the RF model derived from the culprit lesion was 2 years. The model's ability to distinguish ACS events within this period was superior to that of the stenosis severity model (AUC: Cohort 3: 0.79 [95% Cl: 0.66-0.93] vs. 0.67 [95% Cl: 0.52-0.82], = 0.02; Cohort 4: 0.67 [95% Cl: 0.56-0.77] vs. 0.60 [95% Cl: 0.52-0.68], = 0.05). Additionally, the RF model showed a significant improvement in performance compared to the stenosis severity model in both Cohorts 3 and 4 (all net reclassification improvement [NRI] values > 0).

Interpretation: The ACS culprit lesion model developed using RF algorithms demonstrated superior diagnostic performance and was effective for short-term (2-year) ACS risk stratification.

Funding: This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0521700), and the National Natural Science Foundation of China (No. 82441019 for L.J. Z.).