Diagnostic Performance of Artificial Intelligence Algorithms in Prediction of Acute Coronary Syndrome Based on White Blood Cell Properties

The AI-ACS clinical trial is an observational, prospective, single-center case-control study conducted at the Medical University of Graz. The trial is designed to assess the diagnostic performance of AI algorithms using hematology data to predict or exclude ACS in adult subjects.

The primary focus of the study is the use of WBC data generated by routine hematology testing. In addition, EC and/or TC data may be used, where available, to explore whether these data can predict or exclude ACS independently or improve WBC-based ACS prediction. The diagnostic performance of the AI algorithms will be compared with high-sensitivity cardiac troponin (hs-cTn), and the performance of combined AI and hs-cTn approaches will also be evaluated.

The AI-ACS trial consists of two main phases: training of AI models and testing of AI models.

For training of AI models, subjects will be assigned to the control cohort, case cohort, supplementary cohort, or rule-out cohort. The control cohort includes subjects with suspected ACS but exclusion of a culprit lesion during coronary angiography. The case cohort includes subjects with suspected ACS and identification of a culprit lesion during coronary angiography. The supplementary cohort includes subjects with no or stable angina pectoris and no indication for revascularization during coronary angiography. The rule-out cohort includes subjects with suspected non-ST elevation ACS and NSTEMI rule-out who did not undergo coronary angiography within 72 hours. WBC data from the control, case, and supplementary cohorts will be used to train AI models, while WBC data from the rule-out cohort may be used to optimize AI training by semi-supervised learning. Whenever available, EC and/or TC data may also be used for exploratory AI training, either alone or in combination with WBC data.

For testing of AI models, a separate all-comer cohort will be used. This cohort includes subjects presenting to the emergency department with suspected ACS. WBC and/or EC/TC data from the all-comer cohort will be used to evaluate the diagnostic performance of trained AI models. Data used for testing will not be included in the training set, in order to avoid data leakage and to ensure unbiased evaluation of model performance.

The anticipated maximum number of subjects to be recruited is 3,350. Of these, 2,350 subjects will be recruited for training of AI models and 1,000 subjects will be recruited for testing of AI models. The training population includes the control cohort, case cohort, supplementary cohort, and rule-out cohort. The testing population consists of the separate all-comer cohort.

Hematology data are collected from routine blood tests performed as part of usual clinical care, using hematology analyzers. Clinical assessment, ECG interpretation, hs-cTn measurement, coronary angiography, and treatment decisions are performed according to current clinical guidelines and are independent of study participation. The presence or absence of ACS is determined based on clinical diagnostic procedures, including coronary angiography and review board evaluation where applicable.

The diagnostic performance of the AI models will be evaluated using receiver operating characteristic (ROC) curve analysis, area under the ROC curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value. The diagnostic performance of hs-cTn and combined AI plus hs-cTn approaches will also be assessed. Training of AI models is planned to be repeated at regular intervals as the number of included datasets increases.

The study includes procedures for data validation, source data verification, data management, and quality control to support the accuracy, completeness, and integrity of the data collected. Missing or inconsistent data will be addressed according to the statistical analysis and data management procedures defined in the protocol.