Table 1.

Selection criteria of predictive modelling studies in PICOS format.

Criteria	Participants (P)	Intervention (I)	Comparison (C)	Outcomes (O)	Timeframe (T)	Settings (S)	Other limitations
Inclusion criteria	Patients with a confirmed diagnosis of aortic dissection (AD). AD Type: Any anatomical classification (Stanford Type A or B; DeBakey Type I, II, or III). Phase of Disease: Both acute and chronic/subacute phases were considered, provided the study clearly defined the cohort.	Studies that developed and/or validated (internally or externally) a risk prediction model for outcomes in aortic dissection (e.g., in-hospital mortality, composite adverse events) using machine learning (ML) algorithms*.	Not applicable (Systematic review of prediction model studies, typically without a direct comparator intervention).	Studies reporting at least one quantitative measure of model performance. Mortality: In-hospital, 30-day, or long-term (e.g., 1-year, 3-year) mortality. Specific Complications: e.g., Acute kidney injury (AKI), neurological complications (stroke, cerebral complications), gastrointestinal bleeding, mesenteric malperfusion, pulmonary complications. Composite Outcomes: e.g., Major adverse events (MAEs), a composite of mortality and aortic-related events. Performance Metrics: Primarily discrimination† (e.g., AUC/C-statistic), but also accuracy, sensitivity, specificity, and calibration measures (e.g., calibration plots, Brier score, Hosmer-Lemeshow test).	From each database's inception to September 30, 2025	All clinical settings involved in the management of aortic dissection, including emergency departments, cardiology wards, intensive care units, and surgical centers, across any country income level.	Language: English and Chinese. Study Type: Full-text articles in peer-reviewed journals.
Exclusion criteria	Patients with other aortic conditions (e.g., aortic aneurysm, intramural hematoma without dissection, penetrating aortic ulcer). Patients with traumatic aortic injury. Studies focused exclusively on non-aortic cardiovascular diseases.	Studies that only used traditional statistical methods* (e.g., logistic regression) without an ML component; studies that applied existing models without developing or validating a new one.	N/A	Studies that did not report quantitative performance metrics for the prediction model.	Studies published after the search date.	No restrictions based on clinical setting.	Reviews, editorials, conference abstracts, case reports (<10 patients), animal studies, and studies in languages other than English or Chinese.

^*We defined “machine learning” (ML) models as those employing algorithms capable of learning complex, non-linear relationships from data without explicit rule-based programming, including but not limited to random forests, support vector machines, gradient boosting machines (e.g., XGBoost, LightGBM), neural networks, and ensemble methods. In contrast, “traditional statistical models” were defined as regression-based approaches relying on pre-specified functional forms and assumptions about the data distribution, such as standard (unregularized) logistic regression or Cox proportional hazards models.

^† In this review, the terms C-statistic (concordance statistic) and AUC (area under the receiver operating characteristic curve) were treated as equivalent measures of discriminative performance. For binary outcomes (e.g., early mortality, AKI), the AUC was derived directly from ROC analysis. For time-to-event outcomes (e.g., long-term mortality), studies reported the C-index, which is analogous to the AUC for censored data and quantifies the probability that a model correctly ranks the survival times of two randomly selected patients. Both metrics were pooled together under the common label of “C-statistic” in our meta-analysis, as they share the same interpretation (range 0.5–1.0, with higher values indicating better discrimination) and are mathematically equivalent when no censoring is present. To ensure comparability, we only included C-indices derived from models that were evaluated using time-dependent ROC or Harrell's C-index with appropriate handling of censoring. Studies reporting only Harrell's C without confidence intervals or standard errors were excluded unless these could be calculated from reported sample sizes and event counts using validated methods (e.g., based on the standard error formula for the C-index).