Machine Learning Performance for Aortic Dissection Risk Prediction

Forty studies were systematically reviewed and meta-analyzed to evaluate machine learning models for predicting adverse outcomes in aortic dissection. The analysis synthesized performance metrics like C-statistics for outcomes including early mortality, long-term mortality, and acute kidney injury. It was conducted by Yijun Mao following PRISMA guidelines, with searches up to September 30, 2025.

Use Cases

• Compare pooled C-statistics for early mortality (0.891) and long-term mortality (0.847) across different ML algorithms like random forest and XGBoost.
• Analyze the relationship between PROBAST risk-of-bias assessments (100% high/unclear) and reported model performance metrics.
• Investigate reporting adherence gaps, such as the 82.5% unreported sample size justifications, against model AUC values.
• Assess heterogeneity metrics (I² = 61.3–78.8%) in meta-analyses for complication-specific predictions like AKI.

Strengths

• Meta-analysis synthesizes data from 40 included studies following PRISMA and CHARMS guidelines.
• Reports specific pooled performance metrics, such as a C-statistic of 0.891 for early mortality prediction.
• Provides detailed methodological critique, including that 100% of studies were at high or unclear risk of bias per PROBAST.

Limitations

• The dataset is a 23.3 KB document summarizing review findings, not the raw study data, limiting direct analysis.
• Significant heterogeneity (I² up to 78.8%) and high risk of bias in all included studies affect result generalizability.
• Critical model reporting is incomplete, with 72.5% of studies lacking full model specifications and only 5 of 40 performing external validation.

Provenance

Source

figshare, authored by Yijun Mao.

Collection Method

Systematic review and meta-analysis of six databases (PubMed, Web of Science, Cochrane Library, Embase, CNKI, Wanfang).

Time Range

Literature search from database inception to September 30, -2025.

Freshness

Last updated March 26, 2026, with literature search coverage up to September 30, -2025.

Geography

null