Table 4 Meta-regression analysis of factors associated with AUC performance.

From: Global performance of machine learning models to predict all-cause mortality: systematic review and meta-analysis

Variable

Comparison

β coefficient

Standard error

p-value

Country income level

High-income versus low/middle-income

0.0561

0.1041

0.591

Population type

General versus disease-specific

− 0.1919

0.0257

< 0.001

Sample size

 ≥ 2000 vs < 2000 participants

0.0402

0.1609

0.803

TRIPOD + AI score

 ≥ 35 vs < 35 points

− 0.1341

0.0312

< 0.001

Algorithm type

Neural Networks versus tree-based

0.1355

0.0267

< 0.001

Algorithm type

Linear/statistical versus tree-based

− 0.1067

0.1097

0.3334

Algorithm type

Ensemble/hybrid versus tree-based

− 0.0043

0.1057

0.9674

Algorithm type

Other models versus tree-based

− 0.0740

0.2242

0.7421

Confidence interval

Imputed versus non-imputed CI

0.1090

0.0410

0.0093

Outcome prevalence

20–39% versus 0–19%

0.0818

0.0425

0.0578

Outcome prevalence

40% or more versus 0–19%

− 0.0253

0.0735

0.7318

  1. Results of univariate meta-regression analyses examining the relationship between study characteristics (country income level, population type, sample size, quality score, algorithm type, confidence interval imputation, and outcome prevalence) and AUC performance, showing β coefficients, standard errors, and p-values.
  2. Note β coefficient represents the difference in AUC between comparison groups; Reference categories: Low/Middle-income countries, Disease-specific populations, < 2000 participants, < 35 TRIPOD + AI points, Tree-based models; Positive β indicates higher AUC in the first group; negative β indicates lower AUC in the first group; Other models category includes algorithms used by single studies: DSM (n = 1), DLS-MSM (n = 1), ICISS (n = 1), Support Vector Machine (n = 1), and Bayesian Network (n = 1).
Back to article page