Demographic Baseline for Pediatric Forearm Fracture Localization: A Machine- Learning Benchmark Using the PediURF Dataset
Abstract Background Image-based artificial intelligence models for pediatric fracture localization may exploit demographic signals that are correlated with fracture distribution. Quantifying the discrimination available from demographic metadata alone provides a reference for interpreting image-only and multimodal models. Methods We performed a secondary analysis of the open PediURF dataset, preserving its original patient-level training (n = 4,212) and test (n = 1,053) split. Only age and sex were used as predictors. Logistic regression (LR), Gaussian naive Bayes, radial-basis-function support vector machine, random forest, and gradient boosting were tuned exclusively in the training set. Standard training, inverse-frequency cost-sensitive training, and a post hoc equal-prior decision rule applied to standard LR were compared. The primary discrimination measure was macro one-vs-rest ROC-AUC. Test uncertainty was quantified with 2,000 bootstrap resamples. Paired ΔAUC comparisons were Holm-adjusted. Calibration, class-specific recall, and a training-estimated saturated age-by-sex reference were also evaluated. A test-label-derived saturated model was reported only as an optimistic empirical oracle reference, not as a deployable or theoretical upper bound. Results Among 5,265 children with confirmed fractures, 64.1% were distal, 25.1% midshaft, and 10.9% proximal. Standard LR achieved a test macro AUC of 0.704 (95% CI 0.674–0.734), comparable with the training-estimated saturated reference (0.698). Random forest (0.701) and gradient boosting (0.699) were within the prespecified analytic equivalence framework relative to LR, whereas SVM was lower (0.648; ΔAUC − 0.056; Holm-adjusted p < 0.004). Standard models had poor proximal-class recall. Cost-sensitive training increased proximal recall to 0.450 for most algorithms but worsened probability calibration. Modifying the LR decision rule using equal-prior rescaling reproduced the balanced classification profile (balanced accuracy 0.528; proximal recall 0.450) while retaining the original LR probabilities for probability-based discrimination and calibration assessment. Conclusions Age and sex contain moderate information about fracture location in PediURF, but flexible machine-learning algorithms did not materially improve on a parsimonious LR reference. The principal effect of imbalance handling was on class decisions rather than on the amount of demographic information available. This demographic benchmark should be interpreted as a non-imaging reference for future image-only and multimodal studies, not as a diagnostic substitute for radiography.
Authors
- Ali Levent (ORCID: https://orcid.org/0000-0002-3666-1084)
- Canan Aslanyürek (ORCID: https://orcid.org/0000-0001-7513-9953)
- Murat Aslanyürek (ORCID: https://orcid.org/0000-0002-3296-4395)
- Volkan Kizilkaya (ORCID: https://orcid.org/0000-0002-7472-3303)
Institutions
- Kırklareli University (TR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11195635/v1
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint