Clinical machine learning for mortality classification and deep learning for chest radiograph classification in neonatal pulmonary hemorrhage: a retrospective case–control study
Background Neonatal pulmonary hemorrhage (PH) is a life-threatening condition associated with substantial morbidity and mortality. This study evaluated two separate artificial intelligence approaches: machine-learning (ML) models using clinical and demographic variables for in-hospital mortality classification and deep-learning (DL) models for classifying chest radiographs according to clinically recognized PH status. Methods This retrospective, single-center case–control study included 101 neonates (50 with PH and 51 controls); 42 died during hospitalization. Controls were selected considering broad similarity in gestational age and birth weight, without formal matching or predefined tolerances. Clinical models classified all-cause in-hospital mortality using repeated cross-validation, fold-specific preprocessing, and a held-out test set ( n = 26) for exploratory internal assessment. Separately, 204 chest radiographs (102 per group) were evaluated using convolutional neural networks with repeated patient-level grouped cross-validation. Radiographs from PH cases were obtained within 48 h before clinically recognized bleeding; control radiographs were selected at corresponding postnatal ages. An audit for non-anatomical shortcuts prompted image harmonization to reduce acquisition-related bias. Results The prespecified logistic regression model incorporating PH status, gestational age, and birth weight achieved a cross-validated area under the receiver operating characteristic curve (AUC) of 0.80 (95% CI, 0.70–0.88) and a test-set AUC of 0.86 (95% CI, 0.67–1.00), with 73% sensitivity and specificity. There was no evidence that another model outperformed the primary logistic regression model. Before harmonization, non-anatomical image properties alone distinguished radiograph groups (AUC, 1.00); after harmonization, negative-control performance fell to near-chance levels (AUC, 0.44–0.59). On harmonized radiographs, the pre-specified final-analysis ensemble achieved a patient-level AUC of 0.78 (95% CI, 0.68–0.87), while ConvNeXt-Tiny achieved a repeat-averaged AUC of 0.80 (95% CI, 0.70–0.88). Conclusions A parsimonious clinical model classified in-hospital mortality with performance comparable to more complex models. Harmonized pre-event radiographs showed moderate discrimination according to PH status. The analyses were separate; no integrated clinical–imaging model was evaluated. Radiographic classification does not establish a PH diagnosis or prospective predictive utility. These preliminary findings require validation in larger, independent cohorts.
Authors
- Aslı Okbay Güneş (ORCID: https://orcid.org/0000-0003-4041-0648)
- İhsan Yıldırım (ORCID: https://orcid.org/0000-0002-7126-7963)
- Selin Ercan (ORCID: https://orcid.org/0000-0002-9356-3042)
- Fatih İşleyen (ORCID: https://orcid.org/0000-0002-5015-621X)
- Mete Yağanoğlu (ORCID: https://orcid.org/0000-0003-3045-169X)
- Mehmet Kılıç (ORCID: https://orcid.org/0000-0001-5950-886X)
- Halil Aslan (ORCID: https://orcid.org/0000-0001-8111-121X)
Publication Details
- Journal
- BMC Pediatrics
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1186/s12887-026-08011-x
- Primary Topic
- Neonatal Respiratory Health Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00