Bouzalas, Ioannis (2026) Predicting microbiologically confirmed neonatal bacterial infection in general neonatal care using a transformer-based survival model. thesis, Medicine.
Full text available on request.Abstract
Background Neonatal infections represent a major challenge in newborn care, with delayed diagnosis and treatment leading to adverse outcomes. Early identification is difficult because clinical signs are often subtle and existing diagnostic tests have limited accuracy and delayed availability. Machine learning approaches offer the potential to enhance early detection by identifying complex patterns in routinely collected clinical data that are not readily apparent to clinicians, enabling earlier risk stratification and more timely intervention. This pilot study primarily focuses on applying machine learning models to predict neonatal infections in settings outside the neonatal intensive care unit (NICU). Methods In this retrospective study, data from 344 neonates admitted to the general neonatal ward were used to develop a transformer-based survival model to predict culture-proven bacterial infection. An accelerated failure time model with Weibull parameterisation estimated infection risk within 24 hours after the last observation. Data were split at the patient level into training (80%) and test (20%) sets with five-fold cross-validation. Model performance was assessed using discrimination and calibration metrics. Results The model achieved strong discrimination and calibration, with C-indices of 0.85 in cross-validation and 0.83 on independent testing. Time-dependent performance remained high over 24 hours, and predictions were driven primarily by dynamic physiological characteristics, particularly oxygen saturation, heart rate, and temperature. Conclusions This exploratory study demonstrates the potential of transformer-based models to accurately predict neonatal infection risk at the individual level. These preliminary findings highlight the feasibility of using such models to support decision-making while maintaining interpretability and clinical relevance
| Item Type: | Thesis (UNSPECIFIED) |
|---|---|
| Supervisor name: | Ploegstra, Dr. M.J. MD PhD |
| Faculty: | Medical Sciences |
| Date Deposited: | 02 Nov 2026 13:25 |
| Last Modified: | 02 Nov 2026 13:25 |
| URI: | https://umcg.studenttheses.ub.rug.nl/id/eprint/3993 |
Actions (login required)
![]() |
View Item |
