Conformal Tensor State Modeling for Early Neonatal Bilirubin Rebound Prediction After Phototherapy Cessation in Heterogeneous Nurseries

• Qiang Chen

  School of Information Science and Engineering, Yunnan University, Cuihu North Road, Wuhua District, Kunming 650091, China

  Author

• Yifan Gao

  Department of Computer Science, North China University of Technology, Jinyuanzhuang Road, Shijingshan District, Beijing 100144, China

  Author

• Ming Zhao

  School of Data Science and Technology, Heilongjiang University, Xuefu Road, Nangang District, Harbin 150080, China

  Author

Neonatal hyperbilirubinemia is often managed with phototherapy, yet a clinically difficult decision remains after treatment stops: whether bilirubin concentration will rebound enough to require renewed monitoring or retreatment. Early discharge can reduce family burden and hospital crowding, but missed rebound may lead to delayed recognition of unsafe bilirubin rise. This paper reports an empirical study of conformal tensor state modeling for bilirubin rebound prediction after phototherapy cessation. We assembled 46,320 neonatal phototherapy courses from five nursery settings and manually reviewed 6,840 courses with uncertain rebound status. The prediction target was rebound within 24 hours, defined by either retreatment, bilirubin rise above a risk-adjusted threshold, or clinician-documented return for bilirubin management. The proposed model represents each infant as a time-indexed tensor containing bilirubin values, age in hours, gestational age, weight trajectory, feeding pattern, hemolysis indicators, and treatment exposure. A low-rank dynamical encoder estimates latent bilirubin production and clearance states, while split conformal calibration produces nursery-specific prediction sets. Compared with gradient boosting, recurrent neural networks, and ordinary transformer encoders, the proposed approach improved area under the precision-recall curve from 0.421 to 0.566, reduced missed high-risk rebound from 12.8% to 6.1%, and maintained 91.4\% conformal coverage under temporal shift. The model retained moderate specificity, with 14.7 additional observation recommendations per 100 treated infants. Results suggest that time-resolved bilirubin physiology and distribution-aware uncertainty can improve rebound screening without converting all post-phototherapy infants into high-alert cases.

• PDF