Fetal Heart Rate Analysis from a Multi-task Learning Perspective with Gaussian Processes

Assessments of fetal heart rate tracings by obstetricians suffer from inter- and intra-observer variability whereas computerized fetal heart rate analysis lacks consensus on labels that have diagnostic capability. There are different measurements that carry important information about fetal well-being, although in the literature the most adopted one has been the umbilical cord blood pH value at birth. In this paper, instead of relying on pH-based labeling only, we propose Gaussian process-based multi-task learning that is able to learn multiple fetal well-being measurements simultaneously by explicitly modeling similarity between the tasks. We tested the proposed approach with different intrapartum databases on both regression and classification tasks. Our experimental results show that the proposed approach achieves superior performance compared to popular single-task learning models for fetal heart rate analysis.