NewbornTimeLine: Automated Video-Based Timelines for Neonatal Resuscitation in a Hospital Pilot Study

Accurate documentation of neonatal resuscitation events is critical for quality improvement and evidence-based research, yet current manual methods are labor-intensive and lack temporal precision. This paper presents NewbornTimeLine, an AI-based system that automatically generates detailed timelines of birth and resuscitation activities from thermal and visible light video recordings. The system consists of (1) a locally deployed activity recognition system for automated video analysis, using thermal imaging for privacy-preserving time-of-birth detection and visible light video for Neonatal Resuscitation Algorithm (NRA) activity recognition, and (2) a web-based dashboard that provides clinicians with access to both manually annotated and AI-generated timelines. For time-of-birth detection, the proposed two-stream fusion architecture achieves 93.80% accuracy within a 10-second tolerance and 100% birth identification. For NRA activity recognition, the ROI-centered MoViNet approach achieved weighted-average and macro-averaged F1- scores of 0.97 and 0.77, respectively, for key resuscitation activities, including ventilation, stimulation, and suction. The web-based dashboard enables dataset exploration, timeline visualization, and comparison between automated and manual documentation. NewbornTimeLine was deployed as a single-hospital pilot, demonstrating that automated timeline generation for neonatal resuscitation is feasible in a single-centre setting. The system may further support retrospective analysis, clinical debriefing, and quality improvement research.