AI-Based Thermal Video Analysis in Privacy-Preserving Healthcare: A Case Study on Detecting Time of Birth

TL;DR

The study tackles the challenge of precise Time of Birth documentation by introducing a privacy-preserving, AI-driven system that analyzes thermal video to detect ToB. By combining GMM-based normalization, sliding-window spatiotemporal CNNs (including MoViNet variants), and FIR-based post-processing, the approach achieves high ToB-detection precision (≈91.4%), recall (≈97.4%), and a median ToB error of about 1 second. The method demonstrates strong performance on a 321-video thermal dataset while preserving privacy, and shows potential to automate NRAA timelines through the NewbornTimeline platform. Practically, this work enables accurate, privacy-conscious ToB logging and broader application to privacy-sensitive clinical video documentation.

Abstract

Approximately 10% of newborns need some assistance to start breathing and 5\% proper ventilation. It is crucial that interventions are initiated as soon as possible after birth. Accurate documentation of Time of Birth (ToB) is thereby essential for documenting and improving newborn resuscitation performance. However, current clinical practices rely on manual recording of ToB, typically with minute precision. In this study, we present an AI-driven, video-based system for automated ToB detection using thermal imaging, designed to preserve the privacy of healthcare providers and mothers by avoiding the use of identifiable visual data. Our approach achieves 91.4% precision and 97.4% recall in detecting ToB within thermal video clips during performance evaluation. Additionally, our system successfully identifies ToB in 96% of test cases with an absolute median deviation of 1 second compared to manual annotations. This method offers a reliable solution for improving ToB documentation and enhancing newborn resuscitation outcomes.

Figures (5)

• Figure 1: Overview of our proposed video-based, AI-driven Time of Birth detector. Birth episodes are recorded by a thermal camera installed on the ceiling. First, thermal videos undergo adaptive normalization using Gaussian Mixture Models, followed by trimming into video clips. At each second, a sliding window containing the previous 25 frames (3 seconds) is utilized as input $x(t)$ to our model for prediction. All the model's output scores $\hat{y}(t)$ are then concatenated and post-processed using a Finite Impulse Response (FIR) filter with a rectangular window. Finally, Time of Birth $\hat{T}_{birth}$ is inferred based on the model's confidence in successfully detecting the birth in the video clips.
• Figure 2: Illustration of thermal frames with Max-Min normalization (first row) and our proposed GMM normalization (second row).
• Figure 3: Visualization of the ToB probability score in four thermal videos in the range $\pm$10 minutes centered around the manual annotated ToB (in blue). In red, the raw score generated by our MoViNet-A2. In black, the postprocessed score using a FIR filter.
• Figure 4: FPR
• Figure 5: ToB difference