AI-powered thermal imaging: A privacy-preserving innovation in neonatal care

The key clinical benefits of this AI-based system include more accurate newborn resuscitation timelines, ensuring interventions occur within the critical golden minute. Additionally, automated and unbiased ToB documentation reduces errors caused by manual recording delays, improving medical record accuracy.

In neonatal care, accurate documentation of the Time of Birth (ToB) is critical, as it directly impacts the effectiveness of resuscitation efforts. Approximately 10% of newborns require assistance to breathe, and 5% need proper ventilation, with birth asphyxia remaining a leading cause of neonatal mortality. However, traditional ToB documentation relies on manual recording by healthcare professionals, which introduces inconsistencies due to delays and human error.

A groundbreaking study titled “AI-Based Thermal Video Analysis in Privacy-Preserving Healthcare: A Case Study on Detecting Time of Birth”, conducted by Jorge García-Torres, Øyvind Meinich-Bache, Siren Rettedal, and Kjersti Engan, from the University of Stavanger, Norway, and Laerdal Medical AS, proposes an AI-driven solution to automate ToB detection using thermal imaging. Their approach not only improves accuracy and efficiency but also ensures privacy by avoiding the use of identifiable visual data. The study reports that the AI model achieved 91.4% precision and 97.4% recall, successfully detecting ToB within a 1-second median deviation compared to manual annotations, offering a new standard for neonatal resuscitation documentation.

AI-powered thermal imaging: A privacy-preserving innovation in neonatal care

Traditional methods for ToB documentation depend on manual recording by nurses or midwives, typically with minute-level precision. Such imprecision can negatively impact the ability to follow Neonatal Resuscitation Algorithm Activities (NRAA), which require interventions within the “golden minute” after birth to improve newborn survival rates. To address this, the study presents an AI-powered video-based system capable of detecting ToB automatically, accurately, and in real-time.

Unlike conventional video-based AI models that process RGB or grayscale images, this system uses thermal imaging, capturing heat emissions rather than identifiable facial or bodily features. By leveraging infrared (IR) imaging, the system ensures patient privacy, adhering to regulations like General Data Protection Regulation (GDPR) while preserving crucial birth event data.

The study is part of the NewbornTime project, an initiative aimed at developing a fully automated AI-driven system for newborn care. It builds upon previous research that utilized individual thermal frames for ToB estimation but lacked temporal context. By introducing a spatiotemporal AI model, the researchers improved accuracy by analyzing movement dynamics and temperature variations rather than relying solely on static images.

How AI detects time of birth in thermal video

The study utilized 321 birth videos recorded at Stavanger University Hospital, Norway, using a ceiling-mounted thermal camera. These cameras were triggered when temperatures exceeded 30°C, ensuring that only relevant data was captured. The AI model was trained using a sliding window approach, where each second of the video was analyzed within a 3-second frame sequence.

To ensure precision, the researchers developed an adaptive Gaussian Mixture Model (GMM) normalization technique, which filters out irrelevant temperature variations caused by room temperature fluctuations or automatic sensor calibrations. This method improves the AI’s ability to consistently recognize the warmer temperature of a newborn as it emerges.

The system then applies deep learning-based video analysis using three different convolutional neural network (CNN) architectures: I3D, X3D, and MoViNet. Among these models, MoViNet-A2 outperformed the rest, achieving the highest precision (91.4%) and recall (97.4%), ensuring that it could detect the ToB event with minimal false positives or negatives.

Performance and clinical impact of AI-based ToB detection

The study tested the AI model’s ability to detect ToB accurately and reduce discrepancies in neonatal resuscitation documentation. Results showed that 96% of test cases were correctly identified, with a median absolute deviation of 1 second compared to manual annotations. This is a significant improvement over traditional methods, where manual logging can be delayed due to the chaotic nature of the delivery room.

By implementing a confidence thresholding mechanism, the system also reduces false positives, ensuring that ToB is not mistakenly detected from other movements in the delivery room. The Finite Impulse Response (FIR) filter used in the system further refines the probability scores, reducing noise and improving detection accuracy.

The key clinical benefits of this AI-based system include more accurate newborn resuscitation timelines, ensuring interventions occur within the critical golden minute. Additionally, automated and unbiased ToB documentation reduces errors caused by manual recording delays, improving medical record accuracy. This system also enhances post-birth research and debriefing by providing reliable data for assessing newborn resuscitation performance. Importantly, it maintains privacy compliance as it does not rely on identifiable video footage, ensuring patient confidentiality in sensitive healthcare settings.

Challenges, future improvements, and broader applications

While the system shows great promise, the study highlights challenges and areas for improvement. One issue is that in certain maternal birthing positions (e.g., hands-and-knees delivery), visibility is reduced, making detection more difficult. Future work may involve combining thermal and conventional video for improved robustness.

Another limitation is the potential for missing detections in cases where the newborn is partially obscured by medical staff. A possible solution is to integrate a fallback system using both image-based and video-based AI models, allowing for more comprehensive analysis.

Looking ahead, the researchers plan to incorporate this AI-powered ToB detection system into a larger AI-driven newborn care framework, NewbornTimeline, which will track not just the birth event but also resuscitation activities. This could revolutionize neonatal care by providing real-time decision support for clinicians.

Beyond newborn care, the concept of AI-based thermal video analysis has potential applications in other privacy-sensitive healthcare settings, such as emergency rooms for tracking critical patient movements, elderly care for fall detection, and surgical monitoring for documentation without intrusive video recording.

Advancing newborn care with AI and thermal imaging

The study represents a significant step forward in automating newborn care with AI. By leveraging thermal imaging and deep learning, researchers have developed a privacy-compliant, highly accurate ToB detection system that enhances clinical decision-making and neonatal resuscitation timelines.

With an accuracy rate of over 97%, this system has the potential to revolutionize how newborn resuscitation data is recorded, ultimately leading to better medical outcomes for infants. As AI continues to advance in healthcare, privacy-preserving techniques like thermal imaging will play a crucial role in ensuring both patient confidentiality and improved medical efficiency.

This breakthrough highlights the growing role of AI in transforming neonatal care and paves the way for more reliable, efficient, and ethical applications of AI in healthcare settings worldwide.