Related Experiment Video
Updated: Jul 6, 2026
Multi-Modal Home Sleep Monitoring in Older Adults
Published on: January 26, 2019
Privacy-Preserving Infant Sleep-Wake Detection in NICUs based on Federated Learning: A Multi-center Clinical Study
Insights
Federated learning enhances infant sleep-wake cycle monitoring in NICUs, improving accuracy by 12-15% while protecting patient privacy. This multi-center approach overcomes limitations of single-hospital training.
Area of Science:
- Neonatal Intensive Care Unit (NICU) research
- Computational neuroscience
- Medical informatics
Background:
- Monitoring infant sleep-wake cycles is vital for assessing physiological and neurodevelopmental status.
- Camera-based infant sleep-wake detection in NICUs has shown promise but faces privacy concerns due to centralized data aggregation.
- Traditional single-hospital training models lack generalizability and raise privacy issues with data transmission.
Purpose of the Study:
- To develop a privacy-preserving, multi-center approach for infant sleep-wake cycle detection using federated learning.
- To evaluate the performance of federated learning models against traditional single-hospital training.
- To address the generalization challenges in developing accurate infant sleep monitoring systems.
Main Methods:
- A multi-center clinical study involving video data from 100 preterm and term infants across four NICUs.
- Distributed model training using federated learning approaches (FedAvg, FedOpt, FedNova).
- Comparison of federated learning models with traditional single-hospital training methods.
Main Results:
- Federated learning significantly outperformed traditional single-hospital training, showing a 12-15% improvement in accuracy, precision, recall, and F1-score.
- The federated learning approach effectively eliminated privacy concerns associated with inter-hospital data transfer.
- Improved model generalization was observed compared to models trained in isolation at a single hospital.
Conclusions:
- Federated learning offers a robust solution for multi-center infant sleep-wake cycle monitoring, enhancing accuracy and generalizability.
- This distributed approach effectively safeguards infant privacy by avoiding direct data transmission between institutions.
- Federated learning represents a significant advancement for developing reliable and secure AI-driven healthcare solutions in NICUs.
Abstract:
Clinically, monitoring infants' sleep-wake cycles provides crucial information of sleep parameters (e.g., sleep duration and frequency), enabling the assessment to their physiological and neuro developmental status. Recently, camera-based infant sleep-wake detection using facial features had been demonstrated in the Neonatal Intensive Care Unit (NICU). However, this approach relied on large-scale data training from different hospitals and ignored the risk of privacy invasion caused by data transmission, as it requires aggregating the data from multiple hospitals at a centralized location for model training. To solve this issue, we organized a multi-center clinical study that uses federated learning to train the model in a distributed way for protecting infant privacy. Specifically, we used video data of 100 preterm and term infants from NICUs of four hospitals to train three separate models with baseline federated learning approaches, including FedAvg, FedOpt, and FedNova. The experiments show that federated learning significantly outperforms traditional single-hospital training, with an improvement of 12-15% in all used metrics including accuracy, precision, recall, and F1-score. More importantly, federated learning can eliminate the privacy issue caused by data transfer between different hospitals, while solving the generalization issue of model training in a single hospital.
Related Concept Videos
01:11Understanding Sleep
The circadian rhythm, a nearly 24-hour cycle, is deeply influenced by environmental light cues. Light exposure directly affects the hypothalamus, which in turn regulates...
01:15REM Sleep Behavior Disorder
RBD is significantly associated with...