Deep learning-based detection of cessation of breathing in pre-term infants

Technical Summary: Deep Learning-Based Detection of Cessation of Breathing in Pre-Term Infants

Problem Statement
Apnoea of prematurity (AOP) is a critical complication in pre-term infants, characterized by recurrent episodes of Cessation Of BrEathing (COBE). Current detection in Neonatal Intensive Care Units (NICUs) relies on bedside monitors that trigger alarms based on fixed thresholds for respiratory rate, heart rate, and oxygen saturation. These rule-based approaches suffer from high false positive rates, often failing to capture short or irregular events, which contributes to alarm fatigue. While deep learning has advanced apnoea detection in adults, its application to neonatal populations is constrained by the scarcity of large, annotated datasets and a lack of systematic comparison between modern architectures and signal modalities. Furthermore, existing neonatal studies often utilize heterogeneous segmentation strategies and handcrafted features, complicating direct performance comparisons.

Methodology
The study utilized a clinical dataset comprising approximately 430 hours of recordings from 24 pre-term infants (mean gestational age 31.1 ± 1.8 weeks) monitored in a high-dependency NICU. The dataset included three routinely acquired physiological signals:

1. Impedance Pneumography (IP): Direct measure of respiratory effort.
2. Electrocardiography (ECG): Used to derive ECG-derived respiration (EDR) via R-peak amplitude modulation.
3. Photoplethysmography (PPG): Used to derive a respiratory envelope from pulsatile arterial blood volume variations.

Data Annotation and Preparation:
COBE events were defined clinically as breathing pauses ≥20 seconds, or ≥10 seconds accompanied by bradycardia (<100 bpm) or oxygen desaturation (SpO2 < 80%). An automated screening algorithm identified candidate desaturation events, which were then manually annotated by three independent reviewers (one clinician, two biomedical engineers) using a custom MATLAB interface. The final dataset consisted of 346 COBE events and 608 non-COBE segments. These were segmented into 20-second windows with a 10-second step size, yielding 6,678 analysis windows.

Deep Learning Architectures:
The authors evaluated three families of 1D convolutional neural networks, adapted from image classification architectures to physiological time-series:

1. Shallow CNN: A six-layer architecture based on Urtnasan et al., modified with specific kernel sizes and pooling strategies for neonatal signals.
2. Residual Networks (ResNets): Variants of ResNet-18, ResNet-34, and ResNet-50, adapted for 1D inputs with reduced feature channels.
3. ConvNeXt: A modern architecture adapted for 1D signals, evaluated with default and larger kernel sizes, and reduced feature widths.

The models were tested in unimodal configurations (using IP, EDR, or PPG envelope individually) and multimodal configurations using both early fusion (concatenating inputs at the network stem) and late fusion (concatenating features before the classifier).

Evaluation Protocol:
To prevent data leakage, the dataset was split at the infant level. A five-fold cross-validation was performed on 20 infants, with a completely independent test set comprising data from 4 held-out infants. Performance was assessed using balanced accuracy, sensitivity, precision, F1 score, and Cohen's κ, with class-weighted loss functions addressing class imbalance.

Key Results

• Signal Modality Dominance: Across all architectures, the choice of input signal had a far greater impact on performance than architectural complexity.
  • IP-based models consistently achieved the highest balanced accuracy, ranging from 86.8% to 88.0% on the independent test set.
  • ECG-derived (EDR) models performed significantly lower, with balanced accuracies between 62.6% and 69.7%.
  • PPG-derived models achieved balanced accuracies between 65.1% and 66.4%.
• Architectural Performance: While deeper and more complex models offered marginal improvements, the gains were modest.
  • The ConvNeXt architecture generally outperformed the shallow CNN and ResNets.
  • The best-performing configuration was a ConvNeXt model using late fusion of IP and PPG inputs, achieving a balanced accuracy of 88.7% and an F1 score of 0.75 on the independent test set.
  • Multimodal fusion provided only incremental improvements over IP alone, suggesting that the direct respiratory signal (IP) contains the majority of discriminative information.

Significance and Claims
The paper claims to provide the first systematic evaluation of a 1D ConvNeXt architecture for neonatal COBE detection, alongside controlled comparisons with CNN and ResNet models under identical preprocessing and evaluation protocols. The primary contribution is the demonstration that signal modality is the primary determinant of detection performance in data-constrained neonatal settings, outweighing architectural complexity.

The authors assert that deep learning models applied to routinely monitored NICU signals can achieve reliable detection of COBE events without requiring additional sensing hardware. However, they maintain a modest stance regarding clinical translation, noting that while detection performance is strong, prospective validation of alarm burden within operational workflows is necessary before implementation. The study highlights that the superior performance of IP-based models reflects their direct measurement of respiratory mechanics, whereas ECG and PPG serve as less reliable indirect surrogates in this specific population.