Improving central apnea detection in preterm infants: Neural network strategies and clinical impact

BACKGROUND AND OBJECTIVE: Preterm infants are characterized by immature cardiorespiratory systems and require continuous monitoring of physiological signals in neonatal intensive care units (NICUs) to assess their clinical condition and return alarms in critical situations. However, many alarms are false or clinically irrelevant, leading to alarm fatigue for nurses and clinicians. A particularly high false alarm rate is reported for central apneas (CAs), with precision as low as 0.35. This study proposes neural networks to improve CA detection by increasing precision.

METHODS: We used a reference dataset of 10 preterm infants (951 annotated CAs across 480 h) and a hold-out dataset of 10 preterm infants characterized by fewer CAs (254 annotated CAs across 480 h). CA detection models were developed using features extracted from the electrocardiogram, chest impedance, and peripheral oxygen saturation, considering four consecutive 30-second moving windows with a 5-second time shift (four-window sets) as individual records. These were trained, optimized, and tested using different neural network architectures, hyperparameter tuning, and leave-one-patient-out cross-validation. Their evaluation was performed through different metrics, including precision, computed by considering all CAs and apneic events of different origins, and sensitivity for all CAs, CAs paired with bradycardia (heart rate, HR ≤ 80 bpm), and CAs paired with low HR (80 bpm < HR ≤ 100 bpm).

RESULTS: The best-performing CA detection model, using convolutional neural networks (CNN), was trained and tested on the reference dataset. By retaining alarms detected in 5 four-window sets, it achieved a precision of 0.54 and a sensitivity of 0.75 for all CAs, 0.87 for CAs paired with bradycardia, and 0.88 for CAs paired with low HR, outperforming results found in previous studies and current clinical practice. Testing on the hold-out dataset showed a decrease in precision and sensitivity for all CAs, equal to 0.28 and 0.69, respectively. However, the sensitivity for CAs paired with bradycardia and with low HR remained high at 0.83 and 0.85.

CONCLUSIONS: Larger datasets with more diverse CA distributions are needed to enhance generalization. Neural networks demonstrated potential to improve CA detection and reduce false alarms in NICUs, supporting more accurate monitoring strategies for preterm infants.