Machine learning model with output correction: Towards reliable bradycardia detection in neonates

Area of Science:

• Biomedical Engineering
• Neonatal Medicine
• Artificial Intelligence in Healthcare

Background:

• Bradycardia is a common and serious condition in premature infants, requiring accurate detection for timely intervention.
• Excessive false alarms from current detection systems erode clinician trust and disrupt patient care workflows.
• Machine learning (ML)-based clinical decision support tools face challenges with false alarm rates.

Purpose of the Study:

• To develop and evaluate an ML-based approach with an output correction element to minimize false alarms in bradycardia detection.
• To improve the reliability and accuracy of automated bradycardia detection in preterm infants.
• To reduce alarm fatigue and enhance healthcare professionals' engagement with clinical decision support tools.

Main Methods:

• Applied five ML-based autoencoder techniques: RNN, LSTM, GRU, 1D CNN, and a CNN-LSTM combination.
• Utilized approximately 440 hours of real-time preterm infant data for analysis.
• Incorporated an output correction element into the ML models to reduce false alarms.

Main Results:

• The proposed approach achieved high performance metrics, including AUC-ROC of 0.978, AUC-PRC of 0.73, recall of 0.992, and F1 score of 0.671.
• Demonstrated a significant reduction in false positive rate (FPR) to 0.007.
• Achieved a 36% reduction in false alarms compared to methods without the output correction approach.

Conclusions:

• The developed ML approach effectively minimizes false alarms in bradycardia detection for preterm infants.
• Reducing false alarms is crucial for maintaining clinician trust and improving the quality of patient care.
• This study highlights the need for solutions that alleviate alarm fatigue and promote healthcare professional engagement.