454 - Prediction of Bradycardia Episodes in Very Preterm Infants

Poster Number: 454
Publication Number: 454.228

Colm P. Travers, University of Alabama at Birmingham, Birmingham, AL, United States; Arie Nakhmani, University of Alabama at Birmingham, Birmingham, AL, United States; Pravinkumar G. Kandhare, University of California, San Francisco, San Francisco, CA, United States; Nikolay M. Sirakov, Texas A&M University Commerce, Commerce, TX, United States; Deborah J. Laney, University of Alabama School of Medicine, Birmingham, AL, United States; Waldemar A. Carlo, University of Alabama at Birmingham, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States

Background: Bradycardia episodes, frequently associated with apnea, are common in very preterm infants and associated with bronchopulmonary dysplasia and neurodevelopmental impairment. Prediction of bradycardia events before they occur may allow development of novel interventions to reduce the frequency or severity of these episodes, potentially improving outcomes.

Objective: To develop models to predict episodes of bradycardia using heart rate data from patient monitors.

Design/Methods: We included infants less than 29 weeks’ gestation breathing either room air or oxygen without support (via low-flow cannula or incubator) at the neonatal intensive care unit at the University of Alabama at Birmingham. We excluded infants with major malformations, neuromuscular conditions affecting respiration, terminal illness, or decision to withhold or limit support. We collected real-time numeric (1 Hz) and waveform (125 - 500 Hz) cardiorespiratory data from Philips MP70 or MP50 patient monitors using Bedmaster (Anandic Medical Systems Inc., Feuerthalen, Switzerland). Episodes of bradycardia, defined as heart rate < 100 beats per minute for ≥10 seconds, were analyzed with MATLAB using predictors based on computational mathematical models as well as deep learning models. We developed early bradycardia prediction algorithms from 1 to 10 seconds ahead of the real-time data. We used 80% of the data to train and validate and 20% for final testing to determine the true prediction and false alarm rates.

Results: We evaluated data from 134 preterm infants with a mean (SD) gestational age of 27w 0d (10d) and birth weight of 958 (273) grams. There were 2380 days of data included: 57,120 hours or approximately 205 million data points and 26,515 episodes of bradycardia. Nine predictors were compared. The finite difference algorithm predicted 90% of bradycardia events 5 seconds ahead with 20% false alarms and 80% of bradycardia events 5 seconds ahead with a 10% false alarm rate (Figure 1). Linear and linear regression algorithms were also relatively good predictors of bradycardia episodes with acceptable false alarm rates. Deep learning models did not perform as well as the computational mathematical models.Conclusion(s): Bradycardia episodes in spontaneously breathing very preterm infants can be predicted from heart rate signals using computational mathematical models and machine learning. Computational mathematical models including finite difference, linear, and linear regression algorithms had high predictive accuracy for bradycardia episodes with acceptable rates of false alarms.
Figure 1Comparison of true prediction (TPred) versus false alarm (FA) rates (5 seconds ahead) averaged on 500 days for bradycardia prediction models. LSTM - long short-term memory; Bi-LSTM - bidirectional LSTM; GRU - gated recurrent unit; RNN - recurrent neural network; LOCF - last observation carried forward.