797 - A Machine Learning Approach to Intrapartum Fetal Heart Rate Tracing Analysis to Identify Neonates Requiring Therapeutic Hypothermia

Publication Number: 797.3641

Dillon S. Joiner, University of Florida College of Medicine, Alachua, FL, United States; Lauren Silva, University of Florida, Gainesville, FL, United States; Jaber Qezelbash-Chamak, University of Florida, Gainesville, FL, United States; Michael D. Weiss, University of Florida College of Medicine, Gainesville, FL, United States; John C. Smulian, University of Florida College of Medicine, Gainesville, FL, United States; Karen T. Hicklin, Icahn School of Medicine at Mount Sinai, New York, NY, United States; Marie Berg, University Of Florida, University Of Florida College Of Medicine, FL, United States

Background: Approximately one-third of cases of moderate to severe neonatal encephalopathy (NE) due to hypoxia-ischemia (HI) are attributable to an identified sentinel event during labor. High rates of intra- and interobserver variability in the interpretation of category II intrapartum fetal heart rate tracing (FHRT) make this borderline classification an opportunity for improved identification of at-risk cases.

Objective: Utilize a machine learning (ML) approach to analyze category II FHRT to identify neonates who developed moderate to severe NE and were treated with therapeutic hypothermia (TH).

Design/Methods: This retrospective case-control study included neonates born at a single academic institution and admitted to the NICU with low cord pH (≤7.15), category II FHRT, gestational age ≥35 weeks, and birth weight ≥2 kg. They were assessed for NE to determine the need for TH. The "cooled" (C) group met TH criteria based on NE from a modified Sarnat exam, while the "not cooled" (NC) group did not meet these criteria and did not receive TH. Cases having a sentinel event other than FHRT anomalies were excluded. The FHRT for eligible neonates was analyzed to develop a ML model using a sliding time window. Isolation Forest was used to identify and cluster windows with unusual patterns as anomalies or typical patterns as normal. A Random Forest classifier was then trained to predict abnormal pattern labels. The ML model’s performance was evaluated by calculating the percentage of anomalous time windows during the entire FHRT for cases in both subject groups and compared by student t-test.

Results: Of 479 cases screened, 44 cases (22 [C] and 22 [NC]) were included, with both groups showing similar maternal and neonatal clinical characteristics, including maternal age, parity, race/ethnicity, comorbidities, pregnancy complications, gestational age at delivery, birthweight, infant sex, and cord arterial pH. Using a 90-minute sliding time window, the ML model identified a significant difference in the proportion of windows with abnormal patterns between the NC and C groups (4.7% vs 73.3%, p < 0.001). The ML classifier applied to these 90-minute windows achieved 100% accuracy in distinguishing between C and NC cases.

Conclusion(s): The ML model utilizing a sliding time window accurately identified neonates requiring TH based solely on intrapartum FHRT in a cohort where risk factors were comparable between the two groups. This approach shows promise in assessing category II FHRT, with the potential to identify cases at risk for NE due to HI and enable timely intervention.

Table 1
Characteristics of the cooled and not cooled population.

Figure 1
Percentage of windows predicted as "cooled" for cooled (C) and not cooled (NC) groups.

Table 1
Characteristics of the cooled and not cooled population.

Figure 1
Percentage of windows predicted as "cooled" for cooled (C) and not cooled (NC) groups.