Longitudinal Fetal Heart Rate Analysis Identifies Neonates with Metabolic Acidemia Requiring Therapeutic Hypothermia.

Abstract

BACKGROUND Despite widespread use of intrapartum fetal monitoring, rates of fetal brain injury remain unchanged. Neonatal encephalopathy due to hypoxia-ischemia is a leading cause of neonatal morbidity and mortality, often occurring in the absence of acute intrapartum events. Current fetal heart rate interpretation guidelines rely on static, cross-sectional assessments and offer limited guidance for managing "indeterminate" patterns. Machine learning provides an objective, longitudinal analysis of the entire intrapartum fetal heart rate tracing and may enhance the identification of fetuses at risk for neurological injury. OBJECTIVE To evaluate whether a longitudinal machine learning-based analysis of abnormal fetal heart rate patterns-and their duration-throughout labor can identify neonates requiring therapeutic hypothermia among those born with acidemia. STUDY DESIGN We conducted a retrospective case-control study of neonates born with metabolic acidemia (umbilical or neonatal arterial pH ≤7.15 or base deficit ≥10 mmol/L) who met clinical, biochemical, and neurologic criteria for therapeutic hypothermia (study group). These were compared to a control group of neonates who met clinical and biochemical criteria, but did not meet neurologic criteria for therapeutic hypothermia. Neonates with acute sentinel events were excluded. Maternal and neonatal characteristics were abstracted from electronic medical records. Fetal heart rate tracings for the entire labor course were extracted from the perinatal monitoring system for machine learning analysis. The optimal reading window size was determined by incrementally increasing fetal heart rate tracing reading window sizes by 5-minute intervals until optimal model performance was achieved. After the machine learning-based model was trained for anomaly detection, a two-step process using Isolation Forest followed by Random Forest Classifier was applied to label each window as "normal" or "abnormal." The primary outcome was the proportion of fetal heart rate windows labeled as "abnormal" during the entire intrapartum course for each fetus. T-tests, Wilcoxon rank-sum tests, chi-square tests, or Fisher's exact tests were used as appropriate. RESULTS There were 44 mother-infant pairs included, with 22 in each group. There were no significant differences between the study and control groups in umbilical cord pH or base deficit levels. Maternal and neonatal characteristics, duration of tracing analyzed, intrapartum events, and delivery outcomes were similar between groups. The study group had significantly lower APGAR scores, higher Sarnat scores, increased seizure incidence, elevated initial AST and ALT levels, and longer lengths of stay (all p < .03). The proportion of "abnormal" 90-minute intrapartum windows was significantly higher in the study group (73.3% ± 15.2%) compared to the control group (4.9% ± 3.3%; p < .001). CONCLUSIONS The higher proportion of "abnormal" 90-minute windows in cases requiring therapeutic hypothermia provides evidence that the duration of abnormal patterns is a key indicator of significant fetal compromise. Our findings also suggest that traditional perinatal clinical characteristics may have limited predictive value for identifying fetuses at risk of neurologic injury. Continuous, objective analysis of the entire intrapartum tracing may improve detection of evolving fetal compromise, enhance interpretation of fetal heart rate patterns, and reduce the rate of unnecessary interventions.