Consultant Neonatologist La Paz University Hospital Madrid, Madrid, Spain
Background: Cardiovascular support (CVS) is associated with morbidity and mortality in very low birth weight (VLBW) infants. Treatment response (TR) to CVS occurs in only 50% of the cases and treatment failure (TF) is associated with mortality. Preliminary predictive models based on machine learning (ML) techniques have shown good TR predictive capacity (PAS2023). Objective: We aim to refine these algorithms and validate them using prospective data. Design/Methods: VLBW infants that received CVS during the neonatal period were prospectively included (Feb 2022-Feb 2023). Perinatal, cardiorespiratory, echocardiographic and clinical data were recorded. TR occurred if blood pressure raised at least above the 5th centile for gestational age (systolic, diastolic or mean), or SVC flow > 55 ml/kg/min, while decreased serum lactate and less negative base excess were observed with only the initial CVS within the first 24 hours after treatment started. TR predictive models were refined from our preliminary models. The impact of CVS effectiveness on morbidity and mortality was also analysed. Results: 84 VLBW infants (retrospective n=68 and prospective n=16 infants, [median GA 27 weeks and median BW 864 g] received Dopamine (Dp) (n=48), Dobutamine (Db) (n=30) or Adrenaline (Ad) (n=6) as first line CVS. TR was observed in 45% of the cases (Dp 57%, Db 40%, Ad 33%). Groups were comparable respect to perinatal variables except for higher SNAPPE-II score (p=0.03) and basal serum lactate (p < 0.05), and lower chorioamnionitis (p < 0.05) in TF compared to TR. ML identified that a logistic regression (LR) model using L1 regularization (L1) technique for Db (accuracy 77%), random forests 50 trees 5 leaves for Dp (accuracy 70 %) and LR L1 for Ad (accuracy 83%) achieved the best TR predictions. Twenty-two infants died (TR 10% and TF 42%; p= 0.001). Multiple regression showed that mortality was associated with TF (p=0.006), but not with SNAPPE score or basal serum lactate. Other neonatal morbidities were not different between the study groups
Conclusion(s): Lack of CVS effectiveness increases the risk of mortality in VLBW infants. Refined predictive models for CVS associated TR built with ML techniques can help in decision-making process improving treatment effectiveness and reducing mortality.
