444 - Classifying Multi-center Ventilation Strategies for Severe Bronchopulmonary Dysplasia by Hierarchical Clustering Analysis

Poster Number: 444
Publication Number: 444.431

Matthew Kielt, Nationwide Children's Hospital, Columbus, OH, United States; Leon D. Hatch, Vanderbilt University Medical Center, Nashville, TN, United States; Jonathan C. Levin, Boston Children's Hospital, Boston, MA, United States; Natalie Napolitano, Children's Hospital of Philadelphia, Philadelphia, PA, United States; Christopher D. Baker, University of Colorado School of Medicine, Aurora, CO, United States; Laurie C. Eldredge, University of Washington School of Medicine, Seattle, WA, United States; Sharon McGrath-Morrow, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Joseph M. Collaco, Johns Hopkins Medical Institutions, Baltimore, MD, United States; Steven Abman, University of Colorado School of Medicine, denver, CO, United States; Khanh Lai, University of Utah School of Medicine, Salt Lake City, UT, United States; Rebecca S. Rose, Indiana University/Riley Hospital for Children at IU Health, Indianapolis, IN, United States; Martin Keszler, Alpert Medical School of Brown University, Providence, RI, United States; Richard Sindelar, Department of Women's and Children's Health, Uppsala University, Uppsala, Uppsala Lan, Sweden; Leif D. Nelin, Nationwide Children's Hospital, Coumbus, OH, United States; Robin L. McKinney, Hasbro Children's Hospital at Rhode Island Hospital, Providence, RI, United States

Background: Evidence-based ventilation strategies for children with severe bronchopulmonary dysplasia (sBPD) remain unknown due to patient-level heterogeneity, significant inter-center practice variation, and the absence of sufficient research. Determining whether current ventilation approaches cluster as specific sBPD strategies may better characterize care and enhance the design of clinical trials.

Objective: To test the hypothesis that unsupervised clustering analysis of respiratory support data would identify distinct ventilation strategies in a multi-center cohort of children with sBPD.

Design/Methods: We performed a secondary analysis of a multi-center point prevalence study of children with sBPD treated with invasive mechanical ventilation (type 2 sBPD). We clustered the cohort by mean airway pressure (MAP), positive end expiratory pressure (PEEP), set respiratory rate, and inspiratory time (Ti) using Ward’s hierarchical clustering analysis (HCA). Optimal cluster number (k) was determined via dendrogram visualization. We determined the strength of the clustering structure quantitatively using the agglomerative coefficient and tested cluster stability by bootstrapping Jaccard coefficients with 100 replacements. Following the clustering procedure, we used descriptive statistics to compare ventilation settings, clinical characteristics, and medication exposures stratified by cluster.

Results: Seventy-eight patients with type 2 sBPD were included. HCA classified 3 dissimilar clusters (k=3) (Figure 1, Panel A) as determined by an agglomerative coefficient of 0.97 (Figure 1, Panel B). Cluster stability was relatively strong as determined Jaccard coefficient means of 0.81, 0.91, and 0.74 for clusters 1, 2, and 3, respectively. The median PEEP, MAP, rate, Ti, and PIP differed significantly between clusters P< 0.0001 for each comparison by Kruskall-Wallis testing (Figure 2). Patients in cluster 1 were more likely to be small for gestational age at birth and more likely to be exposed to sedative medications on the study day as compared to patients in cluster 2 and cluster 3 (Table 1).Conclusion(s): Unsupervised hierarchical clustering analysis of ventilator setting data identified 3 distinct ventilation strategies in a multi-center cohort of children with type 2 sBPD. We observed significant differences in median PEEP, MAP, rate, Ti, and PIP between clusters. Prospective trials are needed to determine whether these ventilation strategies improve respiratory outcomes for patients with type 2 sBPD and associate with sBPD respiratory phenotypes.
Figure 1. Ward’s hierarchical clustering analysis identified 3 clusters.Panel A - Dendrogram of cohort following Ward’s hierarchical clustering analysis using PEEP, rate, inspiratory time, and mean airway pressure. Three clusters (k=3) were identified following visual inspection of the dendrogram. Cluster 1 is demarcated in red, cluster 2 is demarcated in green and cluster 3 is demarcated in blue. Panel B - Cluster plot following principal component analysis. Clusters were quantitatively dissimilar as determined by an agglomerative coefficient of 0.97.
Figure 2. Ventilator settings differed significantly between clusters.Panel A - Peak inspiratory pressure (PIP), Panel B - Positive end expiratory pressure (PEEP), Panel C – Set respiratory rate, Panel D – Inspiratory time. Panel E, Mean airway pressure (MAP), Panel F – Fraction of inspired oxygen (FiO2). Data were compared using the Kruskall-Wallis test with Benjamini-Hochburg procedure for multiple comparisons. NS – not significant, **P < 0.01, ***P < 0.001, ****P < 0.0001. The dashed line represents the median for each ventilator setting across the cohort.