A Vital Signs-Driven Approach for Clustering the Responses to Diuretics Treatment in Premature Newborns

Chronic lung disease (CLD) is a growing problem worldwide. Pulmonary edema, an excess of fluids in the alveolar or proto-alveolar epithelium, is a frequently reported pathology impacting premature newborns affected by chronic lung disease (CLD), and also respiratory distress syndrome (RDS). Both conditions are generally treated with diuretics. Despite the widespread administration, their associated short and long-term effects are still unclear. We aim to characterize the response to a 3-day diuretic trial in a cohort of premature newborns (N = 198, gestational age 26.44 ± 2.04 weeks) leveraging the richness of the continuously recorded vital signs as well as the information abstracted from the electronic health records (EHR). In particular, we aim to examine the concordance between the doctors’ recommendations on the diuretic treatment (interruption versus continuation) versus the novel unsupervised learning analysis, which exploits vital signs, such as heart rate (HR), respiratory rate (Resp), and oxygen saturation (SpO2), and electronic health records (EHR)-derived features. We utilized (a) the derivative dynamic time warping (DDTW) and (b) the spectral clustering. Findings obtained by analyzing the standard deviation (SD) series of the vital signs suggest to reconsider the traditional classification of respondent or not respondent. Finally, retrospective analysis on remaining length of stay (LOS) in vital signs trend-based clustering confirms our primary hypothesis that the dynamics of physiological signals offer additional insights on treatment response (compared to the sole utilization of EHR) and may be instrumental in developing more reliable monitoring systems.