Mortality Risk Factors in Pediatric Onco-Critical Care Patients and Machine Learning Derived Early Onco-Critical Care Phenotypes in a Retrospective Cohort

OBJECTIVES: To use supervised and unsupervised statistical methodology to determine risk factors associated with mortality in critically ill pediatric oncology patients to identify patient phenotypes of interest for future prospective study. DESIGN: This retrospective cohort study included nonsurgical pediatric critical care admissions from January 2017 to December 2018. We determined the prevalence of multiple organ failure (MOF), ICU mortality, and associated factors. Consensus k-means clustering analysis was performed using 35 bedside admission variables for early, onco-critical care phenotype development. SETTING: Single critical care unit in a subspeciality pediatric hospital. INTERVENTION: None. PATIENTS: There were 364 critical care admissions in 324 patients with underlying malignancy, hematopoietic cell transplant, or immunodeficiency reviewed. MEASUREMENTS: Prevalence of multiple organ failure, ICU mortality, determination of early onco-critical care phenotypes. MAIN RESULTS: ICU mortality was 5.2% and was increased in those with MOF (18.4% MOF, 1.7% single organ failure [SOF], 0.6% no organ failure; p ≤ 0.0001). Prevalence of MOF was 23.9%. Significantly increased ICU mortality risk was associated with day 1 MOF (hazards ratio [HR] 2.27; 95% CI, 1.10–6.82; p = 0.03), MOF during ICU admission (HR 4.16; 95% CI, 1.09–15.86; p = 0.037), and with invasive mechanical ventilation requirement (IMV; HR 5.12; 95% CI, 1.31–19.94; p = 0.018). Four phenotypes were derived (PedOnc1–4). PedOnc1 and 2 represented patient groups with low mortality and SOF. PedOnc3 was enriched in patients with sepsis and MOF with mortality associated with liver and renal dysfunction. PedOnc4 had the highest frequency of ICU mortality and MOF characterized by acute respiratory failure requiring invasive mechanical ventilation at admission with neurologic dysfunction and/or severe sepsis. Notably, most of the mortality in PedOnc4 was early (i.e., within 72 hr of ICU admission). CONCLUSIONS: Mortality was lower than previously reported in critically ill pediatric oncology patients and was associated with MOF and IMV. These findings were further validated and expanded by the four derived nonsynonymous computable phenotypes. Of particular interest for future prospective validation and correlative biological study was the PedOnc4 phenotype, which was composed of patients with hypoxic respiratory failure requiring IMV with sepsis and/or neurologic dysfunction at ICU admission.