Objective: Develop an inpatient predictive model of parental post-traumatic stress (PTS) following their child's care in the Pediatric Intensive Care Unit (PICU). Design: Prospective observational cohort. Setting: Two tertiary care children's hospitals with mixed medical/surgical/cardiac PICUs. Subjects: Parents of patients admitted to the PICU. Interventions: None. Measurements and Main Results: Preadmission and admission data from 169 parents of 129 children who completed follow up screening for parental post-traumatic stress symptoms at 3-9 months post PICU discharge were utilized to develop a predictive model estimating the risk of parental PTS 3-9 months after hospital discharge. The parent cohort was predominantly female (63%), partnered (75%), and working (70%). Child median age was 3 years (IQR 0.36-9.04), and more than half had chronic illnesses (56%) or previous ICU admissions (64%). Thirty-five percent (60/169) of parents met criteria for PTS (>9 on the Post-traumatic Stress Disorder Symptom Scale-Interview). The machine learning model (XGBoost) predicted subjects with parental PTS with 76.7% accuracy, had a sensitivity of 0.83 (95% CI 0.586, 0.964), a specificity of 0.72 (95% CI 0.506, 0.879), a precision of 0.682 (95% CI 0.451, 0.861) and number needed to evaluate of 1.47 (95% CI 1.16, 1.98). The area under the receiver operating curve was 0.78 (95% CI 0.64, 0.92). The most important predictive pre-admission and admission variables were determined using the Local Interpretable Model-Agnostic Explanation, which identified seven variables used 100% of the time. Composite variables of parental history of mental illness and traumatic experiences were most important. Conclusion: A machine learning model using parent risk factors predicted subsequent PTS at 3-9 months following their child's PICU discharge with an accuracy of 76.7% and number needed to evaluate of 1.47. This performance is sufficient to identify parents who are at risk during hospitalization, making inpatient and acute post admission mitigation initiatives possible.
Read full abstract