Abstract Details

Title Advanced Predictive Modeling of Children with Neurological Injury in the PICU: A Machine Learning Approach

Topic Neuro Trauma, Critical Care, and Sports Neurology

Presentation(s) S48 - Neurocritical Care: Traumatic Brain Injury and Goals-of-care Decision-making (2:24 PM-2:36 PM)

Poster/Presentation
Number 008

Objective We hypothesize that nonlinear analytics will demonstrate greater predictive performance for morbidity and mortality compared to established linear methods among critically ill children with neurological injury.

Background PICU patients with acute neurological injury are thought to be at higher risk for poor outcomes, though no predictive outcome models are validated for this population. The Trichotomous Outcomes in Pediatric Critical Care (TOPICC) study produced multicenter outcomes data for all ICU patients, enrolling 10,078 patients <18y from 2011-2013 and created a logistic regression model predicting mortality and new morbidity from admission physiologic dysfunction.

Design/Methods TOPICC subgroup analysis of patients with acute neurological injury using techniques including Gradient Boosting (GB), Random Forest (RF), and neural networks (NN).

Results 1860 patients had acute neurological injury at PICU admission. A higher percentage developed new morbidity (8.2 vs 4.3%) and died (6.2 vs 2.7%) during admission versus the general PICU population, p<0.001. The RF regressor performed best, with a C-statistic of 0.89 [95% CI 0.88, 0.91] and Average Precision (AP) of 0.56 [0.53,0.59] for mortality, similar for simultaneous mortality and morbidity (0.81 [0.81, 0.82] and 0.55 [0.54, 0.57]). The GB model was similar, C-statistic of 0.88 [0.87, 0.90] (p=0.18) and AP of 0.56 [0.53, 0.58] (p=0.83). The TOPICC regression applied to this subset had a C-statistic of 0.92 [0.90, 0.94] but an inferior AP of 0.50. Feature importance analysis using mean decrease in impurity on the RF model showed that hyperthermia, hypertension, thrombocytopenia, tachycardia, hypercarbia, and hyperglycemia were the most contributory features.

Conclusions

PICU patients with acute neurological injury are at higher risk for mortality and morbidity versus general PICU patients. Many recognized secondary, potentially modifiable, CNS insults were identified by our machine learning approaches. Future work will extrapolate these methods to a more robust set of institutional electronic record data with the goal of bolstering predictive performance.

Authors/Disclosures
Neil Munjal, MD (University of Wisconsin - Madison) PRESENTER	No disclosure on file
	No disclosure on file
Dennis W. Simon, MD (University of Pittsburgh)	Dr. Simon has nothing to disclose.
Patrick M. Kochanek, MD, MCCM (University of Pittsburgh)	Patrick M. Kochanek, MD, MCCM has received personal compensation in the range of $10,000-$49,999 for serving as an Expert Witness for Johns Hopkins Health System. The institution of Patrick M. Kochanek, MD, MCCM has received research support from Chuck Noll Foundation. The institution of Patrick M. Kochanek, MD, MCCM has received research support from NIH. Patrick M. Kochanek, MD, MCCM has received publishing royalties from a publication relating to health care.
	No disclosure on file

��ɫ��

��ɫ��