
Contact the VUNO team anytime.
APICS 2024
Objectives:
The primary aim of this study is to enhance the predictive capabilities of early warning systems for
pediatric patients through cross-age transfer learning. By adapting models pre-trained on adult patient
data to pediatric patients, we seek to significantly improve the accuracy of predicting clinical
deterioration, thereby enabling clinicians to better identify at-risk pediatric patients and optimize the
allocation of medical resources to improve patient outcomes.
Methods:
This retrospective study analyzed 50,019 pediatric admissions to Asan Medical Center Children’s
Hospital in South Korea from January 2012 to December 2018. The dataset from 2012 to 2016 was
used for model development, while data from 2017 to 2018 were used for validation. The primary
outcome targeted early prediction of clinical deterioration, including unplanned transfers to the pediatric
intensive care unit (PICU) and cardiopulmonary arrests, using the four classic vital signs as input
features. We analyzed the performance of three variants of the DeepCARSTM deep-learning models:
one trained exclusively on adult data, another on pediatric data alone, and a third model that was first
trained on adult data then fine-tuned with pediatric data. Our evaluation focused on determining the
effectiveness of cross-age transfer learning for improving the early prediction of deterioration in pediatric
patients.
Results:
Predicting deterioration in pediatric patients is a challenging problem due to the limited data available
when compared to adult populations. In this regard, we developed a model with improved predictive
performance for pediatric patients using pre-existing deep learning model trained on adult patients.
Conclusions:
Of the 21,236 patients in the validation cohort, 430 (2.02%) experienced deterioration. The model
trained exclusively on adult data achieved an AUROC of 0.810. Meanwhile, the model trained solely on
pediatric data reached an AUROC of 0.915, and the model pre-trained on adult data and fine-tuned with
pediatric data attained the highest AUROC of 0.951.