Introduction: In-hospital cardiac arrest (IHCA) is an important cause of morbidity and mortality in children. IHCA prediction to prevent events remains a challenge in the pediatric intensive care unit (PICU). Machine learning algorithms have been used successfully for clinical predictions tasks. An emerging approach is the use of foundation models, pre-trained on unrelated datasets, for this purpose. Tabular Prior-data Fitted Network (TabPFN) is a recently published foundation model designed for tabular data, such as electronic health record data. Our objective was to evaluate the ability of TabPFN to predict IHCA in PICU patients 6 hours after admission compared to traditional machine learning algorithms. Methods: Retrospective cohort study of PICU patients from 1/6/2016-30/5/2022. Features included demographics, vital signs, assessments, laboratory values and medications from the first 6 hours of admission. We compared TabPFN to multiple traditional machine learning algorithms (K-Nearest Neighbor, Random Forest, LightGBM, XGBoost and Tabular Neural Network) for model development. Results: IHCA occurred in 175/14,050 encounters (1.2%). TabPFN had the best performance (AUROC 0.84, AUPRC 0.16) followed by Random Forest (AUROC 0.83, AUPRC 0.12). Negative predictive value was >98.8% for all algorithms, however positive predictive value was low (2.9-5.1%). Conclusions: We have demonstrated a modest ability to identify children at risk of IHCA early in their ICU admission, this may be improved by developing models that re-evaluate risk over time. Though the TabPFN model was not developed using pediatric ICU data, it outperformed the other algorithms highlighting the potential of this approach.
Brown et al. (Sun,) studied this question.