(1529) Cardiac Arrest Prediction in the Pediatric CICU: A Fine-Tuned Language Model Approach

Introduction: Electronic health record (EHR) data, spanning various data modalities (i.e., tabular, time-series, and text), store rich clinical information for cardiac arrest (CA) prediction. Conventional artificial intelligence models are developed to handle one specific data modality, which poses great challenges for improving patient health outcomes. Language models (LMs) pretrained on large-scale biomedical corpora have proven effective in analyzing clinical texts. This study aims to develop a CA prediction model using an LM that can effectively integrate multimodal clinical data.

Methods: A retrospective analysis of pediatric patients admitted from 2010 to 2023 to a single quaternary, academic pediatric Cardiac ICU (CICU) was performed. Patients with CA were identified using the code blue event of chest compressions performed. Patient demographics, vital sign and assessment flowsheets, laboratory values, and admitting diagnosis within the first 24 hours were extracted from the EHR to predict the probability of developing CA during the rest of the CICU stay. We followed a case-control design to create two datasets: 1) a balanced incidence of CA (35%) for prototype validation and 2) an unbalanced incidence of CA (3.9%) for pseudo real-world validation. All datasets were split into 80%-20% training-testing sets. We fine-tuned an LM on the training set by presenting the multimodal clinical data in the textual format (e.g. Age: 81 days. Resp: 20-41. Admitting Diagnosis: Hypoxemia).

Results: There were 15,145 admissions (177 with CA events), including a total of 204 clinical features, that met the inclusion criteria. The fine-tuned LM model for CA prediction achieved an area under the receiver operating (AUROC) of 0.86 and 0.94 in the balanced and unbalanced datasets, respectively. Furthermore, our fine-tuned LM achieved a positive predictive value (PPV) of 75% with a sensitivity of 82% in the balanced dataset, and a PPV of 54% with a sensitivity of 73% in the unbalanced dataset.

Conclusions: This study demonstrates the capability of our fine-tuned LM to integrate multiple types of clinical data to predict CA in pediatric patients. Future work will involve improving our proposed model by expanding our analysis to various data sources across different sites.