An "All-Data-on-Hand" Deep Learning Model to Predict Hospitalization for Diabetic Ketoacidosis in Youth With Type 1...

Area of Science:

• Artificial Intelligence in Medicine
• Pediatric Endocrinology
• Computational Health

Background:

• Diabetic ketoacidosis (DKA) poses significant risks for youth with type 1 diabetes (T1D).
• Existing risk factors for DKA lack clinic-ready predictive models.
• Deep learning, specifically Long Short-Term Memory (LSTM) networks, offers a potential solution for risk prediction.

Purpose of the Study:

• To develop and validate a Long Short-Term Memory (LSTM) model for predicting the 180-day risk of DKA-related hospitalization in pediatric patients with T1D.
• To assess the model's accuracy and generalizability across different patient cohorts.

Main Methods:

• Utilized 17 quarters of clinical data (2016-2020) from 1745 youth (ages 8-18) with T1D.
• Input data included demographics, clinical observations, medications, visit history, prior DKA episodes, and natural language processing-derived features from clinical notes.
• Trained and validated the LSTM model using partial and full out-of-sample cohorts.

Main Results:

• The LSTM model demonstrated validity in predicting 180-day DKA hospitalization risk.
• Precision in identifying high-risk individuals improved significantly with list ranking (e.g., 100% for top 10 in OOS-P cohort).
• DKA admissions occurred at a 5% rate per 180 days in validation cohorts.

Conclusions:

• The developed LSTM model is a valid tool for predicting DKA hospitalization risk in youth with T1D.
• The model enables clinics to identify at-risk youth for proactive intervention development.
• Further validation in diverse populations is recommended to address potential health inequities.