Serial 12-Lead Electrocardiogram-Based Deep-Learning Model for Hospital Admission Prediction in Emergency Department...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Health Informatics

Background:

• Emergency department (ED) crowding is a significant issue, often exacerbated by challenges in predicting patient disposition.
• Accurate and timely prediction of hospital admission for patients with cardiac conditions is crucial but difficult due to evolving patient risk.
• Existing predictive scores often rely on single-time-point data and may not perform optimally as patient status changes.

Purpose of the Study:

• To develop and validate a real-time deep-learning model for early prediction of hospital admission in ED patients with cardiac issues.
• To integrate serial 12-lead electrocardiogram (ECG) waveforms with sequential vital signs and clinical data.
• To improve the accuracy and timeliness of disposition predictions compared to existing methods.

Main Methods:

• Retrospective cohort study using MIMIC-IV databases (ED, ECG modules).
• Included adults with chest pain, dyspnea, syncope, or presyncope and at least one ECG.
• Developed and compared a multimodal deep-learning model against baseline tabular (random forest) and ECG-only models using data available up to the last ECG.

Main Results:

• The multimodal deep-learning model achieved an AUROC of 0.911 (all stays) and 0.924 (subset with ≥2 ECGs), outperforming baseline models.
• The model predicted disposition significantly earlier than ED departure (median 4.6 hours prior).
• Baseline models showed lower performance: ECG-only (AUROC 0.852-0.859) and tabular RF (AUROC 0.886-0.911).

Conclusions:

• Serial ECGs combined with evolving vital signs and clinical data enable accurate, early prediction of ED disposition for cardiac patients.
• The developed multimodal deep-learning framework is open-source and reproducible.
• This approach has the potential to streamline ED operations, prioritize high-risk patients, and detect critical conditions early.