A Reproducible Post-Valve-Replacement EHR Cohort for Comparative AI Studies.

Area of Science:

• Artificial Intelligence in Medicine
• Health Informatics
• Cardiovascular Surgery

Background:

• Patients undergoing valve replacement (VR) face significant postoperative complication risks.
• Reproducible Electronic Health Record (EHR) benchmarks for evaluating sequential AI models in VR patients are currently lacking.
• This study addresses the need for standardized datasets and benchmarks for AI model development in this critical patient population.

Purpose of the Study:

• To develop a reproducible pipeline for extracting and preparing EHR datasets from MIMIC-IV for valve replacement patients.
• To create a predictive benchmark dataset for postoperative risk assessment using ICU readmission as a surrogate outcome.
• To compare the performance of sequential AI models against traditional machine learning baselines for postoperative risk prediction.

Main Methods:

• A cohort of 3890 valve replacement patients was curated from MIMIC-IV, incorporating diagnoses, procedures, lab results, medications, and physiological data.
• A sequential Transformer model was trained on tokenized longitudinal EHR sequences.
• Performance was compared against non-sequential Transformer and XGBoost models trained on aggregated feature statistics, using rigorous statistical testing.

Main Results:

• The sequential Transformer model achieved an AUROC of 0.87 and AUPRC of 0.69 for predicting ICU readmission.
• The sequential model significantly outperformed the non-sequential Transformer model.
• A favorable trend was observed compared to the XGBoost baseline, though not statistically conclusive.

Conclusions:

• Leveraging longitudinal EHR data sequences enhances predictive performance for postoperative risk in valve replacement patients compared to static feature summaries.
• The developed preprocessing pipeline and cohort-construction code are publicly released to facilitate reproducibility and benchmarking for future AI research in this domain.
• This work provides a foundation for developing and comparing advanced time-series models for critical care in post-valve replacement patients.