ShouqiangZhu/IOH_Transformer: v1.0 – IOH_Transformer: Transformer-Based Model for Real-Time Prediction of Intraoperative Hypotension

IOH_Transformer v1.0

This repository contains the implementation of a Transformer-based deep learning model for real-time prediction of intraoperative hypotension (IOH) using dynamic time-series vital sign data.

Study Overview

Transient intraoperative hypotension (IOH) is associated with adverse postoperative outcomes, yet many existing prediction models rely on high-resolution waveform data that are not routinely available in clinical practice. This project presents a Transformer-based deep learning framework designed to predict IOH in real time using routinely collected vital sign time-series data.

Dataset

The model was developed using a retrospective dataset of 319,699 surgical cases from a tertiary hospital in China (2013–2023) and externally validated using an independent dataset from South Korea.

Model Performance

The Transformer model demonstrated strong predictive performance:

• 5-minute prediction horizon: AUC = 0.904
• 10-minute prediction horizon: AUC = 0.892
• 15-minute prediction horizon: AUC = 0.882
• Recall ≥ 88.3%

Compared with XGBoost:

• Transformer showed higher recall and better probability calibration
• XGBoost achieved higher accuracy and specificity

External validation confirmed the generalizability of both models.

Clinical Relevance

A nested cohort analysis showed that IOH burden (cumulative MAP ≤65/60/55 mm Hg in mm Hg·min) was significantly associated with postoperative acute kidney injury (AKI) and acute kidney disease (AKD).

Key Features

• Transformer-based deep learning architecture
• Real-time IOH risk prediction
• Time-series vital sign modeling
• External validation
• Clinical outcome association analysis

Citation

If you use this code in your research, please cite:

Zhu S., et al.
Transformer-Based Deep Learning Model for Real-Time Prediction of Intraoperative Hypotension Using Dynamic Time-Series Vital Signs: a retrospective study

License

Please refer to the repository license for usage terms.