Deep learning methods for clinical workflow phase-based prediction of procedure duration: a benchmark study.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Cardiac catheterization laboratory (cath lab) procedures require precise scheduling.
• Predicting procedure end times is crucial for optimizing cath lab workflow and patient throughput.

Purpose of the Study:

• To evaluate deep learning models for predicting cardiac catheterization procedure end times using video-derived clinical phases.
• To identify the most accurate deep learning architectures for this predictive task.

Main Methods:

• Utilized clinical phases from video analysis as input for deep learning models.
• Evaluated various models including InceptionTime, LSTM-FCN, LSTM with attention, standard LSTM, and Transformer.
• Assessed prediction accuracy using Mean Absolute Error (MAE) and Symmetric Mean Absolute Percentage Error (SMAPE).

Main Results:

• InceptionTime and LSTM-FCN demonstrated the highest prediction accuracy.
• InceptionTime achieved MAE below 5 minutes and SMAPE under 6% at 60-s intervals.
• CNN-based models, especially InceptionTime, excelled in feature extraction for time-series prediction.
• Transformer models offered the fastest inference times for real-time applications.
• An ensemble model showed low error rates but required longer training.

Conclusions:

• Deep learning models, particularly CNN architectures like InceptionTime, are effective for accurately predicting cath lab procedure end times.
• These models can form the basis of an automated tool to predict the optimal time for calling the next patient, with potential average errors around 30 seconds.
• Integration into clinical scheduling systems can enhance cath lab efficiency.