Prediction of lactate concentrations after cardiac surgery using machine learning and deep learning approaches

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning

Background:

• Conventional surgical patient prediction models often overlook intraoperative time-series data.
• Deep learning excels at integrating complex, time-varying data, crucial for understanding physiological dynamics.
• Blood lactate concentration is a key indicator of systemic perfusion adequacy during cardiac surgery.

Purpose of the Study:

• To predict maximum postoperative blood lactate concentrations following cardiac surgery.
• To evaluate the efficacy of machine learning and deep learning models using intraoperative data.
• To determine if minute-level intraoperative data improves predictive performance over static models.

Main Methods:

• Adult patients undergoing cardiac surgery with cardiopulmonary bypass were included.
• Three model classes were evaluated: static (preoperative variables), augmented static (with intraoperative statistics), and dynamic (integrating time-series data).
• The primary outcome was maximum lactate concentration within 24 hours postoperatively, with mean absolute error as the performance metric.

Main Results:

• Models incorporating intraoperative data significantly reduced prediction error compared to baseline models.
• Dynamic models using recurrent neural networks and transformers achieved the lowest prediction error (median 1.96 mmol/L).
• Intraoperative lactate concentration, anemia, and weight were identified as key predictors.

Conclusions:

• Postoperative lactate levels can be moderately predicted using baseline and intraoperative data.
• The integration of intraoperative data, especially time-series information, enhances predictive accuracy for clinical outcomes.
• These findings underscore the value of real-time data for guiding perioperative management.