Integration of machine learning with patient-specific cardiovascular computational fluid dynamics models for predicting hemodynamic instability and identifying underlying cause

Abstract Background Although experience plays a crucial role in decision-making for critically ill patients following cardiac surgery in the paediatric intensive care unit (PICU), standardised patient care is equally essential. Patient-specific cardiovascular computational fluid dynamics modelling (CFD) can replicate a patient’s detailed haemodynamics as a digital twin. When combined with machine learning, the patient-specific cardiovascular CFD model can identify which haemodynamic parameter is responsible for the disturbance. Purpose By combining machine learning with patient-specific cardiovascular CFD models of children who have undergone cardiac surgery, we aim to predict lactate levels as a marker of haemodynamic stability or instability and identify the most influential haemodynamic parameters affecting lactate level prediction. Methods Twenty-two postoperative children (13 females and 9 males, median age of 7 months, and median weight of 3.2 kg) who underwent definitive biventricular repair were included. In total, 290 blood draws and corresponding arterial waveforms were collected. The radial artery waveforms were reconstructed by minimising the differences between the actual and re-created waveforms derived from patient-specific cardiovascular CFD models. For the machine learning process, three non-linear regression algorithms (Random Forest, Gradient-Boosting Decision Tree, and Support Vector Machines) were compared to determine the best predictive performance for lactate levels, using mean absolute error (MAE) as the evaluation metric, while feature importance was evaluated by Shapley Additive explanation (SHAP) score. The input data included patient characteristics, blood gas data (lactate level), arterial waveform features (heart rate, systolic/diastolic pressure, area under the curve, and peak angle sharpness), and estimated cardiovascular parameters from the patient-specific CFD models (systemic/pulmonary resistance and compliance, and ventricular elastance). The output variable was the lactate level at an arbitrary timing of blood draw. Results The Random Forest demonstrated the lowest mean absolute error (MAE) of 2.93 mg/dL when the recent lactate level was included as an input alongside the patient-specific CFD model parameters. Feature importance analysis revealed that the recent lactate level had the highest SHAP score of 10.05 (83%), followed by heart rate at 0.34 (2.8%), duration since previous blood draw at 0.21 (1.7%), and systemic arterial compliance at 0.21 (1.0%). Conclusion By combining machine learning with patient-specific CFD models, including measured data, lactate levels can be predicted with high accuracy. With this approach, patient-specific CFD models can function as digital twins and help explain the underlying mechanisms of haemodynamic stability and instability in postoperative care in the PICU.Prediction of lactate level Mean absolute error of three algorithms