A Deep Learning Model to Guide Personalized Mechanical Circulatory Support Use in Cardiogenic Shock Patients Undergoing PCI.

BACKGROUND

Cardiogenic shock (CS) in patients undergoing percutaneous coronary intervention (PCI) involves rapidly changing clinical, hemodynamic, and metabolic factors that current models cannot effectively integrate.

OBJECTIVES

The purpose of this study was to develop and validate a deep learning (DL) model to guide personalized use of mechanical circulatory support (MCS) devices in CS patients undergoing PCI.

METHODS

The authors analyzed data from 1,408 CS patients treated with intra-aortic balloon pump (IABP) or microaxial flow-pump (mAFP) at multiple hospitals (2004-2019). Clinical, hemodynamic, and metabolic variables from admission to MCS placement were compiled into longitudinal vectors. Using Python 3.7 and TabNet in PyTorch, the OPtiMCS DL model was developed and validated in Google Colab to predict 30-day mortality, bleeding, acute kidney injury (AKI), 1-year mortality, and 1-year stroke. The model also simulated alternate outcomes by switching between IABP and mAFP to support a patient-centered approach.

RESULTS

The rates for 30-day mortality, bleeding, AKI, 1-year mortality, and 1-year stroke were 31%, 43%, 35%, 35%, and 5%, respectively. OPtiMCS achieved AUCs from 83% (AKI) to 98% (mortality) and identified key predictive features (eg, cardiac arrest predicted mortality). OPtiMCS predicted benefit of one MCS device over the other (eg, IABP switched to mAFP was predicted to reduce mortality by 1.8% and AKI by 5.2%).

CONCLUSIONS

Among CS patients undergoing PCI, we developed and validated a DL model to predict outcomes and facilitate a patient-centric approach. If it can be externally validated and implemented in clinical practice, it has the potential to distinguish the risk-benefit balance of MCS devices, reduce adverse outcomes, improve survival, and ultimately advance CS care.