Physiological control for left ventricular assist devices based on deep reinforcement learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Physiology

Background:

• Heart failure (HF) poses a significant health burden, necessitating advanced Left Ventricular Assist Device (LVAD) technology.
• Current LVAD control strategies have limitations, and reinforcement learning (RL) applications are underexplored.
• This research focuses on improving LVAD control for better patient outcomes in severe HF.

Purpose of the Study:

• To introduce a novel preload-based deep reinforcement learning (DRL) controller for LVADs.
• To enhance LVAD performance by optimizing patient hemodynamics.
• To address limitations in current LVAD control strategies using advanced AI.

Main Methods:

• Developed a DRL controller using the proximal policy optimization algorithm.
• Utilized a high-fidelity cardiorespiratory simulator with varied physiological parameters to model patient variability.
• Trained the DRL controller to prevent ventricular suction and ensure aortic valve opening using critical LV pressure signals.

Main Results:

• The DRL controller demonstrated superior end-diastolic volume (EDV) stability (5 mL SD) compared to constant speed LVAD (9 mL SD).
• Achieved higher aortic flow rates (average 1.1 L/min) with the DRL controller versus constant speed LVAD (0.9 L/min).
• The controller effectively managed hemodynamic parameters in a simulated severe HF population.

Conclusions:

• A DRL-based controller was successfully implemented and validated in a sophisticated cardiorespiratory simulator.
• The DRL controller significantly improved aortic valve flow and EDV stability over a standard constant speed LVAD.
• This approach shows promise for advancing LVAD technology and managing heart failure.