Optimizing point-of-care ultrasound video acquisition for probabilistic multi-task heart failure detection

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cardiology

Background:

• Point-of-care ultrasound (POCUS) requires efficient data acquisition under time constraints.
• Assessing heart failure (HF) involves key biomarkers like aortic stenosis (AS) severity and left ventricular ejection fraction (LVEF).

Purpose of the Study:

• To develop a personalized, budget-constrained data acquisition strategy for POCUS using reinforcement learning (RL).
• To enable an RL agent to select optimal views or terminate acquisition for HF assessment.
• To jointly predict AS severity and LVEF with calibrated uncertainty upon study termination.

Main Methods:

• Modeled POCUS as a sequential, cost-constrained acquisition problem with an RL agent selecting from five standard views or terminating.
• Employed a multi-view transformer for joint AS classification and LVEF regression upon termination.
• Trained the RL agent using online RL in a partial-observation simulator.

Main Results:

• The method matched full-study performance using 32% fewer videos on a dataset of 12,180 studies.
• Achieved 77.2% mean balanced accuracy across AS severity classification and LVEF estimation.
• Demonstrated robust multi-task performance under acquisition budgets.

Conclusions:

• Patient-tailored view selection preserves diagnostic quality while reducing acquired views in retrospective POCUS.
• The framework supports adaptive acquisition policies and is extensible to other cardiac endpoints.
• Prospective evaluation in live bedside POCUS is warranted.