Multimodal multi-instance learning for cardiopulmonary exercise testing performance prediction.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Heart failure (HF) affects millions, with rising prevalence and limited predictive tools.
• Cardiopulmonary exercise testing (CPET) is vital for HF prognosis but faces practical limitations.
• Accurate prediction of functional capacity is crucial for guiding HF treatment.

Purpose of the Study:

• To develop a multimodal AI framework for predicting peak oxygen consumption (peak VO₂) from transthoracic echocardiography (TTE) and electronic health records (EHR).
• To enhance the prediction of survival outcomes and identify high-risk HF patients.
• To overcome the limitations of traditional CPET in clinical practice.

Main Methods:

• A multimodal multi-instance learning framework was designed to integrate TTE and EHR data.
• The model specifically addressed cross-modal interactions and the multi-instance nature of TTE studies.
• Performance was evaluated based on R² for peak VO₂ prediction and AUROC for high-risk patient identification.

Main Results:

• The AI framework achieved an R² of 0.603 for peak VO₂ prediction, outperforming previous methods (R² = 0.529).
• It demonstrated strong performance in identifying high-risk patients with an AUROC of 0.849.
• External validation showed superior results with R² = 0.541 and AUROC = 0.870 compared to prior work.

Conclusions:

• The developed AI framework offers a more accurate and accessible method for assessing HF patient functional capacity and prognosis.
• This approach can improve the identification of patients eligible for advanced heart failure therapies.
• The study highlights the potential of integrating TTE and EHR data using advanced AI for better HF management.