Transparent chest radiograph foundation model enables explainable human disease profiling

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Chest radiography (CXR) is a widely accessible imaging modality.
• Its potential for detecting subtle systemic disease manifestations is underexplored.
• Multimodal foundation models offer new avenues for analyzing complex medical data.

Purpose of the Study:

• To develop and evaluate a contrastively pretrained multimodal CXR foundation model.
• To assess the model's capacity for predicting diverse human diseases using electronic health records.
• To enhance model interpretability through shared embedding spaces for diseases and CXR features.

Main Methods:

• Developed a multimodal CXR foundation model using large-scale image-report pairs.
• Trained linear probes on frozen image embeddings using 1074 phecodes from electronic health records.
• Validated model performance across three independent cohorts (totaling over 230,000 patients).
• Utilized co-embedding analysis of diseases and 57 radiologist-curated CXR features.

Main Results:

• The model significantly predicted 554 prevalent and 457 incident phenotypes.
• High discrimination was observed for 60 prevalent and 42 incident phenotypes across all datasets.
• Co-embedding identified 28 phenotype clusters linked to recognizable imaging patterns (e.g., cardiomegaly, atherosclerosis).
• Model embeddings captured imaging signatures for near-term cardiovascular events and critical illness.
• Reconstruction of predictions showed strong explanatory performance (R² ≥ 0.85).

Conclusions:

• CXR foundation model embeddings encode rich, clinically relevant information beyond conventional interpretation.
• The model provides a scalable and interpretable framework for comprehensive disease profiling.
• Findings support the potential of AI in extracting deeper insights from routine chest X-rays.
• Further prospective validation is warranted to confirm clinical utility.