Interpretable Artificial Intelligence for Detecting Acute Heart Failure on Acute Chest CT Scans

AI can help address radiologist shortages: The model achieved an area under the ROC of 0.87 in detecting acute heart failure on chest CT scans, demonstrating that AI could help flag critical findings for emergency physicians when radiologist reports are delayed due to staffing shortages.

Explainability matters for clinical adoption: The researchers used Shapley Additive explanations (SHAP) to show how each of twelve key measurements contributed to the AI's predictions, making the model's decision-making transparent rather than a "black box"—an important requirement for doctors to trust and safely use AI in patient care.

AI errors may sometimes reflect limitations in training data, not model failure: When experts reviewed the model's mistakes, they found that 38% of false positives and 39% of false negatives were actually correct predictions, suggesting the errors came from mislabeled training data in radiology reports rather than true model failures.

Automated measurements enable AI analysis: The model relied on automatically segmented cardiac and pulmonary structures from CT scans rather than requiring manual measurements, making the approach scalable and reproducible across different hospitals and radiologists.

AI performance approaches expert-level accuracy: By achieving diagnostic performance comparable to thoracic radiologists, this model demonstrates AI's potential to serve as a complementary tool that could enhance diagnostic precision and speed in emergency settings.