Artificial intelligence-enabled cardiac CT for population-wide type 0-bicuspid aortic valve screening: a model development and validation study

Background

We developed Cardioverse-BAV, an AI platform for population-wide Type 0-bicuspid aortic valve (BAV) screening. It leverages ECG-gated cardiac CT (CTA) while circumventing its need for specialized radiologists.

Methods

Derived from the Three-Province transcatheter aortic valve replacement (TAVR) Registry, Cardioverse-BAV platform leverages a deep learning model trained on 1156 core lab-confirmed ECG-gated CTAs from 1156 TAVR patients. The development set was randomly assigned to training and internal test sets at a 7:3 ratio. Its diagnostic performance was validated in an external surgical aortic valve replacement (SAVR) cohort from a Hospital University between Jan 2018 and Sept 2025 (n=736 with both preoperative ECG-gated CTA and surgical confirmation) and a large-scale cohort receiving screening coronary CTA and transthoracic echocardiography (TTE) (n=16,770) from two tertiary hospitals between Jan 2020 and Aug 2025.

Results

In the internal test set, Cardioverse-BAV platform accurately predicted BAV (AI model: 32.2% vs. core lab: 30.0%), achieving weighted aggregated scores as follows: accuracy 98.9%, recall 96.7%, specificity 100%, precision 100%, and F1-score 98.3%.

In external test sets, the AI model identified BAV in 5.3% of cases, outperforming TTE (2.4%) and comparable to surgical confirmation (5.0%). It achieved 99.2% accuracy, with recall of 94.6%, specificity of 99.4%, precision of 89.7%, and F1-score of 92.1%. Compared to TTE, the AI model showed a significant improvement in recall (+81.1%, P < 0.001), achieving 94.6% recall vs. 13.5% for TTE. While its specificity (99.4%) was slightly higher than TTE (98.1%), the difference did not reach statistical significance (P = 0.999).

In a cohort of 16,770 patients, ECG-gated CTA identified BAV in 1.3% of cases versus 0.4% by TTE (P < 0.001), with weak agreement between methods (kappa = 0.193).

Conclusion

The Cardioverse-BAV platform is a tool for BAV diagnosis with the potential to enhance early detection, as evidenced by its validation in multiple distinct cohorts.