Screening performance of a deep learning model versus an automated wall thickness score in biopsy-proven cardiac amyloidosis
Y Nomura, Y Hirata, S Nishio, Y Saijo, R Nonaka, N Yamaguchi, R Zheng, Y Okushi, T Takahashi, K Kusunose, H Yamada, M SataAbstract
Background
Cardiac amyloidosis (CA) is a progressive cardiomyopathy associated with poor prognosis. The recent emergence of disease-modifying therapies has underscored the importance of early diagnosis. Because definitive diagnosis requires multiple tests, an accessible screening approach is essential. Echocardiography plays a pivotal role in screening. An artificial intelligence (AI)-based fully automated echocardiographic analysis system has been developed to detect CA from an apical four-chamber (A4C) cine clip using a deep learning model.
Purpose
To compare the screening performance of a deep learning model (AI-CNN) with a model that determines CA by calculating an increased wall thickness score from AI-derived automated measurements (AI-IWTS), using endomyocardial biopsy as the reference standard.
Methods
We retrospectively screened 171 consecutive patients who underwent endomyocardial biopsy at our institution between 2016 and 2025. Patients with AI-measured left ventricular (LV) wall thickness ≥ 12 mm were included in the analysis because this threshold is required for both AI-CNN and AI-IWTS. CA was defined as amyloid protein positivity on pathological examination. Diagnostic performance was evaluated by sensitivity, specificity, predictive values, and c-index with 95% confidence intervals. Paired classifications were compared using McNemar’s test.
Results
Among the 92 patients with AI-measured LV wall thickness ≥ 12 mm, 58 had biopsy-confirmed CA, including 48 with transthyretin amyloidosis (ATTR), 9 with light-chain amyloidosis (AL), and 1 with apolipoprotein A-I amyloidosis (ApoA-I). AI-CNN demonstrated a c-index of 0.82 (95% CI 0.72–0.89), with a sensitivity of 98% and a specificity of 65%. AI-IWTS demonstrated a c-index of 0.59 (95% CI 0.48–0.69), with a sensitivity of 26% and a specificity of 91%. Positive predictive value (PPV) was 83% for both methods, while negative predictive value (NPV) was higher for AI-CNN than for AI-IWTS (96% vs. 43%). McNemar’s test showed a significant difference in classifications between the two methods (P<0.001).
Conclusions
In a biopsy-referenced cohort selected by AI-measured LV wall thickness ≥ 12 mm, AI-CNN showed higher sensitivity and discrimination than AI-IWTS. These findings suggest that AI-CNN may serve as a practical echocardiographic screening tool to support early detection of CA in routine clinical practice.Diagnostic Comparison of Two AI Models