Diagnostic Performance Evaluation of Clinical and Artificial Intelligence Risk Models in Patients Referred for Cardiac Amyloidosis Testing.

BACKGROUND

To improve screening for cardiac amyloidosis (CA), several models using artificial intelligence (AI) and conventional statistics have been developed. However, few data are available to compare the relative utility of these tools. In this study, models were compared to determine their potential roles in optimizing diagnostic algorithms.

METHODS

In this retrospective cohort study at a tertiary medical center, patients referred for cardiac scintigraphy for the detection of transthyretin amyloid CA (ATTR-CA) who underwent electrocardiography (ECG) and transthoracic echocardiography within 6 months and had clinical characteristics for risk score calculation were included. The performance of previously developed and validated clinical and AI risk models for ATTR-CA, including the transthyretin amyloid cardiomyopathy (ATTR-CM) clinical score and AI models applied to ECG (AI-ECG) and echocardiography (AI-Echo), was compared in a population referred for cardiac scintigraphy. Previously defined thresholds were used for each model. As the ATTR-CM score was validated following the exclusion of monoclonal immunoglobulin light chain (AL) amyloidosis, 28 patients with AL amyloidosis were excluded. AL amyloidosis and ATTR-CA were defined per guideline criteria.

RESULTS

Among 598 patients (median age, 76 years; interquartile range, 67-82 years; 72.6% men), 181 (30%) had ATTR-CA. AI-Echo identified ATTR-CA with 86% sensitivity and 85% specificity, compared with 80% and 64% for AI-ECG and 86% and 69% for the ATTR-CM score. The area under the receiver operating characteristic curve was 0.93 (95% CI, 0.91-0.95) for AI-Echo, 0.79 (95% CI, 0.76-0.83) for AI-ECG, and 0.87 (95% CI, 0.84-0.90) for ATTR-CM score (P < .001). In this cohort, the use of AI-Echo could have avoided more unnecessary scintigraphy than AI-ECG or ATTR-CM score (45 vs 24 vs 37 per 100, respectively) at a threshold probability of 0.25 (1 case of ATTR-CA per 4 referrals for scintigraphy).

CONCLUSIONS

Within a population at high risk for CA, the AI-Echo model demonstrated superior diagnostic discrimination and clinical utility for the identification of ATTR-CA compared with the AI-ECG model and the ATTR-CM clinical score.