Diagnostic accuracy of a machine learning approach applied to delayed [ 18F]-Florbetaben positron emission tomography in patients with suspected light-chain cardiac amyloidosis

Purpose The diagnosis of AL-CA is often difficult and requires invasive assessment by tissue biopsy. The purpose of the study was to evaluate the diagnostic accuracy of a machine learning approach applied to delayed [ ¹⁸ F]-florbetaben positron emission tomography (PET) uptake in identifying patients with light-chain cardiac amyloidosis (AL-CA). Methods 32 patients (age 67 ± 10 years, 9 women) with biopsy-proven diagnosis of AL-CA and 45 control subjects, referred with the initial clinical suspicion (age 74 ± 11 years, 7 women) and later diagnosed with non-AL-CA pathology, underwent a cardiac PET/computed tomography scan. Cardiac [ ¹⁸ F]-Florbetaben PET uptake was assessed using static acquisition 110 min after radiotracer injection. Results Semiquantitative radiotracer uptake showed higher SUV values in patients with AL-CA than in control subjects (p < 0.001). Machine Learning, specifically unsupervised Fuzzy C-means algorithm, proved to be an optimal methodology for classification, with sensitivity of 0.87 and specificity of 0.90 for SUV _mean , and sensitivity of 0.87 and specificity of 0.90 for SUV _max . These values are similar to the ones obtained by statistical analysis (sensitivity 0.87, specificity 0.90 for SUV _mean and sensitivity 0.96, specificity 0.80 for SUV _max ), but the cut-off determined by fuzzy C-means analysis better separates the two groups of subjects. Conclusion Machine learning analysis of delayed [ ¹⁸ F]-Florbetaben cardiac uptake may discriminate AL-CA from mimicking conditions and may represent a noninvasive tool for the diagnosis of AL-CA promising to avoid tissue biopsy for a certain diagnosis.