Transthyretin cardiac amyloidosis (ATTR-CA) is an increasingly recognised disease, especially among older patients with heart failure with preserved ejection fraction and aortic stenosis, with an estimated prevalence of 12-15% among patients with aortic stenosis (1). Early initiation of disease-specific therapy reduces all-cause mortality and cardiovascular-related hospitalisations (2). However, early identification of patients with ATTR-cardiac amyloidosis remains a clinical challenge, with a median time from symptom onset to diagnosis of 2 years (3). Among diagnostic methods for the identification of patients, bone scintigraphy imaging has emerged as a highly accurate method for identifying patients with ATTR-CA without biopsy. The spectrum of referral indications for scintigraphy is broad, hence, cardiac uptake might occur as an incidental finding. Therefore, the diagnosis of cardiac amyloidosis can be lost by nuclear medicine specialists not informed by the treating physician. Methods to automate opportunistic screening for ATTR-CA could potentially address this care gap.
Recently, in The Lancet Digital Health Spielvogel and colleagues (4) present a solution for large-scale, opportunistic screening of bone scintigraphy imaging. The authors developed a multistep approach to identifying patients with abnormal uptake on planar imaging. First, image counts were normalized; then a deep learning algorithm was used to crop images of the thorax. Such images were used as the input for a convolutional neural network (CNN). This AI approach classified images as positive (abnormal planar uptake) or negative (no abnormal uptake). The overall predictive performance of the model was great (area under the curve [AUC] 1·000 [95% CI 1·000–1·000] during internal cross-validation) and remained relatively consistent across cohorts and radiotracers (AUC ranging from 0·925 to 1·000). The authors evaluated sources of false positive and negative results to further improve their solution, thereby improving the positive predictive value from 0.886 to 0.932.
Before Spielvogel and colleagues’ approach addressed this important clinical gap and showed excellent results, a similar work was published by Delbarre and colleagues (5). They trained a CNN to identify abnormal image uptake on planar imaging, with AUC of 0.999 during internal cross-validation. The added value of the study by Spielvogel and colleagues is not only the increase in sample size (16,241 vs 3,048 patients), but specifically the external testing of their solution in multiple centres with a variety of imaging protocols. External testing is a critical step for any model but particularly so for bone scintigraphy imaging, where there are differences in image sets between sites related to radiotracers and imaging protocols. Additionally, it is worth noting that the solidity of the results presented by Spielvogel and colleagues was the fact that they used a core laboratory approach, which involved each study being reviewed in triplicate (with another two readers for the adjudication of discrepancies).
In summary, the authors’ research question addressed an important clinical gap for a disease in which the diagnostic delay is still measured in years, by developing and rigorously evaluating a model for the detection of abnormal radiotracer uptake on bone scintigraphy imaging. Moreover, the external validation and the consistency demonstrated among centres suggest that this model could be implemented clinically to assist in detecting patients with ATTR-CA. Therefore, incorporating the AI system as an on-line tool, which sits directly on the scanning software and has the ability for immediate image analysis during image acquisition, would enable automated real-time detection of cardiac amyloidosis-suggestive uptake and trigger further diagnostic steps.
Our mission: To reduce the burden of cardiovascular disease.