Introduction: A biomarker-based staging system has recently been described for prognosis in wild-type transthyretin cardiac amyloidosis (ATTRwt). We validated this staging system in consecutive ATTRwt patients and compared its predictive accuracy to machine learning algorithms.
Methods: Clinical data was extracted from medical records of patients diagnosed with ATTRwt over the past 5 years. Diagnosis was based on technetium pyrophosphate scanning, or histopathological analysis of cardiac biopsy specimens. Data on death was collected from medical records as well as social security death index. The following machine learning algorithms were tested – Random Forest learner, AdaBoost, Naïve Bayes, Stochastic Gradient Descent (SGD). The model was built on a random sampling of 80% data, and tested on the remaining 20%, over 100 iterations. For comparison, a recently described staging system was used, which allots one point each for NT-proBNP >3000 pg/mL and troponin- T > 0.05 ng/mL, which results in the stages 1(0 points), 2 & 3 (2 points). The Area under the Curve (AUC) and the classification accuracy (proportion of outcome correctly predicted) were compared.
Results: Among 197 individuals with ATTRwt (mean age 76 ± 6 years, 6 women) 86 were in Stage 1 (44%), 66 in Stage 2 (33%) and 45 in Stage 3 (23%). There were 59 deaths (30%), with a median survival of 5.1 years. The staging system had an AUC of 0.70 (95% CI 0.62-0.77). All machine learning approaches performed better than clinical staging, except for AdaBoost (AUC – 0.71); Table. Naïve Bayes had the highest AUC (0.86), whereas SGD had the highest classification accuracy (0.83).
Conclusion: Biomarker based staging is valid and has moderate accuracy for predicting mortality in patients with ATTRwt cardiac amyloidosis. However, machine learning algorithms can provide superior predictions compared with biomarker staging.