PCA-based feature transfer learning for outcome prediction in syndromic and metabolic cardiomyopathy

Background Risk prediction in syndromic and metabolic cardiomyopathies (SMCs) is limited by the rarity of the disease and small cohort sizes. We evaluated a principal component analysis (PCA)-based feature transfer learning approach using a large hypertrophic cardiomyopathy (HCM) cohort to derive a prognostic score for patients with SMC, with comparison to a traditional approach. Methods The study included 97 patients with SMC and 7,322 patients with HCM sharing 47 harmonized clinical and imaging features. PCA was performed exclusively in the HCM cohort, and principal components (PCs) with eigenvalues >1 were transferred to the SMC cohort for model development. Traditional and PCA-based feature transfer models were developed using Cox regression, random survival forests (RSF), and logistic regression. The best-performing algorithm under each approach was subsequently refitted using the entire SMC cohort to derive individualized prognostic scores. The prognostic scores were evaluated using the concordance index (C-index), time-dependent area under the curve (AUC), calibration, net reclassification improvement (NRI), integrated discrimination improvement (IDI), and decision curve analysis. Results Four PCs explaining 56.3% of the variance were retained. Cox regression and RSF were identified as the best-performing algorithms for the traditional and PCA-based feature transfer learning approaches, respectively. Compared with the traditional prognostic score, the PCA-based feature transfer prognostic score showed comparable but slightly lower discrimination (C-index 0.818 vs. 0.856, P = 0.248), a potential advantage in risk reclassification (continuous NRI 0.584, P = 0.018; IDI 0.114, P = 0.042), and greater net benefit on decision curve analysis. The PCA-based feature transfer prognostic score achieved time-dependent AUCs of 0.933 at 1 year and 0.836 at 3 years. Event rates increased across risk quartiles, from 16.0% in the lowest-risk group to 75.0% in the highest-risk group. Conclusions PCA-based feature transfer learning may offer a practical strategy for prognostic assessment in rare cardiomyopathies. Its clinical utility remains to be established through validation in larger, independent cohorts.