A Machine Learning Model for the Systematic Identification of Wild-Type Transthyretin Cardiomyopathy

Abstract

Introduction Wild-type transthyretin (TTR) amyloid cardiomyopathy (ATTRwt)—a life-threatening but now treatable disease—is often misdiagnosed as more common etiologies of heart failure (HF). As a strategy to identify undiagnosed patients, we sought to develop and validate a machine learning (ML) model to identify ATTRwt patients using medical claims data. Methods We developed an ML model for the identification of ATTRwt using ICD codes from US medical claims data sourced from IQVIA (n =∼300 million patients, up to 10 years of medical history). An ATTRwt cohort was derived using an ICD code for ATTRwt and compared to a random sample of HF patients matched by age, gender, and medical histories [cohort 1a (ATTRwt): N=373, cohort 1b (HF): N=373]. The classification model employed the Random Forest algorithm to learn from data from 80% of the patients (training set) and tested the learnings on the remaining 20% of patients (test set). Model validation was performed by predicting ATTRwt in 3 additional cohorts: ATTRwt and non-ATTRwt HF patients in electronic medical record (EMR) data [cohort 2a (ATTRwt): N=100, cohort 2b (HF): N=100], cardiac amyloidosis (CA) and HF patients in medical claims data [cohort 3a (CA): N=4,485, cohort 3b (HF): N=4,485, and CA and HF patients in EMR data [cohort 4a (CA): N=1,789, cohort 4b (HF): N=1,789]. EMR data was sourced from Optum (n= 88 million US patients, up to 10 years of medical history). The CA cohort comprised patients with organ-specific amyloidosis and HF diagnoses. Results The model delivered robust performance in correctly predicting ATTRwt and HF patients with sensitivity/specificity/accuracy of 94/94/91%, AUC 0.97 (derivation cohort 1, test set). External validation was also successful: cohort 2, 96/95/88%; cohort 3, 64/72/77%; and cohort 4, 75/76/80%. Model output revealed key clinical features associated with ATTRwt patients: pericardial disease (OR 11.8 CI 5.3-30.1), primary intrinsic cardiomyopathy (OR 6.1 CI 4.4-8.5), carpal tunnel (OR 5.7 CI 4.3-11.8), diastolic HF (OR 5.3 CI 3.8-7.4), atrial fibrillation (OR 4.9 CI 3.4-7.2), RBBB (OR 4.5 CI 2.5-5.6), synovitis (OR 3.9 CI 2.4-6.8), pleural effusion (OR 2.7 CI 2.0-3.7), osteoarthrosis (OR 2.5 CI 1.8-3.5), lumbar spinal stenosis (OR 2.1 CI 1.5-3.1), chronic kidney disease (OR 2.1 CI 1.5-2.8), and senile cataract (OR 2.0 CI 1.5-2.8). Conclusion Our findings from a robust ML analysis confirm the clinical profile of ATTRwt patients consistent with published literature and provide additional novel insights to aid in identifying undiagnosed patients. This is the first ML approach that provides a systematic framework to identify ATTRwt patients with a high degree of accuracy among large patient-level databases.