Clinical characteristics and health care resource use of patients at risk for wild-type transthyretin amyloid cardiomyopathy identified by machine learning model

BACKGROUND: Transthyretin amyloid cardiomyopathy (ATTR-CM) is a progressive, life-threatening systemic disorder that is an underrecognized cause of heart failure (HF). When the diagnosis of wild-type ATTR-CM (ATTRwt-CM) is delayed, patients often undergo additional assessments, deferring appropriate management as symptoms potentially worsen. Prompt recognition of patients at risk for ATTRwt-CM is essential to facilitate earlier diagnosis and disease-modifying treatment. A previously developed machine learning model performed well in identifying ATTRwt-CM in patients with HF vs controls with nonamyloid HF using medical claims/electronic health records, providing a systematic framework to raise disease suspicion. OBJECTIVE: To further evaluate this model’s performance in identifying ATTRwt-CM using a large claims database of older adults with HF and confirmed ATTRwt-CM or nonamyloid HF; and to explore the characteristics and health care resource utilization (HCRU) of patients with confirmed and suspected ATTRwt-CM. METHODS: In this retrospective study, the prior model was applied using Humana administrative claims for patients diagnosed with ATTRwt-CM (cases) and nonamyloid HF (controls [1:1]). Patients were aged 65-89 years, had at least 2 claims for HF diagnosis (2015-2020), and were continuously enrolled in a Medicare Advantage prescription drug plan for at least 12 months before and at least 6 months after HF diagnosis. For the assessment of characteristics and HCRU, the suspected risk level was categorized based on the predicted probability (PP) from model output (high, moderate, and low risk: PP≥0.70; ≥0.50 and < 0.70; and < 0.50, respectively). RESULTS: Of 267,025 eligible patients, 119 (0.04%) had confirmed ATTRwt-CM; of 266,906 patients with nonamyloid HF, 10,997 (4.1%), 68,174 (25.5%), and 187,735 (70.3%) were categorized as high, moderate, and low risk for ATTRwt-CM, respectively. The model demonstrated sensitivity/specificity/accuracy/receiver operating characteristic area under the concentration-time curve of 88%/65%/77%/0.89, respectively, in differentiating ATTRwt-CM from nonamyloid HF. In patients with confirmed ATTRwt-CM, the mean (SD) time between HF and ATTRwt-CM diagnoses was 751 (528) days; 65% and 48% were hospitalized before and after ATTRwt-CM diagnosis, respectively. Atrial fibrillation was more common in patients with confirmed ATTRwt-CM and high risk (39% and 55%) vs low risk (27%). Hospitalization and emergency department visits after HF diagnosis were reported in 57% and 46% of patients with high ATTRwt-CM risk, respectively. CONCLUSIONS: The ATTRwt-CM predictive model performed well in identifying disease risk in the Humana Research Database. Patients at high risk for ATTRwt-CM had high HCRU and may benefit from the earlier suspicion of ATTRwt-CM. The model may be used as a tool to identify patients with a suspected high risk for the disease to facilitate earlier detection and treatment.