Machine Learning Improves the Identification of Individuals With Higher Morbidity and Avoidable Health Costs After Acute Coronary Syndromes
Publicado em: Dec 2020
Autores
Resumo
Traditional risk scores improved the definition of the initial therapeutic strategy in acute coronary syndrome (ACS), but they were not designed for predicting long-term individual risks and costs. In parallel, attempts to directly predict costs from clinical variables in ACS had limited success. Thus, novel approaches to predict cardiovascular risk and health expenditure are urgently needed. Our objectives were to predict the risk of major/minor adverse cardiovascular events (MACE) and estimate assistance-related costs. We used a 2-step approach that: (1) predicted outcomes with a common pathophysiological substrate (MACE) by using machine learning (ML) or logistic regression (LR) and compared with existing risk scores; (2) derived costs associated with noncardiovascular deaths, dialysis, ambulatory-care-sensitive-hospitalizations (ACSH), strokes, and MACE. With consecutive ACS individuals (n = 1089) from 2 cohorts, we trained in 80% of the population and tested in 20% using a 4-fold cross-validation framework. The 29-variable model included socioeconomic, clinical/lab, and coronarography variables. Individual costs were estimated based on cause-specific hospitalization from the Brazilian Health Ministry perspective. After up to 12 years follow-up (mean = 3.3 ± 3.1; MACE = 169), the gradient-boosting machine model was superior to LR and reached an area under the curve (AUROC) of 0.891 [95% CI 0.846-0.921] (test set), outperforming the Syntax Score II (AUROC = 0.635 [95% CI 0.569-0.699]). Individuals classified as high risk (>90th percentile) presented increased HbA1c and LDL-C both at <24 hours post-ACS and 1-year follow-up. High-risk individuals required 33.5% of total costs and showed 4.96-fold (95% CI 3.71-5.48, P < .00001) greater per capita costs compared with low-risk individuals, mostly owing to avoidable costs (ACSH). This 2-step approach was more successful for finding individuals incurring high costs than predicting costs directly from clinical variables. ML methods predicted long-term risks and avoidable costs after ACS.