Abstract 4145658: Early Prediction of Inpatient Worsening Heart Failure Using Machine Learning

Background: In patients with acute heart failure (AHF), worsening heart failure (WHF) predicts mortality, prolonged length of stay and readmission. We aimed to create a predictive model to identify predictors of intensive care unit (ICU) transfer as a proxy for WHF. Methods: We developed and evaluated retrospectively a machine learning model based on the LASSO algorithm to predict ICU transfer. The algorithm was trained on 11 years of data (from 2010-2020) collected from a large regional healthcare system. We included patients admitted with acute HF, an admission BNP value >100, and excluded direct ICU admissions. The input variables include demographics, vital signs, laboratory values, pre-existing diagnoses, and treatment teams. Model performance was evaluated AUROC. Results: Of 19,351 patients with an index AHF hospitalization, 5,829 (30.1%) required ICU transfer. These patients were younger (70 ± 15 vs 73 ± 16 yrs, p<0.001), more likely to be male (54% vs 51%, p<0.001), and be White (52% vs 40%, p<0.001). These patients had higher rates of 30-day mortality (17% vs 6%, p<0.001), 30-day readmission (11% vs 8%, p<0.001), and a longer median length of stay (12 [7, 20] vs 4 [3, 7] days, p<0.001). Model performance on training and validation datasets yielded AUC=0.78 and AUC=0.78, respectively. Internal medicine primary teams predicted lowest odds of ICU transfer (OR: 0.17 [95% CI: 0.15 - 0.18], p<0.001), while atrial fibrillation predicted highest odds of ICU transfer (1.72 [95% CI: 1.59 - 1.88], p<0.001) (Fig). Conclusion: WHF during index hospitalization was associated with worse 30-day outcomes and longer length of stay. Our model has excellent risk prediction for worsening HF defined by ICU transfer during HF hospitalization.