Abstract
There is limited evidence on how social determinants of health (SDOH) and physical frailty (PF) influence mortality prediction in heart failure (HF), particularly for in-hospital, 90-day, and 1-year outcomes. This study aims to develop explainable machine learning (ML) models to assess the prognostic value of SDOH and PF at multiple time points. We analyzed data from adult patients admitted to the intensive care unit (ICU) for the first time with a diagnosis of HF. Key variables extracted from electronic health records included SDOH (e.g., primary language, insurance type), PF indicators (Braden mobility, nutrition, activity, and fall risk scores), vital signs, laboratory tests, and lung sounds (LS) from both ICU admission and discharge. We employed the eXtreme Gradient Boosting (XGBoost) algorithm to build models for short- and long-term mortality prediction, and used SHapley Additive exPlanations (SHAP) to interpret model outputs and quantify the importance of each feature. The observed mortality rates were 14.8% in-hospital (n = 12,856), 7.0% at 90 days (n = 10,990), and 13.5% at 1 year (n = 10,221). The prediction models achieved area under the receiver operating characteristic curve (AUROC) scores of 0.836 (95% CI: 0.831-0.844) for in-hospital, 0.790 (95% CI: 0.780-0.800) for 90-day, and 0.789 (95% CI: 0.780-0.799) for 1-year mortality. These models outperformed baseline ML algorithms and conventional clinical risk scores. Key predictors of HF outcomes included age, fall risk, primary language, blood urea nitrogen, comorbidities, urine output, insurance type, and LS findings. Incorporating PF at ICU admission and discharge, along with SDOH such as language proficiency and insurance status, could enhance the identification of high-risk HF patients and may inform targeted interventions.