Predicting 30-Day Readmission in Patients with Heart Failure and AKI Using a Large National Database

Background: Heart failure (HF) is the leading hospitalization diagnosis in the US and the leading cause of increased healthcare costs. Acute kidney injury (AKI) affects between 10-43% of individuals admitted with HF. Published outcomes and predictive models for those with HF and AKI are lacking, those available have poor performance. TriNetX, a global health research network, has been used to develop predictive models in other disease states. Methods: TriNetX was used to identify patients with HF prior to admission who developed AKI during hospitalization. EHR data, including diagnoses, procedures, vital signs, laboratory values during admission, as well as delta change in creatinine and weight during admission were used to predict 30-day readmission. Data were preprocessed, features were calculated, and multiple feature selection and machine learning models were evaluated. Model performance was assessed using sensitivity, specificity, area under the ROC curve (AUC), and balanced accuracy. Results: 182,765 patients were identified from TriNetX who were hospitalized with HF and had AKI. Of these 20% were readmitted within 30 days, 30% were readmitted at 90 days, 12% died within 30 days, and 17% died within 90 days. The mean age of the cohort was 75 years, with 56% being male, 71% were white, 16% black, 22% had known CKD, and 21% had DM. The median GFR on admission for the cohort was 41 ml/min/1.73 m2. The best overall model predicting 30-day readmission was a logistic regression model with random forest selection. This final model had an area under the curve of 0.63. Features of high importance in the model included creatinine, platelet count, hemoglobin and GFR. Conclusion: We identified a large cohort of patients with HF and AKI during hospital admission. These individuals had high rates of hospital readmission and death. Vast EHR data elements were extracted from the cohort and used for analysis. Our final model showed only modest performance but did illustrate that laboratory values during admission may be important for identifying those at risk for 30-day readmission. Due to limitations with a database of this nature this may not be the most robust way to evaluate outcomes in this patient population. We plan to build on these findings using local data to further solidify risk factors for outcomes in patients with HF and AKI.