Abstract: FR-PO122
A Global Data Approach to Prediction of 30-Day Readmission among Patients with Heart Failure and AKI
Session Information
- AKI: Diagnosis and Outcomes
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical, Outcomes, and Trials
Authors
- McAdams, Meredith C., The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Elnakieb, Yaser, The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Parikh, Samir M., The University of Texas Southwestern Medical Center, Dallas, Texas, United States
Background
Heart failure (HF) is the leading hospitalization diagnosis in the US. Acute kidney injury (AKI) affects between 10-43% of individuals admitted with HF. Published outcomes models for HF complicated by AKI have poor predictive performance due to small patient numbers, single center data, and variable definitions of AKI.
Methods
The global healthcare database TriNetX Research was used to identify patients hospitalized with HF and AKI. Electronic health record (EHR) data including diagnoses, procedures, demographics, vital signs, medications, and laboratory values were obtained. Variables were encoded and pre-processed and the data was split into training and testing cohorts. Various machine learning models were used to build preliminary risk prediction models for 30-day readmission with 5-fold cross-validation. Initial models included only information available during the index hospitalization, no historical data was used. Model performance was evaluated on the hold-out test sets, assessing accuracy, sensitivity, specificity, area under the ROC curve (AUC), and balanced accuracy to comprehensively measure each model’s predictive power.
Results
Of 250 million unique patients in TriNetx, 271,388 patients hospitalized with HF and AKI were identified. Of these individuals, 78,806 (29%) were readmitted within 30 days. A preliminary logistic regression model for 30-day readmission had an area under the curve (AUC) of 0.63. An initial random forest model had an AUC of 0.65. Adding different feature selection techniques to the models did not improve the AUC. Variables with high importance in both the logistic and random forest models included elements of the complete blood count, blood gas, metabolic profile, urinalysis, and electrocardiogram.
Conclusion
We have identified a large cohort of patients hospitalized with HF who have concurrent AKI. Importantly, this group has a high rate of 30-day readmission. Both models performed poorly for the prediction of 30-day readmission despite converging on variables that routinely inform clinical care in this morbid population. A unique opportunity therefore exists to develop a well-powered and robust risk prediction model by adding machine learning techniques, including deep learning and transformers, and including historical and dynamic EHR data elements.