Real-world data-driven risk prediction of hospitalization for heart failure in non-diabetic CKD

BACKGROUND: Heart failure (HF) is a major cause of cardiovascular morbidity and mortality. Despite recent advances in diagnosis and management of HF, the prognosis remains poor. HF and chronic kidney disease (CKD) are interlinked chronic health conditions. The availability of large volume of patient data and modern analytic techniques opens new opportunities for identification of individuals at elevated risk of HF. PURPOSE: Develop risk prediction model for HF hospitalizations (HHF) in patients with non-diabetic CKD by applying data-driven computational intelligence techniques to a US population-based administrative claims database. METHODS: Individual-level data from the US Optum Clinformatics Data Mart for years 2008–2018 were analysed. To be eligible for inclusion, adult individuals were required to have non-diabetic CKD stage 3 or 4 (index event) and one year continuous insurance coverage prior to the index date (baseline period). Selection criteria and the main clinical outcome, hospitalisation for heart failure (HHF), were identified by using laboratory tests results and/or specific codes from common clinical coding systems. Risk prediction model for HHF was built on patient data in the baseline period composed to more than 6,000 variables. Computational intelligence method based on ant colony optimization was used to develop a time-to-first-event risk prediction model for HHF. RESULTS: Of the 64 million individuals in the database, 504,924 satisfied the selection criteria. Median age was 75 years, 60% were female. Among most common baseline comorbidities were hypertension (85%) and hyperlipidaemia (68%). Coronary artery disease, HF, atrial fibrillation and peripheral artery disease were recorded in 24%, 16%, 15% and 14% of individuals. Over a median follow-up of 744 days, 53,282 (11%) patients had recorded HHF, the corresponding incidence rate was 3.95 events/100 patient-years. The developed risk prediction model for HHF in non-diabetic CKD contained 20 risk factors. The five strongest risk factors were history of HF, intake of loop diuretics, severely increased albuminuria, atrial fibrillation or flutter and CKD 4 as observed “yes/no” in the baseline period. Fig. 1 depicts the final risk prediction model. To assess model performance, all patients in the cohort were stratified into five HHF risk groups. For each group, a Kaplan-Meier curve was built based on the HHF outcome data in the database. Fig. 2 shows clear separation between the curves, demonstrating high performance of the developed risk prediction model. CONCLUSION: Despite many existing scores to predict HHF, their use is limited. Some scores rely on availability of rarely collected information, some are applicable for specific patient populations only. Risk prediction model for HHF in non-diabetic CKD is presented, which contains risk factors routinely collected by healthcare providers. Therefore, it might be applicable for HHF risk estimation in various settings. FUNDING ACKNOWLEDGEMENT: Type of funding sources: Other. Main funding source(s): Bayer AG