Evaluating the kidney disease progression using a comprehensive patient profiling algorithm: A hybrid clustering approach

Background Among 35.5 million U.S. adults with chronic kidney disease (CKD), more than 557,000 are on dialysis with incurred cost ranges from $97,373 to $102,206 per patient per year. Acute kidney injury (AKI) can lead to an approximate ninefold increased risk for developing CKD. Significant knowledge gaps exist in understanding AKI to CKD progression. We aimed to develop and test a hybrid clustering algorithm to investigate the clinical phenotypes driving AKI to CKD progression. Methods This retrospective observational study utilized data from 90,602 patient electronic health records (EHR) from 2010 to 2022. We classified AKI into three groups: Hospital Acquired AKI (HA-AKI), Community Acquired AKI (CA-AKI), and No-AKI. We developed a custom phenotypic disease and procedure network and a complementary variable clustering to examine risk factors among three groups. The algorithm identified top three matched clusters. Results Among 58,606 CKD patients, AKI group had a higher prevalence of heart failure (21.1%) and Type 2 Diabetes (45.3%). The No-AKI group had a higher comorbidity burden compared to AKI group, with average comorbidities of 2.84 vs. 2.04; p < 0.05; 74.6% vs. 53.6%. Multiple risk factors were identified in both AKI cohorts including long-term opiate analgesic use, atelectasis, history of ischemic heart disease, and lactic acidosis. The comorbidity network in HA-AKI patients was more complex compared to the No-AKI group with higher number of nodes (64 vs. 55) and edges (645 vs. 520). The HA-AKI cohort had several conditions with higher degree and betweenness centrality including high cholesterol (34, 91.10), chronic pain (33, 103.38), tricuspid insufficiency (38, 113.37), osteoarthritis (34, 56.14), and removal of GI tract components (37, 68.66) compared to the CA-AKI cohort. Conclusion Our proposed custom patient profiling algorithm identifies AKI phenotypes based on comorbidities and medical procedures, offering a promising approach to identify early risk factors for CKD using large EHR data.