Application of group LASSO regression based Bayesian networks in risk factors exploration and disease prediction for acute kidney injury in hospitalized patients with hematologic malignancies

Abstract

Background: This study aims to explore a novel machine-learning algorithm, Bayesian networks (BNs), to delineate the interrelationships between acute kidney injury (AKI) and its associated risk factors among patients with hematologic malignancies (HM), to assess the prediction ability of BNs model, and to infer the probability of AKI under different clinical settings. Methods: From 1 October 2014 to 30 September 2015, 2501 hospitalized patients diagnosed with HM in Zhongshan Hospital, Fudan University, Shanghai of China, were recruited in this retrospective study. Data on demographics, comorbidities, and baseline clinical lab records were exported from the electronic medical records. The group-LASSO (gLASSO) regression was performed to select the candidate predictors of AKI, which were further presented in BNs analysis for interrelationship exploration and disease prediction. Results: Among 2395 eligible patients, 370 episodes were diagnosed with AKI (15.4%). Patients with multiple myeloma (24.1%) and leukemia (23.9%) shared a higher AKI incidence than lymphoma (13.4%). Screened by the gLASSO regression, variables as age, gender, diabetes, hemopathy category, anti-tumor treatment, hemoglobin, serum creatinine (SCr), estimated glomerular filtration rate (eGFR), serum uric acid, serum sodium and potassium level were found with significant associations with AKI occurrence. BNs model revealed a complex interrelationship between these factors, among which there were direct connections between AKI and age, hemoglobin, eGFR, serum sodium and potassium. Moreover, HM category and anti-tumor treatment were indirectly linked to AKI through hemoglobin and eGFR; diabetes was connected with AKI through serum sodium level. BNs inferences indicated that when poor GFR, anemia and hyponatremia occurred simultaneously, the probability of AKI might reach 78.5%. The area under the receiver operating characteristic curve (AUC) of BNs model was 0.835, higher than that in logistic score model (0.763) and showed a robust performance in 10-fold cross-validation. Conclusion: AKI is common in HM hospitalized patients and is affected by multiple risk factors. The application of Bayesian networks and gLASSO regression can provide a novel strategy to explore the potential risk factors integrally and deep into their interrelationships, in accordance with the tide of e-alert and big-data for AKI early detection.