ML-Based AKI Prediction in Acute Pancreatitis: Innovative Models from MIMIC-IV Database

Abstract

Abstract Background Acute kidney injury (AKI) constitutes a prevalent and deleterious complication in the context of severe acute pancreatitis (AP), underscored by elevated mortality rates and substantial disease burden. Given its substantial clinical ramifications, the early anticipation of AKI assumes paramount significance, facilitating prompt intervention and ultimately engendering an improved prognosis. This study is poised to forge novel avenues by crafting and validating predictive models hinged upon innovative machine learning (ML) algorithms, tailored to discern the emergence of AKI among critically ill individuals grappling with acute pancreatitis. Methods The dataset encompassing patients beset by acute pancreatitis was meticulously extracted from the comprehensive repository, Medical Information Mart for Intensive Care IV (MIMIC- IV) database. Within this construct, feature selection was diligently executed via the employment of the random forest methodology. The orchestration of model construction hinged upon an ensemble of ML algorithms—namely, random forest (rf), support vector machine (svm), k-nearest neighbors (knn), naive Bayes (nb), neural network (nnet), logistic regression (glm), and gradient boosting machine (gbm). This orchestration was facilitated through the meticulous deployment of tenfold cross-validation. The discriminatory capacity of each model was rigorously gauged by assessing the cross-validated area under the receiver operating characteristic curve. Subsequently, the model attaining superior performance was meticulously fine-tuned, and its ultimate prowess was comprehensively assessed via split-set validation. Results An aggregate of 1,235 critically ill patients afflicted by acute pancreatitis were meticulously encompassed within our analytical purview, within which 667 cases (54%) manifested the onset of AKI during the trajectory of hospitalization. A comprehensive selection of 50 variables was marshaled for the elaborate edifice of model construction. The constellation of models encompassing gbm, glm, knn, nb, nnet, rf, and svm was meticulously instantiated, yielding area under the receiver operating characteristic curves quantified at 0.814 (95% CI, 0.763 to 0.865), 0.812 (95% CI, 0.769 to 0.854), 0.671 (95% CI, 0.622 to 0.719), 0.812 (95% CI, 0.780 to 0.864), 0.688 (95% CI, 0.624 to 0.752), 0.809 (95% CI, 0.766 to 0.851), and 0.810 (95% CI, 0.763 to 0.856) respectively. In this constellation, the gradient boosting machine model emerged as the vanguard, standing testament to its preeminent predictive proficiency across both discrimination and calibration domains. The gradient boosting machine's performance in the test set was mirrored by an area of 0.867 (95% CI, 0.831 to 0.903). Conclusions The triumph engendered by this methodological paradigm, culminating in the anticipation of AKI within acute pancreatitis patients, augurs well for the viability and promise of machine learning models as potent instruments for predictive analytics within the critical care arena. Efficacy concomitant with the selected model and its judicious fine-tuning stands as a pivotal determinant in this predictive orchestration. Notably, the gbm model, distinguished by its optimal predictive precision, proffers an invaluable compass for clinicians, facilitating the discernment of high-risk patients, and, in tandem, instating timely interventions with a view to curbing mortality rates.