Back
The application of methods of artificial intelligence (AI) has been widely used in several fields, and recently it has resulted to be an essential tool to speed up the process of data analysis, diagnosis, and prediction for the study of diseases. In particular, determining the sub-phenotification of patients with acute kidney injury (AKI) has had an impact not only in handling large amounts of data but also because of the different types of data generated. In this respect, the development of computational methods aimed at resolving and proposing solution schemes that characterize clinical data for this multi-systemic syndrome is of great relevance. In this study we design and implement an innovative computational methodology that automates the process of pattern identification and the construction of AKI patient groups by associating clinical data variables, and at the same time efficiently integrating AI algorithms.
alt="Figure 1. Input data of Acute Kidney Injury patients with COVID-19">
The implemented methodology utilizes as input data clinical variables and calculated metrics associated with AKI patients. This methodology is presented in 6 modules which begin with the data collection obtained from different sources of COVID-19 patients. The data normalization and standardization were carried out and imputation methods were used in the data preprocessing module. The identification of relevant variables associated with mortality was computed by feature selection algorithms and validated by an expert in the clinical nephrology area. A central part is the application of the k-means method for the construction of patient clusters, and subsequently the determination of critical states according to the clinical characteristics per cluster. The use of silhouette index and statistical tests were used for the validation of clusters. The patient groups obtained are associated with the statistical metrics of relevant variables that define the critical stratification of patients. The integration of AI algorithms, functions, and computational scripts for each task of the strategy was implemented and executed in Matlab (V. 2022).
To test the functionality of the methodology and the assessment of patients with renal disease different sets of clinical variables and parameter configuration of the algorithms were designed. A database of 2,934 records of COVID-19 patients with 278 clinical variables was utilized. For each set of initial variables selected all the steps of the methodology were applied and validated by a nephrologist. The optimal number of relevant features according to the selection algorithms were 8 and 10 variables which allowed to identification of the critical clusters of AKI patients. In the execution of k-means, different sizes of clusters (k=2,..,10) were applied according to several relevant features, obtaining a better characterization with 3 clusters which resulted in a major impact on the construction of relevant groups of patients.
We propose a new methodology for identifying patterns in critical groups of patients with AKI based mainly on the association of relevant variables with mortality. This approach presents an automated analysis of data handling and the integration of efficient AI algorithms that permit strengthening the analysis and the formation of clusters. This research contributes to subsequent decision-making that is of interest in the evolution of AKI as well as the implementation of specialized treatments.