K-means Clustering for Shock Classification in Pediatric Intensive Care Units
Egileak: María Rollán-Martínez-Herrera
Data: 10.08.2022
Diagnostics
Abstract
Shock is described as an inadequate oxygen supply to the tissues and can be classified in multiple ways. In clinical practice still, old methods are used to discriminate these shock types. This article proposes the application of unsupervised classification methods for the stratification of these patients in order to treat them more appropriately. With a cohort of 90 patients admitted in pediatric intensive care units (PICU), the k-means algorithm was applied in the first 24 h data since admission (physiological and analytical variables and the need for devices), obtaining three main groups. Significant differences were found in variables used (e.g., mean diastolic arterial pressure p < 0.001, age p < 0.001) and not used for training (e.g., EtCO2 min p < 0.001, Troponin max p < 0.01), discharge diagnosis (p < 0.001) and outcomes (p < 0.05). Clustering classification equaled classical classification in its association with LOS (p = 0.01) and surpassed it in its association with mortality (p < 0.04 vs. p = 0.16). We have been able to classify shocked pediatric patients with higher outcome correlation than the clinical traditional method. These results support the utility of unsupervised learning algorithms for patient classification in PICU.
BIB_text
author = {María Rollán-Martínez-Herrera},
title = {K-means Clustering for Shock Classification in Pediatric Intensive Care Units},
journal = {Diagnostics},
pages = {1932},
volume = {12},
keywds = {
shock; pediatric; unsupervised learning; k-means; stratification
}
abstract = {
Shock is described as an inadequate oxygen supply to the tissues and can be classified in multiple ways. In clinical practice still, old methods are used to discriminate these shock types. This article proposes the application of unsupervised classification methods for the stratification of these patients in order to treat them more appropriately. With a cohort of 90 patients admitted in pediatric intensive care units (PICU), the k-means algorithm was applied in the first 24 h data since admission (physiological and analytical variables and the need for devices), obtaining three main groups. Significant differences were found in variables used (e.g., mean diastolic arterial pressure p < 0.001, age p < 0.001) and not used for training (e.g., EtCO2 min p < 0.001, Troponin max p < 0.01), discharge diagnosis (p < 0.001) and outcomes (p < 0.05). Clustering classification equaled classical classification in its association with LOS (p = 0.01) and surpassed it in its association with mortality (p < 0.04 vs. p = 0.16). We have been able to classify shocked pediatric patients with higher outcome correlation than the clinical traditional method. These results support the utility of unsupervised learning algorithms for patient classification in PICU.
}
doi = {10.3390/diagnostics12081932},
date = {2022-08-10},
}
