This paper aims to predict the risk of Acute Kidney Injury (AKI) in intensive care units (ICUs) using machine learning techniques and statistical approaches. The data used in the study are derived from Medical Information Mart for Intensive Care (MIMIC) III, which is a freely accessible database of de-identified ICU-related data. The paper focuses on two different phases. The first one consists of a scrupulous phase of extraction and transformation of MIMIC data to retrieve all the criteria specified in the Kidney Disease Improving Global Outcomes (KDIGO) clinical practice guideline definition of AKI. The main features included are demographics, medications, comorbidities, charted vital signs, and laboratory events. In the second phase, we used several different techniques already used to predict AKI, and we also added a complex temporal feature, called Trend-Event Feature (TE-F). The prediction was performed using a rolling observational window design that includes the data collection window (length: 1 to 6 days) and the prediction window (7 days). The Gradient Boosting Decision Trees (GBDT) method was used, with different groups of features, to predict the risk of AKI. To evaluate the GBDT performances, we used the area under the ROC curve (AUROC), Recall, Precision, and F-score measures. We observed that lab parameters contribute the most to the prediction of the AKI risk. Moreover, adding TE-Fs to different groups of features leads to better performance results.

Acute Kidney Injury Prediction with Gradient Boosting Decision Trees enriched with Temporal Features

Golovco, Stela;Mantovani, Matteo
;
Combi, Carlo
;
Holmes, John H.
2022-01-01

Abstract

This paper aims to predict the risk of Acute Kidney Injury (AKI) in intensive care units (ICUs) using machine learning techniques and statistical approaches. The data used in the study are derived from Medical Information Mart for Intensive Care (MIMIC) III, which is a freely accessible database of de-identified ICU-related data. The paper focuses on two different phases. The first one consists of a scrupulous phase of extraction and transformation of MIMIC data to retrieve all the criteria specified in the Kidney Disease Improving Global Outcomes (KDIGO) clinical practice guideline definition of AKI. The main features included are demographics, medications, comorbidities, charted vital signs, and laboratory events. In the second phase, we used several different techniques already used to predict AKI, and we also added a complex temporal feature, called Trend-Event Feature (TE-F). The prediction was performed using a rolling observational window design that includes the data collection window (length: 1 to 6 days) and the prediction window (7 days). The Gradient Boosting Decision Trees (GBDT) method was used, with different groups of features, to predict the risk of AKI. To evaluate the GBDT performances, we used the area under the ROC curve (AUROC), Recall, Precision, and F-score measures. We observed that lab parameters contribute the most to the prediction of the AKI risk. Moreover, adding TE-Fs to different groups of features leads to better performance results.
2022
978-1-6654-6845-9
AKI , KDIGO , MIMIC , temporal windows , prediction windows , machine learning , trend-event features
File in questo prodotto:
File Dimensione Formato  
Acute_Kidney_Injury_Prediction_with_Gradient_Boosting_Decision_Trees_enriched_with_Temporal_Features.pdf

solo utenti autorizzati

Tipologia: Versione dell'editore
Licenza: Copyright dell'editore
Dimensione 12.47 MB
Formato Adobe PDF
12.47 MB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11562/1074186
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact