Volume 79, Issue 7 (October 2021)
Tehran Univ Med J 2021, 79(7): 568-572 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Mohammadi Taghiabad Z, Ahmadi M, Atashi A. Prediction of mortality in patients admitted to intensive care units, A comparison of three data mining techniques: a brief report.. Tehran Univ Med J 2021; 79 (7) :568-572
URL: http://tumj.tums.ac.ir/article-1-11367-en.html
URL: http://tumj.tums.ac.ir/article-1-11367-en.html
1- Department of Health Information Management, School of Health Management and Information Sciences, Iran University of Medical Sciences, Tehran, Iran.
2- Department of Electronic Health, Virtual School, Tehran University of Medical Sciences, Tehran, Iran.
2- Department of Electronic Health, Virtual School, Tehran University of Medical Sciences, Tehran, Iran.
Abstract: (1264 Views)
Background: Early outcome prediction of hospitalized patients is critical because the intensivists are constantly striving to improve patients' survival by taking effective medical decisions about ill patients in Intensive Care Units (ICUs). Despite rapid progress in medical treatments and intensive care technology, the analysis of outcomes, including mortality prediction, has been a challenge in ICUs. Hence, this study aims to predict the mortality of patients admitted to ICUs using data mining techniques.
Methods: In this study, among the cases of patients who were admitted to ICUs of the Rasoul Akram and Firoozgar hospitals of Tehran City, Iran, from December 2017 to March 2018, the first 24 hours of the ICUs admission data of 874 cases were gathered. A new model based on the standard methodology CRISP was developed. In the modeling section, two well-known data mining techniques called artificial neural network (ANN), K nearest neighbor (KNN) and decision tree (DT) were used. WEKA 3.9.2 open-source software was implemented for data analysis. Finally, according to the accuracy, sensitivity, specificity criteria and AUC-ROC Curve, the superior model was introduced.
Results: Based on the WEKA results, 19 variables had the most impact on the mortality prediction of patients admitted to ICUs including Glasgow Coma Scale (GCS), mechanical ventilation, surgical service at ICUs admission, gender, temperature, serum creatinine, diabetes, Blood urea nitrogen (BUN), age, addiction, International Normalized Ratio (INR), PH, Partial Thromboplastin Time (PTT), albumin, hemoglobin, glucose, pulse rate, hematocrit (HCT), PO2. Based on the created models, some rules have been extracted which can be used as a pattern to predict the probability of mortality. Although the AUC of the three models was acceptable (KNN 81.5%, ANN 77.5% and DT 74.3%), but the accuracy of decision tree J48 (74.2%) was higher.
Methods: In this study, among the cases of patients who were admitted to ICUs of the Rasoul Akram and Firoozgar hospitals of Tehran City, Iran, from December 2017 to March 2018, the first 24 hours of the ICUs admission data of 874 cases were gathered. A new model based on the standard methodology CRISP was developed. In the modeling section, two well-known data mining techniques called artificial neural network (ANN), K nearest neighbor (KNN) and decision tree (DT) were used. WEKA 3.9.2 open-source software was implemented for data analysis. Finally, according to the accuracy, sensitivity, specificity criteria and AUC-ROC Curve, the superior model was introduced.
Results: Based on the WEKA results, 19 variables had the most impact on the mortality prediction of patients admitted to ICUs including Glasgow Coma Scale (GCS), mechanical ventilation, surgical service at ICUs admission, gender, temperature, serum creatinine, diabetes, Blood urea nitrogen (BUN), age, addiction, International Normalized Ratio (INR), PH, Partial Thromboplastin Time (PTT), albumin, hemoglobin, glucose, pulse rate, hematocrit (HCT), PO2. Based on the created models, some rules have been extracted which can be used as a pattern to predict the probability of mortality. Although the AUC of the three models was acceptable (KNN 81.5%, ANN 77.5% and DT 74.3%), but the accuracy of decision tree J48 (74.2%) was higher.
|
Conclusion: The study indicated that in the KNN model, the rules derived from it can be effective in mortality prediction in patients admitted to ICUs.
|
Type of Study: Brief Report |
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
