Data Preprocessing on Triage Data for Predicting ICU Transfer within 24 Hours

Automation in the medical field, particularly in Intensive Care Units (ICU), has become a primary focus of technological research, with efforts directed towards developing early diagnosis systems, condition prediction, and other applications using machine learning technology. Despite initial challenges due to limited access to sensitive patient and hospital data, research has made significant strides in enhancing the efficiency and quality of care through automation. This study evaluates the effectiveness of data preprocessing on triage data for predicting ICU transfer within a 24-hour period, comparing the performance of the Triage in Emergency Department Early Warning Score (TREWS) system with a Random Forest machine learning model. Due to the unavailability of verification for oxygen requirements and consciousness levels, the TREWS system is only partially scored, excluding these two parameters. The overall accuracy of the TREWS system was 54.23%. When dividing the data based on threshold scores of 4 and 7, the accuracy achieved was 50.13% and 55.77%. In contrast, the Random Forest model demonstrated significant improvements over the TREWS scoring system which increased the accuracy to 63% with the inclusion of additional parameters such as oxygen saturation and gender. The results suggest that integrating machine learning models with expanded parameters and effective data preprocessing can significantly enhance prediction accuracy compared to traditional scoring systems. Recommendations include optimizing traditional scoring systems with additional parameters, implementing machine learning models for better prediction, utilizing broader data sets, and providing training for medical teams on the effective use and interpretation of these models.