Chang Yin Chionh, FASN, MBBS
Changi General Hospital, Singapore
"Background: Patients admitted from the Emergency Department (ED) are at risk of sudden clinical deterioration. Scoring systems such as the National Early Warning Score 2 (NEWS2) were developed for mortality prediction but show limited specificity and variable performance across populations.
This study evaluates a machine learning (ML) model to address these limitations, comparing its discriminative performance against NEWS2 for predicting acute deterioration (ICU/high dependency admission, Emergency/Code Blue activation or death) within 72h of admission.
Methodology: ED medical admission data (2020–2023) was obtained, including demographics, admission data (triage class, diagnosis), vital signs, lab results, medications, and historical healthcare attendances. The target was acute deterioration as defined above. A CatBoost classifier was trained using 75% of the dataset and evaluated on the remaining 25%.
Results: The test set had 13,263 medical admissions, with median age 69y and 45.2% females. There were 382 (2.88%) acute deterioration events. Most significant predictive data points include ED diagnosis, triage class, clinical parameters and common lab tests.
Due to low event rate, NEWS2 AUPRC was low at 0.066 as anticipated. Our ML model improved the AUPRC to 0.127.
Conclusion: The ML model demonstrated better discrimination than NEWS2 for predicting early clinical deterioration, which demonstrates the potential of data models in clinical risk stratification."
This study evaluates a machine learning (ML) model to address these limitations, comparing its discriminative performance against NEWS2 for predicting acute deterioration (ICU/high dependency admission, Emergency/Code Blue activation or death) within 72h of admission.
Methodology: ED medical admission data (2020–2023) was obtained, including demographics, admission data (triage class, diagnosis), vital signs, lab results, medications, and historical healthcare attendances. The target was acute deterioration as defined above. A CatBoost classifier was trained using 75% of the dataset and evaluated on the remaining 25%.
Results: The test set had 13,263 medical admissions, with median age 69y and 45.2% females. There were 382 (2.88%) acute deterioration events. Most significant predictive data points include ED diagnosis, triage class, clinical parameters and common lab tests.
Due to low event rate, NEWS2 AUPRC was low at 0.066 as anticipated. Our ML model improved the AUPRC to 0.127.
Conclusion: The ML model demonstrated better discrimination than NEWS2 for predicting early clinical deterioration, which demonstrates the potential of data models in clinical risk stratification."
