Machine learning without borders? An adaptable tool to optimize mortality prediction in diverse clinical settings

Area of Science:

• Trauma care research
• Clinical informatics
• Machine learning applications

Background:

• Mortality prediction is crucial for clinical decision-making and quality improvement.
• Traditional metrics require specific variables and diagnostics, limiting use in resource-limited settings.
• Machine learning is hypothesized to improve mortality prediction across diverse economic contexts.

Purpose of the Study:

• To evaluate the efficacy of machine learning for predicting mortality in trauma patients.
• To compare machine learning performance against standard scoring methods.
• To identify key clinical variables influencing mortality prediction across different healthcare settings.

Main Methods:

• The SuperLearner ensemble machine-learning algorithm was applied to three prospective trauma cohorts (US, South Africa, Cameroon).
• Cross-validation and receiver operating characteristic curves assessed model discrimination of discharge mortality.
• SuperLearner performance was compared with conventional scoring algorithms, and driving variables were analyzed.

Main Results:

• SuperLearner demonstrated superior prediction of discharge mortality in the US (AUC, 94-97%) and Cameroon (AUC, 90-94%).
• Performance was comparable to standard scores in the South African cohort (AUC, 90-95%).
• Site-specific variables (e.g., partial thromboplastin time, hospital distance) and severe brain injury were key predictors.

Conclusions:

• Machine learning offers excellent discrimination for injury mortality across various settings.
• Data-adaptive machine learning methods optimize site-specific predictions, overcoming limitations of traditional scores.
• This approach supports individualized decision-making and enhances quality improvement programs globally.