Multicenter validation of a machine-learning algorithm for 48-h all-cause mortality prediction.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Accurate mortality prediction is crucial for timely clinical intervention.
• Electronic medical record (EMR) data offers a rich source for developing predictive models.
• Existing scoring systems like MEWS and qSOFA have limitations in predictive accuracy.

Purpose of the Study:

• To evaluate the performance of gradient-boosted trees (GBT) for predicting inpatient mortality.
• To compare GBT predictive accuracy against logistic regression (LR), support vector machines (SVM), qSOFA, and MEWS.
• To assess GBT model generalizability across different academic health centers.

Main Methods:

• Retrospective study using EMR data from three academic health centers.
• Inclusion of inpatients aged 18+ with complete vital sign observations.
• Development and evaluation of GBT models for 12-, 24-, and 48-hour mortality prediction using area under the receiver operating characteristic curve (AUROC).

Main Results:

• GBT achieved high average AUROCs: 0.96 (12h), 0.95 (24h), and 0.94 (48h) when trained and tested within the same institution.
• When trained and tested across different institutions, GBT AUROCs reached up to 0.98 (12h), 0.96 (24h), and 0.96 (48h).
• GBT significantly outperformed LR, SVM, MEWS, and qSOFA, particularly for 48-hour predictions (average AUROCs: 0.85, 0.79, 0.86, 0.82 respectively).

Conclusions:

• Gradient-boosted trees demonstrate superior performance in predicting inpatient mortality compared to traditional methods.
• GBT models show strong generalizability across different healthcare institutions.
• These findings suggest GBT can be a valuable tool for identifying patients who may benefit from intensified clinical care.