Multicenter validation of a scalable, interpretable, multitask prediction model for multiple clinical outcomes

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Perioperative Medicine

Background:

• Predicting multiple postoperative complications is difficult, hindering integrated risk assessment.
• Current methods often assess complications individually, limiting comprehensive preoperative evaluation.

Purpose of the Study:

• To develop and validate a multitask learning model for predicting key postoperative complications.
• To assess the model's performance in predicting acute kidney injury (AKI), postoperative respiratory failure (PRF), and in-hospital mortality.

Main Methods:

• Developed a scalable, interpretable, tree-based multitask learning model.
• Utilized 16 preoperative features from electronic health records.
• Externally validated the model on two independent cohorts.

Main Results:

• The model demonstrated high predictive accuracy for AKI, PRF, and mortality across derivation and validation cohorts (AUROCs ranging from 0.789 to 0.925).
• Significant predictive performance was observed for all outcomes (p < 0.001 in most cases).
• The model elucidated variable contributions to predictions for different outcomes.

Conclusions:

• Multitask learning offers a streamlined approach to preoperative risk assessment.
• The developed model provides a scalable, interpretable, and generalizable framework for improving perioperative care outcomes.