Domain-Aware Interpretable Machine Learning Model for Predicting Postoperative Hospital Length of Stay from...

Area of Science:

• Medical Informatics
• Surgical Outcomes Research
• Machine Learning in Healthcare

Background:

• Postoperative hospital length of stay (LOS) is a critical metric for surgical recovery and resource utilization.
• Predicting LOS is challenging due to diverse patient recovery patterns.
• Accurate LOS prediction can optimize hospital resource allocation and patient care planning.

Purpose of the Study:

• To develop and validate an interpretable machine learning framework for predicting postoperative LOS.
• To integrate multimodal perioperative data for enhanced prediction accuracy.
• To identify key clinical drivers of prolonged hospitalizations.

Main Methods:

• Utilized a large dataset of 97,937 adult surgical cases from a perioperative registry.
• Employed a supervised regression model incorporating demographics, comorbidities, lab values, intraoperative physiology, and procedure details.
• Validated the model using internal cross-validation and an independent holdout set, assessing cohort and individual performance.

Main Results:

• The model demonstrated strong predictive performance with R²=0.61 and MAE≈1.34 days on the holdout set.
• Key predictors of LOS included operative duration, diagnostic complexity, intraoperative hemodynamic variability, and preoperative albumin and hematocrit levels.
• Prolonged hospitalization was associated with complex procedures, specific diagnoses (malignant, respiratory), and lower albumin levels.

Conclusions:

• Interpretable machine learning provides accurate and generalizable LOS predictions.
• The framework reveals actionable perioperative insights into factors influencing recovery.
• This approach supports efficient perioperative planning, resource management, and personalized patient recovery strategies.