Predicting Patient No-Shows: Situated Machine Learning with Imperfect Data.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Operations Research

Background:

• Patient no-shows represent a significant financial burden and operational challenge in healthcare settings.
• Accurate prediction of patient no-shows is crucial for optimizing resource allocation and improving patient flow.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting patient no-shows for outpatient surgery.
• To assess the impact of situated work and staff engagement on model performance.
• To identify the best performing ML model for no-show prediction in an endoscopy ward.

Main Methods:

• Utilized historical patient data not originally collected for ML purposes.
• Employed situated work within the hospital to understand data practices and refine models.
• Trained and compared various ML models, including XGBoost with oversampling.
• Evaluated model performance using sensitivity, specificity, and accuracy metrics.

Main Results:

• The best performing model (XGBoost with oversampling) achieved a sensitivity of 0.97, specificity of 0.66, and accuracy of 0.95.
• Situated work and staff engagement led to significant quantitative improvements in model performance.
• The developed models demonstrated high predictive power for patient no-shows.

Conclusions:

• Machine learning models, particularly XGBoost with oversampling, show strong potential for predicting patient no-shows in outpatient surgery settings.
• Collaborative design involving hospital staff is critical for enhancing the performance and applicability of ML models.
• While promising, the generalizability of these models to other hospital wards and institutions requires further validation and adaptation.