Predicting scheduled hospital attendance with artificial intelligence.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Operations Research

Background:

• Hospital appointment no-shows incur substantial costs (£1 billion annually in the UK).
• Simple predictive models perform poorly due to the complexity of absence causes.
• Advanced machine learning models are needed for accurate risk stratification and intervention.

Purpose of the Study:

• To quantify the effectiveness of complex, high-dimensional machine learning models for predicting hospital appointment attendance.
• To compare the performance of various machine learning algorithms in predicting no-shows.
• To assess the potential economic benefits of improved attendance prediction.

Main Methods:

• Trained and evaluated logistic regression, support vector machines, random forests, AdaBoost, and gradient boosting machines.
• Utilized data from 22,318 magnetic resonance imaging appointments at two UCL hospitals.
• Identified optimal predictive performance using 81 variables with high-dimensional Gradient Boosting Machine models.

Main Results:

• High-dimensional Gradient Boosting Machine models achieved state-of-the-art performance (AUC 0.852, average precision 0.511).
• Optimal prediction required 81 variables, indicating model complexity is crucial.
• Simulations projected a net benefit of up to £3.15 per appointment.

Conclusions:

• Complex, high-dimensional machine learning models are essential for accurate hospital appointment attendance prediction.
• These models outperform traditional methods by capturing intricate patient, environmental, and operational factors.
• Effective prediction is achievable using data from a single institution, avoiding large-scale data aggregation.