Predicting pediatric diagnostic imaging patient no-show and extended wait-times using LLMs, regression, and tree...

Area of Science:

• Radiology and Medical Imaging
• Machine Learning in Healthcare
• Health Services Research

Background:

• Patient no-shows and long wait times in diagnostic and interventional radiology lead to resource inefficiency and patient dissatisfaction.
• Predictive modeling can help mitigate these issues by identifying at-risk patients.
• Optimizing appointment scheduling is crucial for improving healthcare access and operational efficiency.

Purpose of the Study:

• To survey machine learning methods for predicting pediatric patient no-shows and long wait times (exceeding 1 hour) for radiology appointments.
• To identify the most effective predictive models and data features for these tasks.
• To propose targeted downstream accommodations for scheduling teams.

Main Methods:

• Experimented with six predictive models: extreme gradient boosting (XGBoost), Random Forest (RF), Support Vector Machine, Logistic Regression, Artificial Neural Network, and a large language model (LLM).
• Utilized 20 features including patient demographics and appointment-related data.
• Employed data balancing methods like instance hardness threshold (IHT) and class weighting.
• Compared performance with and without continuous contextual data in the LLM.

Main Results:

• XGBoost achieved the best AUC (0.96) and F1 score (0.62) for no-show prediction.
• Random Forest yielded the best AUC (0.83) and F1 score (0.61) for wait-time prediction.
• LLMs showed strong performance at 90% probability thresholds and were robustly calibrated.

Conclusions:

• Contextual appointment data is more important than patient demographics for equitable prediction.
• The developed predictive models can inform risk-based scheduling to improve patient access to care.
• Further integration of predictive outputs can optimize resource allocation and reduce appointment delays.