TLDR: This research integrates machine learning (ML) with Answer Set Programming (ASP) to improve operating room scheduling. By using ML (specifically XGBoost) to predict surgery durations and incorporating prediction confidence into the ASP model, the approach generates more robust and efficient schedules. Experimental results on real hospital data show improved operating room usage and reduced overbooking/underbooking compared to traditional methods.
Optimizing daily surgical schedules in operating rooms, known as the Operating Room Scheduling (ORS) problem, is a complex challenge for hospitals. It involves determining surgery start times and allocating crucial resources like operating rooms and beds, all while facing numerous constraints. Traditional solutions, particularly those based on Answer Set Programming (ASP), have shown promise in finding feasible and efficient schedules. However, a significant limitation arises when applying these solutions to real-world scenarios: they often assume surgery durations are known in advance. This assumption prevents the generation of provisional schedules and can lead to less robust outcomes when actual durations differ from estimates.
A recent research paper titled “Improving ASP-based ORS Schedules through Machine Learning Predictions” by Pierangela Bruno, Carmine Dodaro, Giuseppe Galatà , Marco Maratea, and Marco Mochi addresses these limitations by integrating machine learning (ML) predictions into the ASP-based ORS framework. This innovative approach combines inductive (ML-based) and deductive (logic-based) techniques to create more accurate and robust surgical schedules.
The core of their solution involves using machine learning algorithms to predict surgery durations based on historical data. The researchers first performed a thorough analysis of a real-world dataset from ASL1 Liguria in Italy, identifying and addressing data quality issues like distribution skewness. They applied preprocessing steps, including grouping rare diagnoses, filtering extreme duration values using the Interquartile Range (IQR) method, and removing highly correlated features to improve data quality and model stability. This reduced the feature dimensionality from 32 to 23, making the models more robust and interpretable.
Several state-of-the-art ML algorithms were evaluated for predicting surgery durations, including Decision Tree, Random Forest, Gradient Boosting, XGBoost, K-Nearest Neighbors, and Support Vector Regressor. Deep learning models like Multi-Layer Perceptron (MLP), 1D Convolutional Neural Network (1D-CNN), and TabNet were also tested. Among all the models, XGBoost consistently delivered the best performance, achieving the lowest Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), and the highest Coefficient of Determination (R²), indicating its superior accuracy in predicting surgery durations. Even with the inherent skewness in the data, XGBoost’s average error was around 12 minutes, a level of accuracy deemed suitable for hospital operations when combined with confidence estimation.
A key innovation in this paper is the introduction of ‘prediction confidence.’ After predicting surgery durations, the researchers calculated the Absolute Percentage Error (APE) for each prediction. Based on this error, predictions were categorized into four confidence levels: High (APE < 10%), Moderate (10% ≤ APE < 25%), Low (25% ≤ APE < 50%), and Very Low (APE ≥ 50%). This confidence information is then fed into the ASP encoding, allowing the scheduling system to make more informed and robust decisions. The ASP model is extended with rules that prefer balanced and high-confidence scheduling decisions, aiming to distribute patients with high confidence values equally among operating rooms and days.
Experimental evaluations on real-world data from ASL1 Liguria demonstrated the effectiveness of this neuro-symbolic approach. The method incorporating XGBoost predictions with confidence information (referred to as ‘Conf.’) consistently outperformed traditional methods that rely on average durations (average by department or surgical procedure). For instance, in Bordighera, the ‘Conf.’ method achieved a mean OR occupancy of 96%, very close to the ideal 100%, with better control over overbooking and underbooking compared to other methods. Similar improvements were observed in Imperia and Sanremo, where the ‘Conf.’ method led to more balanced and operationally feasible schedules, maximizing OR usage while minimizing disruptions.
Also Read:
- A New Framework for Temporal Reasoning with Constraints in Answer Set Programming
- Optimizing Industrial Assembly Lines with AI: A New Deep Reinforcement Learning Approach
This research highlights the significant potential of integrating machine learning with symbolic reasoning, particularly ASP, to solve complex real-world problems like ORS. By providing more accurate provisional schedules and enhancing their robustness through confidence-aware decision-making, this approach can lead to more efficient hospital operations and better patient care. For more details, you can refer to the full research paper available at https://arxiv.org/pdf/2507.16454.
