How Unsupervised Machine Learning Solves the Problem of Last-Minute Cancellations in Pediatric Surgery

The research team from IRCCS Ospedale Pediatrico Bambino Gesù in Rome, Italy, targeted the pain point of last-minute cancellations in pediatric outpatient/day surgery. They used unsupervised machine learning techniques (Factor Analysis of Mixed Data, FAMD + K-means clustering) to analyze 1773 cancellation cases, identifying three patient group characteristics and providing an interpretable solution for hospital resource optimization. This thread will break down the research background, methods, findings, and application suggestions in separate floors.

In hospital operations, last-minute surgery cancellations are a long-standing pain point, especially in pediatric outpatient/day surgery centers: they not only cause idle waste of operating rooms, medical staff, and equipment but also may delay treatment opportunities for children. The team from Rome IRCCS Hospital, one of Italy's largest pediatric hospitals, developed an analysis solution based on unsupervised machine learning to address this issue.

The study analyzed 1773 canceled pediatric outpatient/day surgery cases from January 2020 to March 2022. The data came from the hospital's appointment system, including multi-dimensional information such as patient age, surgical department, type of surgery, and waiting time. This study uses unsupervised learning—algorithms do not require pre-set labels and independently discover hidden patterns in data.

Mixed Data Clustering Process

1. FAMD Dimensionality Reduction: Handles medical data containing both numerical and categorical variables;
2. K-means Clustering: Evaluates clustering quality using Calinski-Harabasz index (inter-cluster separation), Davies-Bouldin index (intra-cluster compactness), Dunn index (comprehensive effect), and silhouette coefficient (sample assignment rationality);
3. Statistical Validation: Uses Kruskal-Wallis test and paired Wilcoxon test to verify significance, with all p-values <0.001, indicating significant differences among the three groups.

Characteristics of Three Patient Groups

• Group 1: Infants aged 0-1 year, urology department (circumcision, cryptorchidism, etc.), waiting time 12-32 days, cancellation rate 97.92% (highest);
• Group 2: Preschool children aged 1-2 years, plastic surgery department (nevus removal, scar repair, etc.), waiting time 1-2 days, cancellation rate ~45% (affected by pre-operative preparation);
• Group 3: Children aged 2-10 years, maxillofacial surgery department (tooth extraction, cyst removal, etc.), waiting time 58-194 days, cancellation rate ~47% (due to acute infection or parental time conflicts).

Key Findings

1. Nonlinear Relationship Between Age and Cancellation Rate: Infants aged 0-1 have the highest cancellation rate (due to immune development, strict pre-operative fasting, and parental anxiety); cancellations among school-age children/adolescents are related to academic/activity conflicts;
2. Department Differences: The cancellation rate of urology department is significantly higher than other departments (p=0.003), reasons include high elective nature of surgeries, parents' high expectations for timing flexibility, and strict pre-operative infection screening;
3. Golden Window for Waiting Time: Too short (1-2 days) leads to insufficient preparation; too long (>60 days) leads to condition changes or parental loss of patience. It is recommended to control it within 10-30 days.

Interpretability Highlight

Through feature importance analysis, the contribution of variables is clarified: patient age (1), surgical department (2), waiting time (3), type of surgery (4), making model decisions transparent and avoiding the "black box" issue.

Practical Application Recommendations

1. Differentiated Appointments: Provide pre-operative parent education + fast re-appointment channels for infants; optimize preparation checklists + flexible time options for preschool children; develop online self-assessment tools for adolescents;
2. Dynamic Resource Allocation: Reserve backup operating rooms on surgery days for high-risk groups, adjust department appointment quotas, and establish cancellation early warning systems;
3. Quality Improvement Cycle: Incorporate into the CQI framework, update the model quarterly to track effects.

Limitations and Future Directions

• Limitations: Single-center design, data from the pandemic period may affect generalizability;
• Future: Multi-center validation, introduce supervised learning for risk prediction, integrate EHR to enrich variables, develop real-time decision support systems.