​​Clustering Analysis of Traumatic Cervical Spine Injuries in 60,000 Patients using Machine Learning

Introduction: This study employs machine learning-based clustering algorithms to identify unique clinical profiles in patients with cervical spine trauma, providing new insights into the stratification of risk factors and treatment outcomes.

Methods: The 2015-2019 NIS was queried using ICD-10 PCS/CM coding to identify patients with cervical vertebral fracture and spinal cord injury. Machine learning clustering analysis evaluated the population based on 50 comorbidities, complications and clinical covariates. Optimal number of clusters was determined using the Davies-Bouldin Index and Calinski-Harabasz Index. Between-cluster multivariate logistic regression analysis was performed to assess risk of mortality and non-routine discharge. Kruskal-Wallis H-Testing was performed to assess variance in length-of-stay between clusters. Statistical analysis was performed using Python.

Results: 60,955 patients were clustered into 4 groups ranging from 461 to 54,188 patients. Mortality ranged from 4.48% in Cluster 1 to 29.72% in Cluster 4. Clusters 2-4 each displayed significantly higher risk of mortality [OR Range 1.51 - 10.08, p< 0.001] relative to Cluster 1. Individual cluster profiles are visualized in a heatmap. Cluster 1 had the greatest prevalence of Arrhythmia, Dementia, Depression, Diabetes, Hyperlipidemia and Hypertension amongst groups. Cluster 4 had the greatest prevalence of Acute Kidney Failure, Pneumonia, Sepsis and placement on Mechanical Ventilation.
Risk of non-routine discharge was lowest in Cluster 2 [OR 0.61, p< 0.001] and highest in Cluster 4 [OR 3.95, p< 0.001] when compared to Cluster 1. Kruskall-Wallis H testing and post-hoc pairwise testing of length-of-stay distributions showed significant (p < 0.001) differences between all clusters except 1 and 2 with the greatest test statistics occurring when comparing group 4 to all other groups.

Conclusion : Clustering analysis of patients with cervical spinal injury identified 4 distinct groups with unique comorbidity profiles. This clustering approach enables a nuanced understanding of comorbidity interactions, providing a basis for more personalized clinical decision-making and treatment interventions.