Cameron Noorbakhsh: No financial relationships to disclose
Introduction: Identifying patient risk factors for adverse surgical outcomes is vital for improving preoperative counseling and operative planning for lumbar fusions. We apply unsupervised machine learning to cluster patients on a comprehensive dataset based on demographics and comorbidities. This approach overcomes the limitations and bias inherent to manual analysis, aiding in more accurate risk stratification and surgical planning.
Methods: We clustered on patient comorbidities and demographic ranges (age above/below 65, BMI above/below 30) on a sample of 413 patients (187 age above 65, 227 BMI above 30). Comorbidities were grouped by ICD-10 categories to reduce the dimensionality of the data. We used k-modes clustering for its strength with categorical data. Optimal number of clusters was determined by silhouette scores. We compared clusters for 90-day post-operative outcomes using a chi-squared test.
Results: We identified 5 clusters based on silhouette scores, which was 0.539 for 5 clusters. The groups were distinguished as group 1 (153 patients) - older (> 65 yo) with low BMI ( < 30), group 2 (55 patients) - young with high BMI, group 4 (37 patients) - old and young with low BMI. Group 0 (132 patients) and group 3 (36 patients) had similar demographics but differed by whether they had a neoplasm (group 3). Performing regressions on statistically significant complications from the chi-squared test (p < 0.05), we noted group 2 was associated with SSI and wound dehiscence. Group 1 was associated with post-hemorrhagic anemia and anemia of chronic disease. Group 3 was negatively associated with UTIs.
Conclusion : Unsupervised clustering may aid in identifying patient characteristics that would help surgeons determine risk-benefit for surgeries. Clustering patients identifies characteristics that are reflective of the patient populations undergoing lumbar fusion surgeries and clarifies which groups are most at risk for complications.
.jpg)