Machine learning clustering of carotid body tumor patients by clinical features and surgical approaches.

BACKGROUND

To classify CBT patients using machine learning-based clustering analysis to identify subgroups with different risks of complications and outcomes.

METHODS

This retrospective, dual-center study included 127 patients who underwent CBT resection between 2003 and 2019. Clinical data were collected. Clustering analysis was performed using the Gower distance and silhouette coefficient to determine the optimal number of clusters. Clinical characteristics and outcomes were compared between phenotypes.

RESULTS

Phenotype 1 (n = 29) consisted of patients who received preoperative embolization therapy, with a Shamblin III classification in 55.2% of cases and the shortest operative time (110 min). Phenotype 2 (n = 46) included patients with high-grade Shamblin classification without embolization therapy, characterized by longer operative times (240 min) and higher rates of vascular reconstruction (58.7%). Phenotype 3 (n = 52) comprised mainly female patients (80.8%) with Shamblin I tumors (51.9%) and the smallest tumor diameter (3 mm). Phenotype 2 had significantly higher estimated blood loss, postoperative stay, and cranial nerve injury rates compared to other phenotypes.

CONCLUSIONS

Patients in Phenotype 2, with large tumors and high Shamblin classifications, had higher complication rates. Preoperative embolization combined with surgical resection was associated with lower surgical risks in these high-risk patients. This approach may provide a more precise method for identifying high-risk patients and could guide clinical decision-making.