Machine learning for synchronous bone metastasis risk prediction in high grade lung neuroendocrine carcinoma.

Bone metastasis (BM) is common in high-grade lung neuroendocrine tumors (NETs). This study aimed to use multiple machine learning algorithms to exploring the significant factors associated with synchronous BM in these patients. Patients diagnosed with high-grade lung NETs were extracted from SEER 17 registries. Age-standardized incidence rate (ASIR) was calculated. All patients were randomly divided into the training cohort and validation cohort (8:2). Eight machine learning algorithms were used to construct predictive model for synchronous BM in the training cohort, and the optimal model was selected for further validation. Shapley Additive Explanations (SHAP) were used to interpret the importance of each variable in the optimal model. In addition, Kaplan-Meier (KM) survival analysis was performed to evaluate survival in patients with synchronous BM. From 2010 to 2021, the ASIR of synchronous BM in small cell lung cancer (SCLC) showed decreasing incidence (from 1.52 to 1.16 per 100,000 person-years, APC - 2.3, 95% CI - 3.3 to - 1.3, P < 0.001). No significant change was found for large cell neuroendocrine carcinoma (LCNEC). Approximately 23% of patients had synchronous BM at diagnosis. The stochastic gradient boosting (GBM) model, developed using ten-fold cross-validation, showed optimal predictive value both in the training and validation cohorts. The SHAP analysis indicated that liver metastasis had the most significant impact on synchronous BM. The median cancer-specific survival of patents with bone metastasis was 8 months. No survival difference was found between LCNEC and SCLC. The incidence of SCLC with synchronous BM showed a slight but statistically significant decrease over the last decade. These patients experienced poor survival. The selected GBM model could help identify patients at high risk of BM among those with high-grade lung NETs.