Clinical Problem High-grade lung neuroendocrine carcinomas (HGLNEC), comprising small cell lung cancer (SCLC) and large cell neuroendocrine carcinoma (LCNEC), are aggressive tumors with a strong propensity for early systemic dissemination. Bone metastasis occurs in approximately 23% of newly diagnosed HGLNEC patients and is associated with pain, fracture, spinal cord compression, and worse survival.
Importance of Early Detection Identifying patients at high risk for synchronous bone metastasis at diagnosis enables more appropriate staging workup, including bone scintigraphy or PET-CT, and informs prognosis counseling and treatment planning. A reliable machine learning risk model could help clinicians stratify patients who most need intensive staging.
SEER Database Study The authors used the Surveillance, Epidemiology, and End Results (SEER) database, which contains data on 21,809 HGLNEC patients diagnosed between 2010 and 2021. This large population-based dataset enables development and validation of machine learning models with sufficient sample sizes for reliable performance estimation.
Eight ML Algorithms Eight different machine learning algorithms were trained and compared for bone metastasis prediction, including gradient boosting machines (GBM), random forests, logistic regression, support vector machines, and neural networks. The best-performing model was the GBM, which achieved an AUC of 0.723 in validation.
Patient Selection HGLNEC patients (ICD-O histology codes 8041-8045 for SCLC and 8013 for LCNEC) diagnosed between 2010 and 2021 were extracted from the SEER database. After exclusions for incomplete staging data, the analytic cohort contained 21,809 patients, of whom 23.2% had synchronous bone metastasis at diagnosis.
Features Used The models used readily available clinical and pathological variables: age, sex, race, histology (SCLC vs. LCNEC), tumor size, tumor laterality, T and N stage, and concurrent metastases to other sites (brain, liver, lung). Importantly, all features are available from routine clinical workup without requiring additional molecular testing.
Data Splitting The dataset was split into training (70%) and validation (30%) sets using stratified random sampling to maintain proportional class representation. Given the class imbalance (23% bone metastasis positive), synthetic minority over-sampling (SMOTE) was applied to the training set to improve model performance on the minority class.
Outcome Definition Synchronous bone metastasis was defined as bone metastasis documented at the time of initial diagnosis, based on SEER metastasis site coding. This represents the clinically relevant scenario where identifying high-risk patients could influence staging workup decisions.
GBM Model Performance The gradient boosting machine (GBM) achieved the highest AUC of 0.723 in the validation set, with a sensitivity of 0.68, specificity of 0.69, and balanced accuracy of 0.68. While not highly precise, this performance represents meaningful discrimination above chance and provides clinically actionable risk stratification.
Comparison to Other Models Random forest (AUC 0.706), logistic regression (AUC 0.698), and XGBoost (AUC 0.715) showed similar performance to GBM. Decision tree and naive Bayes models performed worst (AUC 0.63-0.65), suggesting that ensemble methods are superior for this prediction task with moderate sample sizes and mixed feature types.
SHAP Importance Analysis SHapley Additive exPlanations (SHAP) identified the most influential predictors: liver metastasis (most important), brain metastasis, lung metastasis, T stage, N stage, tumor size, and age. The finding that presence of other synchronous metastases predicts bone metastasis reflects the systemic dissemination phenotype of HGLNEC.
Liver Metastasis as Top Predictor Liver metastasis was the single most important SHAP feature. Patients with synchronous liver metastasis had a 3-fold higher rate of bone metastasis compared to those without, consistent with the clinical observation that HGLNEC metastatic dissemination patterns tend to be widespread when they occur.
Overall Survival in HGLNEC The entire HGLNEC cohort had a median cancer-specific survival (CSS) of 8 months, reflecting the aggressive biology of this tumor type. Even patients without bone metastasis had limited survival, with median CSS of 10 months, while bone metastasis patients had median CSS of 5 months.
Bone Metastasis Survival Impact Patients with synchronous bone metastasis had significantly worse CSS compared to those without (5 vs. 10 months, p<0.001). Multivariate Cox regression confirmed bone metastasis as an independent predictor of shorter CSS after adjusting for primary tumor T/N stage, histology, age, and other metastatic sites.
SCLC vs. LCNEC Differences SCLC had higher bone metastasis rates (25.1%) compared to LCNEC (17.8%), consistent with SCLC's well-known propensity for early and widespread metastasis. However, survival after bone metastasis was similarly poor in both subtypes, suggesting that while patterns of dissemination differ, the prognostic impact is equivalent.
Impact on Treatment Planning The short median survival and frequent co-occurrence of multiple metastatic sites highlights that comprehensive staging including bone evaluation should be standard for all HGLNEC patients at diagnosis. The ML model could help identify the highest-risk patients for prioritized bone imaging when resources are limited.
SHAP Framework SHAP values provide game-theoretic explanations of individual model predictions, showing how each feature contributes positively or negatively to a given patient's predicted bone metastasis probability. This transforms the GBM from a black-box model into an interpretable clinical tool.
Interaction Effects SHAP interaction analysis revealed that the effect of liver metastasis on bone metastasis risk was amplified in the presence of brain metastasis. Patients with both liver and brain metastases had predicted bone metastasis probabilities exceeding 50%, substantially higher than the 23% population average.
Clinical Nomogram The authors also developed a traditional logistic regression-based nomogram using the top SHAP-identified features. While the nomogram had slightly lower AUC (0.710) than GBM, it provides a visual, interpretable scoring tool that clinicians can use at the bedside without requiring specialized software.
Calibration Analysis Calibration curves comparing predicted probabilities to observed bone metastasis rates showed good calibration for both the GBM and nomogram across the probability range. Hosmer-Lemeshow goodness-of-fit tests confirmed adequate calibration, supporting clinical reliability of the predicted risk scores.
SEER Data Limitations SEER lacks molecular data (EGFR, ALK, RB1, TP53 mutations) that may influence metastatic patterns in HGLNEC. Treatment information is also limited - SEER does not capture chemotherapy regimens or response data - preventing analysis of how treatment affects metastasis development.
External Validation Needed The model was developed and validated on SEER data, which has selection biases related to geographic and socioeconomic factors. External validation on independent institutional cohorts with more complete clinical and molecular data is needed before clinical deployment.
Molecular Biomarker Integration Future models should integrate tumor molecular profiles. SCLC subtype (ASCL1, NEUROD1, POU2F3, YAP1) and specific mutations (RB1 loss, MYC amplification) are associated with metastatic propensity in SCLC and could significantly improve prediction accuracy beyond clinical features alone.
Prospective Impact Assessment Even with adequate predictive accuracy, demonstrating that ML-guided staging decisions actually improve patient outcomes requires prospective clinical trials or real-world evidence studies. A pragmatic trial where ML risk stratification guides bone staging decisions could quantify the clinical impact of the model.