Machine-learning-model based tool for screening bone metastases from lung cancer patients in primary care practice.

PURPOSE: This study aimed to use the machine-learning methods to predict bone metastasis (BM) in patients with lung cancer.

METHODS

This study included 8,612 patients with lung cancer, from whom we collected baseline characteristics and hematological data. BM was diagnosed based on imaging results. We normalized metric data using Z-scores and divided the patients into training and validation sets in a 7:3 ratio. After under-sampling of the training set to relieve the interlocking imbalance, the key variables were then selected using the LASSO method. Nine distinct machine-learning models were developed on the training set, and their performance was evaluated on the validation set. Finally, we created an online tool to support real-time analysis of BM in lung cancer patients.

RESULTS

Among 8,612 lung cancer patients, 947 (11%) were diagnosed with BM. Nine key indicators associated with BM were identified, including ALP, LYM%, HCT, FBG, TT, TBIL, smoking status, DBIL, and D-dimer. Among the evaluated models, the eXtreme Gradient Boosting (XGBoost) model showed consistently strong performance, achieving AUCs of 0.8423 and 0.7871; accuracies of 0.7655 and 0.7309; balanced accuracies of 0.76645 and 0.72385; specificities of 0.7652 and 0.7329; and sensitivities of 0.7677 and 0.7148 in the training and validation sets, respectively, along with higher Youden index and F1 score. Consequently, we identified XGBOOST as the optimal predictive model and developed an online platform based on this model to facilitate prediction of BM in lung cancer patients.

CONCLUSION

A machine learning model was developed to diagnose BM in lung cancer based on hematological indicators, along with a real-time website. This tool cloud guide further diagnostic evaluations and intervention for high-risk patients.