Related Experiment Video
Updated: Jul 15, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Development and validation of prediction model for intrapulmonary metastasis in lung adenocarcinoma based on machine
Zhenglong Wang1,2,3, Xiaobo Liu2, Mingyue Zhang1
1Department of Thoracic Surgery, Affiliated Hospital of Shandong Second Medical University, Weifang, China.
Background:
Lung adenocarcinoma (LUAD) is a leading cause of cancer-related death, with intrapulmonary metastasis (IPM) delaying diagnosis and worsening prognosis. Despite machine learning (ML)'s promise in metastasis prediction, specific models for LUAD-IPM are lacking. The present study aimed to construct a ML algorithm to accurately predict the risk of IPM in patients with LUAD.
Methods:
We analyzed 61,519 LUAD patients diagnosed between 2018 and 2022 from the Surveillance, Epidemiology, and End Results (SEER) database, including 7,696 with IPM. Thirteen clinicopathological and demographic variables were assessed, covering demographics (age, sex, marital status, race), tumor features [size, primary site, tumor stage (T stage), node stage (N stage), bone/brain/liver/lung metastasis], and treatments (radiotherapy, chemotherapy). After excluding patients with missing key data, univariate and multivariate logistic regression (LR) identified independent prognostic factors (P<0.05). The Synthetic Minority Oversampling Technique (SMOTE) addressed class imbalance, and the balanced dataset was split into training (70%) and testing (30%) sets. Twelve ML models were constructed, with 10-fold cross-validation, calibration curves, and decision curve analysis (DCA) for validation. SHapley Additive exPlanations (SHAP) analysis quantified feature contributions and enhanced model interpretability.
Results:
T stage, N stage, bone/brain/liver metastasis, radiation, and chemotherapy were independent prognostic factors. The light gradient boosting machine (LGBM) model outperformed others, achieving a testing cohort area under the receiver operating characteristic (ROC) curve (AUC) of 0.818, sensitivity 0.782, specificity 0.717, and F1-score 0.421. SHAP analysis confirmed T stage, radiation, and N stage as the top 3 influential features shaping predictions.
Conclusions:
This is the first ML model specifically predicting LUAD-IPM. The LGBM model enables accurate risk stratification, supporting personalized surveillance and treatment optimization to improve clinical outcomes by identifying high-risk patients and avoiding unnecessary interventions.
