Related Experiment Video
Updated: Jun 23, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
A highly interpretable machine learning model for predicting lung cancer bone metastasis: uncovering the synergistic
Zi-Feng Jiang1, Zhang-Yan Ke2, Min Wang3
1Department of Geriatric Respiratory and Critical Care Medicine, The First Affiliated Hospital of Anhui Medical University, 218 Jixi Road, Hefei, 230022, Anhui, Province, China. zifeng0915@fy.ahmu.edu.cn.
Background:
Bone metastasis (BM) significantly impairs lung cancer prognosis and patient quality of life. Conventional imaging modalities often face limitations in early detection and cost-effectiveness. This study aimed to develop and validate an interpretable machine learning (ML) model using routine, cost-effective biochemical markers for the early, non-invasive prediction of BM.
Methods:
This retrospective study included 566 lung cancer patients. Clinicopathological and laboratory features such as alkaline phosphatase (ALP), D-dimer, and lactate dehydrogenase (LDH) were collected. The dataset was partitioned into training and independent test sets. Six ML algorithms were evaluated using cross-validation, with the gradient boosting decision tree (GBDT) identified as the optimal model. Robustness and transparency were rigorously assessed via SHAP analysis, 1000 bootstrap resamples, and multi-dimensional subgroup analyses.
Results:
ALP, D-dimer, and LDH were significantly elevated in BM( +) patients (P < 0.001). In the test set, GBDT (gradient boosting decision tree) achieved an overall AUC of 0.774 (95% CI: 0.721-0.827) and an F1-score of 0.762. After subgroup integration, predictive performance improved to an AUC of 0.811 (95% CI 0.752-0.870), significantly outperforming traditional logistic regression (AUC = 0.755). Peak performance was observed in lung adenocarcinoma (AUC = 0.864). SHAP analysis quantitatively revealed a synergistic, non-linear interaction between ALP and D-dimer as a primary, quantifiable driver of BM risk.
Conclusion:
Our routine-marker-based ML model demonstrates high diagnostic accuracy and robust generalizability. By precisely identifying high-risk populations with high transparency, this cost-effective tool provides scientific decision support for implementing personalized bone scan screening strategies and optimizing resource allocation in clinical practice.