Journal of Forensic Medicine ›› 2026, Vol. 42 ›› Issue (1): 8-16.DOI: 10.12116/j.issn.1004-5619.2024.241108

• Original Articles • Previous Articles     Next Articles

Machine Learning Model Based on Structured Injury Features for Knee Mobility Discriminations after Traumatic Injury

Run-ting DOU1(), Shun CHENG2(), Xin ZHOU2, Xing YE1, Zhi-min WANG1, Guang-hui HONG1, Qi ZHANG2, Qing XIA3(), Yi-wen SHEN1()   

  1. 1.Department of Forensic Medicine, School of Basic Medical Sciences, Fudan University, Shanghai 200032, China
    2.The SMART (Smart Medicine and AI-based Radiology Technology) Lab, Shanghai Institute for Advanced Communication and Data Science, Shanghai University, Shanghai 200444, China
    3.Shanghai Key Laboratory of Forensic Medicine, Key Laboratory of Forensic Science, Ministry of Justice, Shanghai Forensic Service Platform, Academy of Forensic Science, Shanghai 200063, China
  • Received:2024-11-29 Online:2026-04-28 Published:2026-02-25
  • Contact: Qing XIA, Yi-wen SHEN

Abstract:

Objective To extract structured injury features of knee trauma from forensic case files, and to calculate the degree of joint mobility limitation based on the injury features using a machine learning model combined with the voting method. Methods A total of 490 forensic cases involving knee trauma leading to motor dysfunction were retrospectively collected and randomly divided into training and testing sets at an 8:2 ratio. Structured injury features were extracted and systematically organized and stored using a MySQL database. Six machine learning models, including support vector classification, random forest, logistic regression, gradient boosting, k- nearest neighbors, and extreme gradient boosting, were applied to select the optimal models. Using a 25% loss of joint range of motion as the threshold, a model for classifying the severity of knee functional impairment was established by combining the selected models with a voting method. The best models were first selected based on their average AUC values, and further validated using 5-fold cross-validation. The SHAP method was used to analyze and interpret the prediction results of the optimal model. In addition, 57 cases were collected as an external validation to evaluate the model's generalization ability. Results The average AUC values for support vector machine, random forest, and extreme gradient boosting all exceeded 0.9. In 5-fold cross-validation, each of the three individual models achieved an average AUC value of 0.89. After integrating these three models using the voting method, the average AUC of 5-fold cross-validation increased to 0.91. Feature contribution plots and decision plots generated based on Shapley values clearly illustrated the model's performance, and the evaluation metrics on the external validation set were comparable to those from internal validation. Conclusion The developed machine learning model based on structured injury features demonstrates good performance in classifying the severity of motor dysfunction following knee trauma, with high model interpretability and strong generalization capability.

Key words: forensic clinics, joint range of motion, machine learning, data structuring, knee joint

CLC Number: