2024, Vol. 28 ›› Issue (33): 5295-5301
Machine learning to analyze risk factors for postoperative failure of proximal humeral fractures with medial column instability
Xu Daxing1, 2, Ji Muqiang2, Tu Zesong2, 3, Xu Weipeng3, Xu Weilong4, Niu Wei5
1Applicants with Equivalent Academic Qualifications for Doctoral Degrees, Guangzhou University of Chinese Medicine, Guangzhou 510006, Guangdong Province, China; 2Department of Orthopedics, Sanshui Branch of Foshan Hospital of Traditional Chinese Medicine, Foshan 528100, Guangdong Province, China; 3Department of Orthopedics, Foshan Hospital of Traditional Chinese Medicine, Foshan 528000, Guangdong Province, China; 4Department of Orthopedics, Shenzhen Pingle Orthopedics Hospital, Shenzhen 518122, Guangdong Province, China; 5Department of Articular Surgery, Guangdong Provincial Hospital of Chinese Medicine, Guangzhou 510120, Guangdong Province, China
Abstract: BACKGROUND: Internal fixation and open reduction with locking plate is the main treatment for proximal humeral fractures with medial column instability. However, reduction failure is one of the main postoperative complications, and accurate risk factor assessment is beneficial for screening high-risk patients and clinical decision selection.
OBJECTIVE: To construct four types of prediction models by different machine learning algorithms, compare the optimal model to analyze and sort the risk variables according to their weight scores on the impact of outcome, and explore their significance in guiding clinical diagnosis and treatment.
METHODS: 262 patients with proximal humeral fractures with medial column instability, aged (60.6±10.2) years, admitted to Foshan Hospital of Traditional Chinese Medicine between June 2012 and June 2022 were included. All patients underwent open reduction with locking plate surgery. According to the occurrence of reduction failure at 5-month follow-up, the patients were divided into a reduction failure group (n=64) and a reduction maintenance group (n=198). Clinical data of patients were collected, and model variables and their classification were determined. The data set was randomly divided into a training set and a test set according to a 7:3 ratio, and the optimal hyperparameters were obtained in the training set according to a 5-fold cross-over test. Four machine learning prediction models of logistic regression, random forest, support vector machine, and XGBoost were constructed, and the performance of different algorithms was observed in the test set using AUC, correctness, sensitivity, specificity, and F1 scores, so as to comprehensively evaluate the prediction performance of the models. The best-performing model was evaluated using SHAP to assess important risk variables and to evaluate its clinical guidance implications.
RESULTS AND CONCLUSION: (1) There were significant differences between the two groups in deltoid tuberosity index, fracture type, fracture end with varus deformity before operation, fragment length of inferior metaphyseal of humerus, postoperative reduction, cortical support of medial column of proximal humerus, and insertion of calcar screw (P < 0.05). (2) The best-combined performance of the four machine models was XGBoost. The AUC, accuracy, and F1 scores were 0.885, 0.885, and 0.743, respectively; followed by random forest and support vector machine, with both models performing at approximately equal levels. Logistic regression had the worst combined performance. The SHAP interpretation tool was used in the optimal model and results showed that deltoid tuberosity index, medial humeral column cortical support, fracture type, fracture reduction quality, and the status of the calcar screw were important influencing fators for postoperative fracture reduction failure. (3) The accuracy of using machine learning to analyze clinical problems is superior to that of traditional logistic regression analysis methods. When dealing with high-dimensional data, the machine learning approach can solve multivariate interaction and covariance problems well. The SHAP interpretation tool can not only clarify the importance of individual variables but also obtain detailed information on the impact of dummy variables in each variable on the outcome.
Key words: proximal humeral fracture, medial column instability, machine learning, influencing factor, SHAP interpretation tool