Disease classification via interpretable machine learning based on multi-center routine coagulation test

Feng Dong¹Yaqiong Zhang²Weibu Chen³Changmin Wang⁴Lei Zhang⁵Xiaoling Gao⁶Xiaoli Zhang⁷Minghua Jiang⁸Guobin Xu⁹Ruichuang Yang¹⁰Yutong Hou^11*Jiandang Ma¹²Zhuanbao Li¹³Jun Wu¹

• ¹Beijing Jishuitan Hospital Affiliated to Capital Medical University, Beijing, China
• ²Taizhou Central Hospital, Taizhou, China
• ³Shenzhen People's Hospital, Shenzhen, China
• ⁴People's Hospital of Xinjiang Uygur Autonomous Region, Urumqi, China
• ⁵Second Affiliated Hospital of Xi'an Jiaotong University, Xi'an, China
• ⁶Hainan General Hospital, Haikou, China
• ⁷The Affiliated Yongchuan Hospital of Chongqing Medical University, Chongqing, China
• ⁸The Second Affiliated Hospital and Yuying Children's Hospital of Wenzhou Medical University, Wenzhou, China
• ⁹Beijing Cancer Hospital, Beijing, China
• ¹⁰Shenzhen Mindray Bio-Medical Electronics Co Ltd, Shenzhen, China
• ¹¹Beijing Jiaotong University, Beijing, China
• ¹²Luoyang Central Hospital Affiliated to Zhengzhou University, Luoyang, China
• ¹³Beijing Hospital, Beijing, China

Background: This study aims to establish an interpretable disease classification model via machine learning and identify key features related to the disease to assist clinical disease diagnosis based on a multi-center CX9000 routine coagulation test. Methods: Data from 11 hospitals were collected. An unsupervised clustering model was used to extract classification patterns, and clinical experts assigned disease labels. Multiple machine learning models, including Random Forest, SVM, Decision Tree, Naive Bayes, MLP, XGBoost, and LightGBM, were trained. Ten-fold cross validation and external validation were performed. For external validation, models were trained with data from 8 hospitals (˜90%) and tested on the remaining 2 hospitals (˜10%). SHAP and Decision Tree analysis were used for interpretability. Results: Clear clustering patterns were observed for valvular heart disease (VHD) and pulmonary infection (PI). LightGBM achieved the best performance in both tasks. In cross validation, the mean F1-scores were 0.8890 and 0.7233, and the mean AUCs were 0.9500 and 0.8023. External validation showed strong generalization, with mean F1-scores of 0.9259 and 0.7464 and mean AUCs of 0.9493 and 0.8297. The sample visualization by t-SNE and the interpretable analysis by SHAP and Decision Trees identified some key classification features, i.e., international normalized ratio (INR) for VHD and age for PI. Conclusion: Machine learning models based on multi-center coagulation tests provide effective and interpretable disease classification, supporting clinical diagnostic automation.