Development and application of machine learning models for hematological disease diagnosis using routine laboratory parameters: a user-friendly diagnostic platform

作者：Jingya Liu, Yang Gou, Wuchen Yang, Hao Wang, Jing Zhang, Shengwang Wu, Siheng Liu, Siheng Liu, Tinglu Tao, Yongjie Tang, Cheng Yang, Siyin Chen, Ping Wang, Yimei Feng, Cheng Zhang, Shuiqing Liu, Shuiqing Liu, Xiangui Peng, Xi Zhang · 发表于：Frontiers in Medicine · 年份：2025 · DOI：10.3389/fmed.2025.1605868 · 被引用次数：1 · 研究领域：Digital Imaging for Blood Diseases、Artificial Intelligence in Healthcare、Imbalanced Data Classification Techniques

Aim: In recent years, with the change of social environment, the incidence and detection rate of hematological diseases have shown an increasing trend. Early diagnosis and detection of hematological diseases are very important to improve the quality of life and prognosis of patients. Methods: In this study, we employed 54 clinical and conventional laboratory parameters. By optimally combining multiple feature selection methods and machine learning algorithms, we developed 7 machine learning models with varying feature set sizes. We comprehensively evaluated the performance of these models, analyzed the interpretability of the optimal and simplified models using SHapley Additive exPlanations (SHAP), and compared these two models with the diagnostic performance of hematologists. Finally, we developed a user-friendly diagnostic platform. Results: The results showed that the ensemble model_1 with 46 feature parameters (EnMod1-46) and the simple ensemble model_2 with 12 feature parameters (EnMod2-12) demonstrated significant performance in diagnosing 16 types of hematological diseases. On the temporally distinct test set_1, the EnMod1-46 achieved an accuracy of 0.804 and an area under the curve (AUC) of 0.964, while EnMod2-12 attained an accuracy of 0.784 and an AUC of 0.961. To further validate the model's generalization performance, EnMod1-46 achieved an accuracy of 0.738 and an AUC of 0.973 on the independent external test set_2, while EnMod2-12 yielded an accuracy of 0.705 and...