XLXuemin Liu

Xuemin Liu

*This is an algorithmically generated author record based on publicly available information.

Xuemin Liu has a total of 17 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Integrating Multimodal Information and Advanced Machine Learning Algorithms for Function Assessment and Rehabilitation Prediction in Patients after Cerebral Hemorrhage

    by Xuemin Liu, Yu He, Ziliang Wang, Mengdi Huang, Zihui Liang, Shenglin Xia, Xiangnan Yuan, Fenghua Zhou, Xueyong Liu

    Background The recovery process for patients post-cerebral hemorrhage is influenced by various factors. Crucially, multimodal information, including results from rehabilitation function assessments, imaging examinations, and laboratory tests, is essential for evaluating motor function and predicting the rehabilitation outcome in these patients. Methods A retrospective analysis was conducted, utilizing data from 315 and 424 patients with post-cerebral hemorrhage. The training set comprised rehabilitation function assessment results, imaging examination data, and laboratory test results, used to establish models for evaluating motor function and predicting rehabilitation outcomes. Clinical characteristics of patients underwent single-factor and multi-factor logistic regression analyses, exploring influencing factors during the recovery process after cerebral hemorrhage. Head CT scans of patients underwent pre-processing, extracting radiomic features for predicting motor function when combined with functional assessment results. Moreover, a GBDT gradient boosting tree model was constructed based on patients' multimodal clinical information and compared with other machine learning models to evaluate rehabilitation outcomes. Separate models for predicting motor function and evaluating rehabilitation were established for patients with cerebral hemorrhage using selected radiomic and clinical features, and the performance of each model was individually assessed. Results Following extensive training and validation with a substantial patient records dataset, the receiver operating characteristic (ROC) analysis reveals the excellent performance of the motor function prediction model we developed for patients post-cerebral hemorrhage. In both five-fold and ten-fold cross-validation, the gradient boosting decision tree (GBDT) algorithm demonstrates superior performance in the post-cerebral hemorrhage rehabilitation assessment compared to other machine learning algorithms. Multifactor analysis indicates that, in predicting motor function after cerebral hemorrhage, factors such as albumin, neutrophil count and percentage, triglycerides, coagulation time, and urea are significant. In rehabilitation assessment, meaningful factors include the age of onset, admission modified Barthel Index (mBI) score, and the initiation of rehabilitation. Conclusion Our study integrates multimodal information and advanced machine learning algorithms, providing a solution for assessing function and predicting rehabilitation in patients after cerebral hemorrhage. The established models hold the potential to provide decision support for clinicians in clinical practice, promoting the realization of personalized rehabilitation treatment.


  • Open Access

    A Novel Deep Learning Model for Automated Mapping of Laboratory Test Terminologies to Logical Observation Identifiers Names and Codes (LOINC) (Preprint)

    by Di-Ge Ai, Yu He, Sheng-Hai Jin, Xue-Min Liu, Nian-Yi Sun, Guang-Ku Tian, Zhi-Qiang Zhang, Xue-Yong Liu

    BACKGROUND A standardized method for identifying medical laboratory observations, such as Logical Observation Identifiers Names and Codes (LOINC), is critical for creating accurate and effective public data models. However, such standards are not being used effectively. Standardized mapping facilitates consistency in medical terminologies and data sharing in multicenter treatment. OBJECTIVE To address the problem of standardizing laboratory test terminologies, a deep learning–based high-precision end-to-end terminology standardization matching system was developed to map laboratory test terms (LTTs) to LOINC. METHODS We manually constructed a laboratory test terminology mapping dataset containing 15,349 data items extracted from the information system of the Shengjing Hospital of China Medical University and matched 2,375 LOINC. We developed Attribute-wised Graph Attention Siamese Network (AGASN), a deep learning–based high-precision laboratory test terminology mapping model, to separately extract LTT features and LOINC term features and calculate the matching rate. We designed an attribute pooling mechanism to convert terminology strings to attribute sequences. Moreover, we developed a graph attention model based on attribute relations, which increased the interpretability of the proposed model. The problem of inconsistency in training and testing objectives was solved by improving the training objectives of the model. RESULTS The proposed a novel deep learning model achieved an accuracy of 82.33% ± 0.6% on the test dataset where the LOINC were visible, corresponding to a 10.9% improvement compared with that obtained using a random forest classifier. Furthermore, the proposed system achieved an accuracy of 63.14% ± 0.2% on the test dataset where the LOINC were invisible, constituting a 10.0% improvement compared with that obtained using SimCSE. Manual validation of the system performance showed accuracies of 82.33% and 70.66% on labeled and unlabeled datasets, respectively. Finally, we constructed a visual attribute relational strength network using an attribute graph attention model. CONCLUSIONS Herein, a Chinese laboratory test terminology mapping dataset was created and a deep learning system for the standardized mapping of LTTs was proposed. The results demonstrate that the proposed system can map LTTs to LOINC with high accuracy.