Development and validation of machine learning models to predict perioperative transfusion risk for hip fractures in the elderly.

BACKGROUND: Patients with hip fractures frequently need to receive perioperative transfusions of concentrated red blood cells due to preoperative anemia or surgical blood loss. However, the use of perioperative blood products increases the risk of adverse events, and the shortage of blood products is prompting us to minimize blood transfusion. Our study aimed to construct a machine learning algorithm predictive model to identify patients at high risk for perioperative transfusion early in hospital admission and to manage their patient blood to reduce transfusion requirements. METHODS: This study collected patients hospitalized for hip fractures at a university hospital from May 2016 to November 2022. All patients included in the analysis were randomly divided into a training set and validation set according to 70:30. Eight machine learning algorithms, CART, GBM, KNN, LR, NNet, RF, SVM, and XGBoost, were used to construct the prediction models. The models were evaluated for discrimination, calibration, and clinical utility, and the best prediction model was selected. RESULTS: A total of 805 patients were included in the study, of whom 306 received transfusions during the perioperative period. We screened eight features used to construct the prediction model: age, fracture time, fracture type, hemoglobin, albumin, creatinine, calcium ion, and activated partial thromboplastin time. After evaluating and comparing the performance of each of the eight models, the model constructed by the XGBoost algorithm had the best performance, with MCC values of 0.828 and 0.939 in the training and validation sets, respectively. In addition, it had good calibration and clinical utility in both the training and validation sets. CONCLUSION: The model constructed by the XGBoost algorithm has the best performance, using this model to identify patients at high risk for transfusion early in their admission and promptly incorporating them into a patient blood management plan can help reduce the risk of transfusion.

Machine learning-based prediction of vitamin D deficiency: NHANES 2001-2018.

Background: Vitamin D deficiency is strongly associated with the development of several diseases. In the current context of a global pandemic of vitamin D deficiency, it is critical to identify people at high risk of vitamin D deficiency. There are no prediction tools for predicting the risk of vitamin D deficiency in the general community population, and this study aims to use machine learning to predict the risk of vitamin D deficiency using data that can be obtained through simple interviews in the community. Methods: The National Health and Nutrition Examination Survey 2001-2018 dataset is used for the analysis which is randomly divided into training and validation sets in the ratio of 70:30. GBM, LR, NNet, RF, SVM, XGBoost methods are used to construct the models and their performance is evaluated. The best performed model was interpreted using the SHAP value and further development of the online web calculator. Results: There were 62,919 participants enrolled in the study, and all participants included in the study were 2 years old and above, of which 20,204 (32.1%) participants had vitamin D deficiency. The models constructed by each method were evaluated using AUC as the primary evaluation statistic and ACC, PPV, NPV, SEN, SPE, F1 score, MCC, Kappa, and Brier score as secondary evaluation statistics. Finally, the XGBoost-based model has the best and near-perfect performance. The summary plot of SHAP values shows that the top three important features for this model are race, age, and BMI. An online web calculator based on this model can easily and quickly predict the risk of vitamin D deficiency. Conclusion: In this study, the XGBoost-based prediction tool performs flawlessly and is highly accurate in predicting the risk of vitamin D deficiency in community populations.

Machine learning algorithms to predict risk of postoperative pneumonia in elderly with hip fracture.

BACKGROUND: Hip fracture (HF) is one of the most common fractures in the elderly and is significantly associated with high mortality and unfavorable prognosis. Postoperative pneumonia (POP), the most common postoperative complication of HF, can seriously affect patient prognosis and increase the burden on the healthcare system. The aim of this study was to develop machine learning models for identifying elderly patients at high risk of pneumonia after hip fracture surgery. METHODS: From May 2016 to November 2022, patients admitted to a single central hospital for HF served as the study population. We extracted data that could be collected within 24 h of patient admission. The dataset was divided into training and validation sets according to 70:30. Based on the screened risk factors, prediction models were developed using seven machine learning algorithms, namely CART, GBM, KNN, LR, NNet, RF, and XGBoost, and their performance was evaluated. RESULTS: Eight hundred five patients were finally included in the analysis and 75 (9.3%) patients suffered from POP. Age, CI, COPD, WBC, HB, GLU, STB, GLOB, Ka+ which are used as features to build machine learning models. By evaluating the model's AUC value, accuracy, sensitivity, specificity, Kappa value, MCC value, Brier score value, calibration curve, and DCA curve, the model constructed by XGBoost algorithm has the best and near-perfect performance. CONCLUSION: The machine learning model we created is ideal for detecting elderly patients at high risk of POP after HF at an early stage.

Type 2 diabetes and the risk of synovitis-tenosynovitis: a two-sample Mendelian randomization study.

Introduction: It has been shown that people with type 2 diabetes have a higher risk of synovitis and tenosynovitis, but previous studies were mainly observational, which may be biased and does not allow for a cause-and-effect relationship. Therefore, we conducted a two-sample Mendelian randomization (MR) study to investigate the causal relationship. Method: We obtained data on "type 2 diabetes" and "synovitis, tenosynovitis" from published large-scale genome-wide association studies (GWAS). The data were obtained from the FinnGen consortium and UK Biobank, both from European population samples. We used three methods to perform a two-sample MR analysis and also performed sensitivity analysis. Results: The results of all three MR methods we used for the analysis illustrated that T2DM increases the risk factor for the development of synovitis and tenosynovitis. Specifically, for the IVW method as the primary analysis outcome, OR = 1.0015 (95% CI, 1.0005 to 1.0026), P = 0.0047; for the MR Egger method as the supplementary analysis outcome, OR = 1.0032 (95% CI, 1.0007 to 1.0056), P = 0.0161; for the weighted median method, OR = 1.0022 (95% CI, 1.0008 to 1.0037), p = 0.0018. In addition, the results of our sensitivity analysis suggest the absence of heterogeneity and pleiotropy in our MR analysis. Conclusion: In conclusion, the results of our MR analysis suggest that T2DM is an independent risk factor for increased synovitis and tenosynovitis.