Interpretable Machine Learning Methods To Predict In-hospital Mortality In Patients With Spinal Fractures In Intensive Care Units: Model Construction And Validation

Objective:There are currently very few criteria for evaluating the severity of patients with spinal fractures in the intensive care unit.To our knowledge,no one has conducted a study on predicting in-hospital mortality in these patients besides us.As such,our current study aims to assess patient severity using clinical indicators within 24 hours of admission and predict the probability of in-hospital mortality.This will allow us to identify high-risk patients early on and guide clinical practice accordingly.Methods:Our patients are sourced from two large critical care medicine databases,the MIMIC-III(Medical Information Mart for Intensive Care version III)database and the e RI(e ICU Research Institute Database)database.We identified patients diagnosed with spinal fractures by disease diagnosis codes and removed duplicate admissions.Data were collected in four major areas: patient demographic characteristics,laboratory tests,co-morbidities,and various scoring systems.Statistical differences in each index were compared between the deceased and surviving groups,as well as regularization methods for dimensionality reduction,and finally meaningful clinical characteristics were filtered by clinical experience for modeling.We used the MIMIC-III database as the training set(internal validation set)and the e RI database as the validation set(external validation set),and we developed and validated six popular machine learning algorithms,namely: k-neighborhood algorithm,random forest,extreme gradient boosting,logistic regression,and neural network.Based on the internal and external validation results of the models,we performed data visualization and filtered the best models.Considering the specificity of the healthcare industry,we also performed model interpretability analysis.Results:A total of 1512 patients were finally included in this study,of which 1190 patients were included in the e RI dataset,117 patients experienced in-hospital death,with a mortality rate of 9.83%;322 patients were included in the MIMIC dataset.25 patients died,with a mortality rate of 7.76%.After screening by the above methods,we finally included a total of 20 clinical variables into the model training.The results of model validation showed that,collectively,the random forest algorithm performed relatively consistently and well for both internal and external validation,with AUC and F1 scores of 0.77 and 0.32 for internal validation and 0.73 and 0.25 for external validation,respectively.in the interpretability analysis,the three most important variables of the model were: body temperature,mean arterial pressure,and age,and other interpretable analyses demonstrated that the model was judged in a manner consistent with our previous medical knowledge.Conclusion:Our study successfully developed and validated several machine learning models for predicting in-hospital mortality in patients with spinal fractures in intensive care units,with the random forest model performing the best and the interpretability analysis of the model being in line with medical knowledge,while external validation confirmed the strong migration and generalization capabilities of our current model.In conclusion,our model can accurately identify high-risk patients early and take active and effective actions in advance that are expected to reduce mortality in admitted patients;at the same time,there are some limitations of our study,especially that this study is a retrospective study and more prospective studies are needed to validate our results afterwards.