基于可解释性机器学习的心力衰竭并发急性肾损伤生存预后模型研究OA

Objective This study aimed to develop and validate an interpretable machine learning-based model for predicting the risk of all-cause death during hospitalization in patients with heart failure(HF)complicated with acute kidney injury(AKI).Methods Study data were collected from MIMIC-Ⅳ database,9987 ICU patients were included,and the cases of HF combined with AKI were screened by ICD-9/10.Multiple interpolation was used to process the missing data,combined with Lasso regression and BorutaShap algorithm for feature screening,and finally identified 13 key predictors,including Charson comorbidibility index,acute physiological and chronic health scores,length of stay,etc.The study compared 10 machine learning models(e.g.XGBoost,random Forest,logistic regression,etc.).Results The AUC of random forest,gradient lift and XGBoost is the best(0.78),XGBoost had the highest accuracy rate(71.76%)and the best F1 score(75.00%).The specificity of the decision tree is the most prominent(90.80%).The gradient lift has the best sensitivity(78.03%).SHAP analysis showed that Charson comorbidity index,acute physiological and chronic health scores and respiratory rate were the core factors affecting the risk of death.The study further revealed dynamic differences in the importance of features in specific subpopulations,such as the prominent role of length of stay and bilirubin scores in local prediction.Based on the model results,an online risk assessment platform was developed to provide clinicians with individualized risk probabilities and support early intervention decisions.Conclusion Through interpretable machine learning model,this study provides accurate tools for prognostic management of patients with HF complicated with AKI,and provides auxiliary decision-making for clinical practice.