Study on machine learning-based mortality prediction models for hypertensive heart failure

WANG Miao; FAN Hong-xuan; WANG Lei-gang; REN Zhao-yu; LIANG Bin

doi:10.3969/j.issn.1672-5301.2025.09.004

Abstract：Objective To explore the risk factors of increased mortality in patients with hypertensive heart failure(HHF)and to develop and verify the clinical significance of the prediction models.Methods A total of 4270 patients with HHF were included from The Medical Information Mart For Intensive Care-Ⅳ(MIMIC-Ⅳ)database.The patients were analyzed for mortality at the following time points:during hospitalization,at 7,14 and 30 days after discharge.The dataset was randomly divided into two parts,with 70%of the data used for model training and 30%for validation.Machine learning algorithms were used to identify the independent risk factors,and multiple Logistic regression analysis was used to establish a predictive model in the training set.The models were then validated in the test set and evaluated for its discrimination,calibration and clinical utility.The best performing models were selected to construct a column chart.The EICU Collaborative Research Database was used for the external validation.Results Among 4270 patients,the mortality rates were as follows:14.6%in-hospital,17.1%at 7 days,18.6%at 14 days,and 21.1%at 30 days.After screening,the final included risk factors were as follows:the white blood cell count,the platelet count,the phosphate,the bicarbonate and the anion gap,the acute physiology score Ⅲ(APSⅢ),the Oxford acute severity of illness score(OASIS),the simplified acute physiology score Ⅱ(SAPSⅡ),the charlson comorbidity index(CCI)and body weight.The ROC values of the four models were 0.817,0.810,0.785 and 0.796 in the training set,while were 0.828,0.819,0.793 and 0.801 in the testing set.The C-index values were 0.817,0.810,0.785 and 0.796 in the training set,while were 0.828,0.819,0.793 and 0.801 in the testing set.The models demonstrated strong and consistent performance in the training and testing datasets.Furthermore,the decision curve analysis showed that applying the model provided a net benefit over the strategies of treating all patients or treating no patients.In the external validation cohort(EICU database),the model for in-hospital mortality achieved a C-index of 0.754,demonstrating good discriminative ability.Moreover,decision curve analysis confirmed the model's clinical utility by showing a net benefit for the model-based strategy across a range of threshold probabilities when compared to the all-treated and treat-none strategies.Conclusions The prediction models can effectively predict the in-hospital,7-day,14-day,and 30-day mortality of hypertensive heart failure patients.They are helpful for clinical decision-making.Compared with established risk scores such as the Get with the Guidelines-Heart Failure(GWTG-HF)and the Charlson Comorbidity Index(CCI),our model demonstrates superior performance in predicting in-hospital mortality.It also shows comparative advantages over general severity-of-illness scores like OASIS,APSⅢ,and SAPSⅢ.