| Part 1:Prognostic value of novel biomarkers in non-ischemic heart failure Section 1:Prognostic value of high-sensitivity cardiac troponin I in non-ischemic heart failureAims:Evidence of the prognostic value of high-sensitivity troponin in patients with non-ischemic heart failure(NIHF)is scarce.This study aimed to assess the distribution and the predictive value of high-sensitivity cardiac troponin I(hs-cTnl)in NIHF.Methods:This study was a retrospective study.Six hundred and fifty NIHF patients admitted to the heart failure care unit(HFCU),Fuwai Hospital,with baseline hs-cTnI results,were consecutively enrolled from December 2006 to October 2017.The primary endpoint was all-cause mortality.A density plot and a boxplot were employed to depict the distribution of hs-cTnI in the whole study population and in NIHF patients with different etiologies,respectively.The Kaplan-Meier survival analysis and the Cox regression analysis were applied to evaluate the association between hs-cTnI and the composite outcome.The concordance index(C-index),the Akaike information criterion(AIC),the Bayesian information criterion(BIC),the Brier score,the integrated discrimination improvement(IDI),and the net reclassification improvement(NRI)were calculated for assessing the increment prognostic value of hs-cTnI based on a well-established model(including age,sex,New York Heart Association class,systolic blood pressure,left ventricular ejection fraction,hemoglobin,sodium,estimated glomerular filtration rate,diabetes mellitus,treatment with angiotensin-converting enzyme inhibitors or angiotensin Ⅱ receptor blockers,treatment with β-blockers,and NT-proBNP)in the NIHF.Decision curve analyses were conducted to evaluate the clinical usefulness of hs-cTnI.Results:During a median follow-up of 1036 days,163 patients reached the primary endpoint.Hs-cTnI was detected in 94.8%of NIHF patients,with 79.7%within the 99th percentile of the upper limit.No significant discrimination of hs-cTnI distribution across different etiologies was observed in the NIHF patients.Hs-cTnI was independently associated with a higher risk of all-cause mortality or heart transplant in NIHF patients[per log2 increase,hazard ratio(HR):1.23,95%confidence interval(CI):1.13-1.33].The cut-off value of hs-cTnI was 0.011 ng/ml.calculated by the maximally selected rank statistics based on survival analysis.Patients with hs-cTnI>0.011ng/ml were also associated with a 1.54 folds higher risk of poor outcome(HR:1.54,95%CI:1.11-2.15).The inclusion of hs-cTnI significantly improved the risk prediction and stratification of the primary outcome(C-index:7.45 vs.7.33,P=0.002;integrated discrimination improvement 1.58%,95%CI 0.38-2.79%,P=0.01;net reclassification improvement 23.41%95%CI 4.52-44.49%,P=0.021)of the well-established model.A higher overall net benefit was also obtained in a threshold probability of 12.5-62.5%.Conclusion:Hs-cTnI provides significant prognostic value and could further remarkably improve risk stratification and prediction capabilities in NIHF patients.Section 2:Association between inflammatory and metabolic biomarkers and their prognostic value in non-ischemic heart failureAims:Inflammation and metabolic dysfunction are common in heart failure(HF)and can jointly affect HF advancement and patients’ prognosis.Few studies have illustrated the prognostic value of inflammatory and metabolic markers in patients with non-ischemic heart failure(NIHF).This study aimed to assess the association among the neutrophil-to-lymphocyte ratio(NLR),platelet-to-lymphocyte ratio(PLR),red blood cell distribution width to albumin ratio(RAR),high-sensitivity C-reactive protein(hs-CRP)and triglyceride and glucose index(TyG)and their prognostic value in NIHF patients.We also assess the combinative predictive performance of the inflammatory and metabolic biomarkers in NIHF.Methods:This study was a retrospective study.Two thousand and seventy-seven NIHF patients admitted to the heart failure care unit(HFCU),Fuwai Hospital,with complete baseline CBC and biochemistry results,were consecutively enrolled from December 2006 to October 2017.The primary endpoint was a composite outcome of all-cause mortality and heart transplant.The correlation between NLR,PLR,hs-CRP,RAR,and TyG was analyzed by the Spearman or Pearson correlation analysis.Their associations with the composite outcome were assessed by the Kaplan-Meier survival analysis and the Cox regression analysis.The concordance index(C-index),the Akaike information criterion(AIC),the Bayesian information criterion(BIC),the Brier score,the integrated discrimination improvement(IDI),and the net reclassification improvement(NRI)were calculated for assessing the increment prognostic value based on a well-established model(including age,sex,New York Heart Association class,systolic blood pressure,left ventricular ejection fraction,hemoglobin,sodium,estimated glomerular filtration rate,diabetes mellitus,treatment with angiotensin-converting enzyme inhibitors or angiotensin Ⅱ receptor blockers,treatment with β-blockers,and NT-proBNP)in the NIHF patients for biomarkers significantly associated with the primary endpoint after multivariable adjustment.Decision curve analyses(DCA)were conducted for assessing the clinical benefit of the inflammatory and metabolic biomarkers respectively and in combination.Results:The median follow-up time in this study was 1433(1341.21,1524.79)days.Since the gender no longer satisfied the Cox proportional risk assumption after 1150 days,we set 1095 days as the follow-up time for analysis.A total of 500 patients reached the composite outcome.NLR,PLR,hs-CRP,and RAR were positively correlated with each other and were negatively correlated with TyG.Multivariable Cox regression showed that per log2 increase of NLR,RAR,and hs-CRP were significantly associated with a 21.1%[hazard ratio(HR):1.211,95%confidential interval(CI):1.068-1.373,P=0.003],132.9%(HR:2.329,95%CI:1.677-3.237)and 7.5%(HR:1.075,95%CI:1.007-1.147)higher risk of all-cause mortality or heart transplant,respectively.TyG was associated with a lower risk of the primary endpoint[per standard division(SD)increase,HR:0.798,95%CI:0.708-0.899].The association between PLR and the primary endpoint was insignificant(per log2 increase,HR:1.029,95%CI:0.881-1.202,P=0.881).Model discrimination,calibration,and reclassification were improved after RAR,or TyG was added to the benchmark model(P<0.001 for all).Hs-CRP and NLR could only improve model discrimination and calibration.Compared with the benchmark model and the one with either RAR or TyG,the performance of risk prediction and stratification further improved after combining these two biomarkers[C-index:0.769 vs.0.758;AIC:1470.16 vs.1495.74;BIC:1554.37 vs.1569.73;Brier score:0.156 vs.0.161;IDI:1.69%(0.88%-2.50%);NRI:24.53%(8.54%-32.81%),P<0.05 for all].A higher overall net benefit was also obtained in a threshold probability of 20%-68.75%.Conclusion:NLR,PLR,hs-CRP,and RAR were positively correlated with each other and were negatively correlated with TyG.The prognostic value,as well as the clinical usefulness of different inflammatory biomarkers,might be different.IR seems to be a protective factor in NIHF patients as higher TyG was associated with a lower risk of all-cause mortality or heart transplantation.Combining inflammatory and metabolic biomarkers could remarkably improve risk stratification and prediction in NIHF patients with satisfied clinical applicability.Part 2:Machine learning in non-ischemic heart failure:model construction and evaluationAims:Machine learning(ML)has shown satisfactory performance in predicting clinical outcomes in heart failure(HF)patients.However,the existing ML models cannot fully reflect the clinical and prognostic characteristics in patients with non-ischemic heart failure(NIHF).Moreover,evidence on the long-term predictive value of ML models is scarce.This study aims to assess the predictive value of the ML model in NIHF patients and further evaluate the effectiveness of the ML algorithm in feature screening and the prediction performance of ML models derived from different feature screening strategies.Methods:This study was a retrospective study.Two thousand and one hundred fifty NIHF patients admitted to the heart failure care unit(HFCU),Fuwai Hospital,were consecutively enrolled from December 2006 to October 2017.The primary endpoint was a composite outcome of all-cause mortality and heart transplant.Eighty-two clinical features were enrolled,including age,gender,body mass index(BMI),systolic blood pressure(SBP),diastolic blood pressure(DBP),heart rate(HR),and laboratory and echocardiography parameters within the first 24 hours after admission,such as complete blood cell count,biochemistry,left ventricular ejection(LVEF),left ventricular end-diastolic diameter(LVEDD),etc.Fifty-six features were finally enrolled based on the missing values,clinical significance,and the results of multicollinearity analysis and univariable Logistic regression analyses:correlation coefficient ≤0.7,variance inflation factor<10,or P value ≤0.2 in Logistic regression analyses.The support vector machine(SVM)algorithm,the random forest(RF)algorithm,and the least absolute shrinkage and selection operator(LASSO)algorithm were introduced for ML-based feature selection.Four ML models were developed by the eXtreme Gradient Boosting(XGBoost)algorithm with 5-fold cross-validation,which were XGBoost-All with 56 features,XGBoost-SVM with 14 features,XGBoost-RF with 30 features and XGBoost-LASSO with 22 features,respectively.A Logistic regression model with 12 features,including age,sex,New York Heart Association class,left ventricular ejection fraction,hemoglobin,sodium,estimated glomerular filtration rate,diabetes mellitus,treatment with angiotensin-converting enzyme inhibitors or angiotensin Ⅱ receptor blockers,treatment with β-blockers,and NT-proBNP,was set up as the metric model.Model performance was assessed by the accuracy,precision,recall,F1 score,the area under the receiver operating characteristics curve(AUC),the area under the precision-recall curve(AUPRC),sensitivity,specificity,and the Brier score.Decision curve analysis(DCA)was employed for clinical usefulness assessment.Results:The median follow-up time was 1433(1341,1525)days.A total of 699 patients reached the primary endpoint.Overall,the ML-based models outperformed the metric Logistic regression model.The XGBoost-All had the best predictive performance[accuracy:0.782;precision:0.638;Recall:76.5%;F1 score:0.695;AUC:0.842(0.824-0.860);PRAUC:0.715;Brier score:0.15],followed by the XGBoost-RF model[accuracy:0.767;precision:0.622;Recall:72.4%;F1 score:0.669;AUC:0.829(0.811-0.847);PRAUC:0.704;Brier score:0.16].Model performance was equivalent between XGBoost-SVM and XGBoost-LASSO.DCA analyses showed that the net benefit of the four XGBoost models was higher than that of the benchmark Logistic model within a threshold probability of 5%-90%.Within a threshold probability of 5%-50%,XGBoost-ALL has a higher net benefit than the other XGBoost model.Clinical usefulness was similar among the four XGBoost models when a threshold probability was over 60%.Conclusion:The machine learning models could significantly improve outcome prediction in NIHF patients with satisfied clinical usefulness,which could further optimize the management of these patients. |
PDF Full Text Request