Development And Evaluation Of Individual Risk-prediction Model For Patients With NSTEMI

Objectives: There are increasing number of patients with non-ST-segment elevation myocardial infarction(NSTEMI)in China recent years.For individual risk assessment and secondary prevention,we aimed to develop and evaluate a novel risk-prediction model using classical Cox proportional hazard regression model and random survival forest algorithm,meanwhile,a lot of candidate clinical predictors were collected for model building.Methods: From 2017.1 to 2017.12,consecutive 1357 diagnosed NSTEMI patients were enrolled in Tianjin Medical Center and followed up 1 year,we collected their baseline demographics information,cardiology test,hematological test,biochemical biomarkers,treatment and medical history.The primary outcome is all-cause death.The secondary outcome was major adverse cardiac events(MACE: including all-cause death,hospital admission for unstable angina,hospital admission for heart failure,nonfatal recurrent myocardial infarction and target lesion revascularization).All cases were randomly split into training cohort(N=945)and validation cohort(N=412)in the proportion of 7:3,the former was used for model fitting and the later was to evaluation.We performed univariate analysis for collected variables at first,then LASSO regression were conducted for selecting variables of prognostic value.Secondly,using these selected variables fit Cox proportional regression model and random survival forest model,denoted as Cox-LASSO and RSF-LASSO,for comparison,we used eight variables involved in GRACE score to fit this two models again,denoted as CoxGRACE and RSF-GRACE.At last,we evaluated above four models from the aspect of discrimination,calibration and improvement on validation cohort,in addition,the corresponding methods were time-dependent concordance index,GND test and calibration plot,NRI and IDI,respectively.Results: 1.The mean±SD of follow-up days of training cohort was 314±112,during which there were 224 MACE events including 40 all-cause deaths,44 patients hospitaloed for unstable angina,and 56 for heart failure,19 patients suffered from nonfatal recurrent myocardial infarction and 4 patients target lesion revascularization among 945 patients.The mean±SD of follow-up days of validation cohort was 311±11,during which there were 104 MACE events including 17 all-cause deaths,20 patients hospitalized unstable angina,26 for heart failure,13 patients suffered from non-fatal recurrent myocardial infarction and 3 patients target lesion revascularization among 412 patients.2.We used ten-fold cross-validation to choose optimal λ in LASSO regression.For all-cause death,there were six predictors left,five of which were statistically significant,i.e.age(per 5 years),HR=1.023(1.006,1.041);heart rates,HR=1.023(1.006,1.041);neutrophil,HR=0.866(0.757,0.991);log-NT-pro BNP,HR=2.390(1.743,3.277);log-CK,HR= 1.488(1.077,2057).For MACE events,there were sixteen predictors left,six of which were statistically significant,i.e.age(per 5 years),HR=1.214(1.105,1.333);diastolic pressure,HR=1.039(1.008,1.071);abnormal ST-segment in ECG,HR=1.639(1.065,2.523);LAD,HR=1.045(1.017,1.074);log-NT-pro BNP,HR=1.589(1.380,1.830);log-CK,HR= 1.488(1.289,1.501).3.We evaluated the four models from the aspect of discrimination,calibration and improvement on validation cohort.As for discrimination,the decreasing order of timedependent C-index of four models for all-cause death were as follow: RSF-LASSO> Cox-LASSO> RSF-GRACE> Cox-GRACE,the decreasing order of that for MACE events remained unchanged.As for calibration,four models were well calibrated in those patients according to GND test,however,the Cox proportional hazard regression tends to underestimate the probability of both all-cause death and MACE event in those patients of low predictive risk.Setting Cox-GRACE as reference model,we found RSF-LASSO could improve the precision of Cox-GRACE model.For all-cause death,IDI =0.03(0.024,0.09),continuous NRI= 0.27(0.01,0.40);as for MACE events,IDI =0.02(0.04,0.11)continuous NRI= 0.18(0.08,0.39).Conclusion: 1.Five predictors for all-cause death in patients with NSTEMI were age,heart rates,neutrophil,NT-pro BNP and CK.Six predictors for MACE events were age,diastolic pressure,abnormal ST-segment in ECG,LAD,NT-pro BNP and CK.The combination of these predictors yield a better discrimination.2.As for calibration,four models were well calibrated in those patients under low-risk,however,the Cox proportional hazard regression tends to underestimate the probability of both all-cause death and MACE event.3.RSF-LASSO model could improve the prediction performance of Cox-GRACE and produce a clinical benefit.