190 likes | 207 Views
Explore predictive modeling of RVAD use in LVAD recipients, including risk factors and model building techniques. Understand the complexity of predicting RVAD use and implications for clinical practice.
E N D
Predictive Modeling of RVAD Use in LVAD Recipients Michael S Kiernan, MD, SM Assistant Professor, Tufts University Medical Director Ventricular Assist Device Program, Tufts Medical Center
Presenter Disclosure Information • I will notdiscuss off label use or investigational use in my presentation. • I havefinancial relationships to disclose. • Employee of: None • Consultant for: None • Stockholder in: None • Research support: None • Honoraria from: ThoratecCorporation
Timing and of RVAD Planned Early Durable Temporary Delayed Morgan Ann Thorac Surg 2004;77:859
Rate RHF: 15-44% Kalogeropoulos JHLT 2015;34(12):1595
Performance of risk prediction models tested in an independent cohort Kalogeropoulos JHLT 2015;34(12):1595
Complex etiology of post-LVAD RH failure RV Dysfunction LVAD & Ventricular Inter-dependence RHF RAP Afterload (PVR) Lampert. JHLT 1025;34:1123 Goldstein. ATS 2003;75:S42
Consort diagram of INTERMACS study cohort All Patients (N=11,162) • 1,186 Patients • 610 pulsatile LVAD • 323 pulsatile BiVAD • 253TAH Primary CF-LVAD (N=9976) Any RVAD (5.2%) (N=521) Early RVAD (≤ 14 d) (3.9%) (N=386) No RVAD (94.8%) (N=9455)
Non-adjusted associations of continuous variables with outcomes: If straight line, then linear fit is ok • Assumes independent variables are linearly related to log odds of the outcome
Model building: Handling of nonlinear relationships between predictors and outcomes Dichotomization median
Model building: Handling of nonlinear relationships between predictors and outcomes Truncate beyond a ‘cut-point’
Building a predictive model: handling missing data 1st pass at a multivariable model included all variables from Table (even those with >20% missing) • Stepwise selection with a p-value of 0.10 to enter the model and p=0.05 to stay in the model
Preliminary model of RVAD predictors • Missing data is often not missing at random • Final model utilized multiple imputation which allowed us to utilize observations on all available subjects • Created 10 new datasets and re-estimated the effect size (ORs)
Sensitive analysis: How well does model predicts the composite endpoint of death or RVAD within 2 weeks of LVAD surgery? [n=608/9976 with composite event (6.1%) vs. n=386 (3.8%) with RVAD]
Conclusions • Predicting RVAD use is complicated due to complex physiology of RHF • Low prevalence of RVAD use will limits clinical utility of risk scores • Many previously identified risk factors from smaller studies were confirmed as risk factors in the INTERMACS model • It takes an expert to build a risk prediction model • (Thank you Robin Ruthazer!) • Next steps: • Create an INTERMACS validation cohort • Describe the late RVAD recipients (2%) • Evaluate less severe RHF not resulting in RVAD