Statistical Analysis
Assessment of the relationship between the routine echocardiographic
diastolic parameters and 2D-STE derived parameters was done using
repeated measures mixed linear regression models with e’ or E/e’ as the
dependent variable, strain measurements as fixed independent variables,
and patient serial number as the random variable. To assess the
predictive ability of Dst on significant GLS reduction, individual
logistic regression models were built with significant GLS reduction as
the dependent variable and relative change in each Dst segment between
T1 and T2 as independent variables. Using the results of the above
models the best predictor of significant GLS reduction was used in a
multivariate model to assess its ability to independently predict
significant GLS reduction. First, a multivariate model was built with
covariates including relative GLS reduction between T1 and T2, baseline
cardiac risk factors, cardiotoxic chemotherapy used and
cardio-protective medication used. The above primary model was then
narrowed using a stepwise forward and backwards Akaike information
criterion (AIC) based method in order to select the best predictive
model which has lowest AIC. To further illustrate the diagnostic
predictive power of Dst alone or in combination with the other model
covariates receptor-operator (ROC) curves were built and AUC with 95%
CI and Youden indexes were calculated. Comparison between AUC of ROC
curves was done using the DeLong & DeLong method. To detect whether
adding Dst data contributed to the multivariate model predictive
ability, net classification index (NRI) was calculated for a logistic
model with and without the added variable, 95% confidence interval for
NRI (and its positive and negative components) was calculated using a
bootstrapping method. Continuous variables are shown as mean±SD, while
discrete variable as n(%). Results were considered significant when
p<0.05. As this is a primary proof of concept investigation,
all assessments were considered hypothesis generating and were not
corrected for multiple comparisons. All calculations were done using R
version 3.5.0, R Foundation for Statistical Computing, Vienna, Austria.