Statistical analysis
For all analyses, SAS v9.4 and R v 3.6.1 and a significance level of
0.05 was used. Descriptive statistics included median and percentiles
(minimum, 25th percentile, 75thpercentile, maximum) for continuous variables, and counts and
percentages for categorical variables. For all inferential analyses,
glucose CV was log transformed due to non-normal distribution. Base 2
was chosen for ease of interpretation and the estimates were back
transformed.
Linear regression was used to test for associations between potential
risk factors for increased glucose CV at each time intervals. Potential
predictors included: age, BMI category, primary diagnosis, conditioning
regimen, and pre-transplant steroids, asparaginase, insulin, and
radiation. For the post-HSCT glucose CV model, additional predictors
were considered: HSCT type, and post-HSCT steroids, radiation, and total
parenteral nutrition (TPN). The final models were chosen based on the
lowest Akaike information criterion (AIC). An additional multivariable
linear regression model evaluated the effects of post-HSCT
immunosuppressive agents (sirolimus, tacrolimus, cyclosporine,
methotrexate) on post-HSCT glucose CV.
Associations between glucose CV and time-to-HSCT outcomes were examined
using Cox proportional hazards models. Glucose values were censored at
the time of the event. For time-to-infection, ICU and GVHD analyses,
death was treated as a competing risk. For TRM, death due to other
causes was treated as a competing risk. Associations between pre-HSCT
and day 0-30 glucose CV and both time to death and TRM were adjusted for
HSCT type, severe GVHD diagnosis, and the need for post-HSCT steroids.
The model evaluating day 0-30 glucose CV and GVHD excluded patients with
GVHD in the first 30 days (n=9) and adjusted for steroids received
post-HSCT but pre-GVHD diagnosis.
Because infections and ICU stays frequently occurred soon after HSCT
(the median days to infection was 19 (interquartile range [IQR]
10.5-32.5)), only pre-HSCT glucose CV could be evaluated for an
association with time to these events. Similarly, due to the short
follow-up window post-HSCT and pre-infection/ICU stay, we were unable to
adjust for steroid exposure, but we did adjust for HSCT type.
Associations between pre-HSCT glucose CV and infection subtypes were
evaluated using logistic regression, as patients could have had multiple
types of infections, requiring significant censoring for each infection
subtype. Viremia was evaluated only in patients who underwent allogeneic
HSCT, as these were the only patients who underwent routine viral
surveillance testing. Interactions between glucose CV and transplant
type as well as glucose CV and steroids were tested for in all models
and included if p<0.10 for the interaction term.
To assess for the effect of possible outliers in glucose values, all
analyses were re-run without glucose values > 500 mg/dL.
This did not significantly change any results, so is not reported.
Results