This is a review of the study by Badhiwala et al. [1], “The Influence of Timing of
Surgical Decompression for Acute Spinal Cord Injury: A Pooled Analysis of Individual
Patient Data” published in the Lancet Neurology Dec 21, 2020.
This study is important for informing practice due both to its content, conclusions,
and state-of-the-art methodology. It further informs the question as to whether the
timing of decompression surgery has a “neuroprotective” influence following spinal
cord injury (SCI). In the study, modern meta-analytical techniques were utilized to
analyze pooled harmonized data from 4 independent prospective multicenter data sources
covering the years 1991 to 2017. The data sets capture acute to longer-term neurological
outcomes following traumatic SCI and utilized the North American Clinical Trials registry
[2], STASCIS (Surgical Timing in Acute Spinal Cord Injury Study) [3], Sygen [4], and
National Acute Spinal Cord Injury Study (NASCIS) III [5]. The main variable assessed
in this study was the impact of the timing of surgical decompression on neurological
recovery. The authors were able to analyze pooled individual patient data (IPD) in
aggregate from source material. The endpoint was the change in motor score from baseline
to 1-year follow-up.
There are several novel aspects of this study including the large sample of aggregate
data that consisted of 1,548 eligible patients. An important analysis technique that
was utilized is the restricted cubic spline method. This method allows testing for
a nonlinear association between the time to decompression and the change in motor
score. Time was treated as a continuous variable and the proportions of the time bins
also shown in Fig. 3 in the study. It is notable that the largest cohorts were in
the 8- to 24-hour time to surgical decompression period frame even including data
from the 1990s. Thus, in the included studies early decompression was more common
than late. A steep nonlinear association was found that flattened out after 36 hours.
This implies that earlier decompression times have more impact on outcome than later.
One issue to bear in mind is that there can be inaccuracy in the very early ISNCSCI
(International Standards for Neurological Classification of Spinal Cord Injury) exam
[6]. This could lead to an overestimate of the effect if, for example, a person was
scored as the American Spinal Cord Injury Association Impairment Scale (AIS) A on
a very early exam (within 4 hours) initially [7]. As conversion (for example to an
AIS B or C) from this early time period occurs at a high rate, this change might be
interpreted as recovery from the surgery but could also be due to natural conversion
from an early exam or due to an inaccurate initial exam in the challenging very early
period after SCI.
Another important data analysis issue was in performing a one-stage hierarchical mixed-effects
regression that allowed the IPD to be assessed against the regression variables at
once. The hierarchical methodology allowed the impact of the individual datasets that
spanned a substantial time period to be examined. One problem that can arise in this
IPD approach is that the aggregate data may take on statistical distribution properties
that were not present in the original studies. This is often described as how the
individual studies cluster for the outcome of interest and the observed variances.
Nuances of studies, historical, and baseline differences can be “washed out” by these
methods. To correct for this, there are 2 main approaches, a stratified or random
intercept [8].
We get some insight into the importance of this modeling from Table S4 that compares
the baseline characteristics between the 4 data sources, where for example one sees
the notable difference in the average age across the studies, and e.g., the fraction
of cervical injury, influenced by the fact that STASCIS enrolled only cervically injured
patients. The Forest plots for the odds ratios dichotomized to early and later surgery
are shown in Fig. 2, where one can appreciate that there are some considerable differences
between the studies, with a smaller effect of early surgery on total motor score and
pin prick score in the NASCIS II and Sygen studies respectively.
The overall IPD meta-analysis variance is estimated by the inverse variance method
to assign greater weight to larger studies. The contribution of the variables: age,
mechanism-of-injury, AIS grade, spinal injury level, and administration of methylprednisolone
was adjusted for as fixed effects covariates to allow the effect of the surgery time
(early/late) to be estimated more accurately.
We are told that the datasets were harmonized but it’s not clear if that presented
challenges and what compromises were needed? As examples of limitations, the original
endpoint of NASCIS III was at 6 months and not 1 year, and only 14 extremity muscles
were assessed versus 20 as is the current standard [5,9]. In addition, the reported
analysis used data from one side of the body only. There was no placebo control in
NACIS III, only methylprednisolone and tirilazad were tested. Surgical decompression
is not mentioned in the index report or for the 1-year follow-up, an indictor perhaps
of the important changes in care perception across the 32 ensuing years. For the Sygen
study [4,10], all patients received methylprednisolone. The outcome assessment used
the Modified Benzel scale that added stratification for the AIS D grade. The authors
were thus reliant on the accuracy of reporting in case report forms especially for
NASCIS III and Sygen where timing of decompression was not a key variable.
The problem with missing data is intrinsic to the exercise of meta-analysis. In this
study, 2 adjustment methods were utilized, one was the last analysis carried forward,
and the other was multiple imputation using a Markov chain analysis. Sensitivity analyses
have the purpose to examine the robustness of the study findings if the methods are
changed [11]. Three comparisons were made, 1 stage versus 2 stage for each key outcome,
changing the method of data imputation, and eliminating data imputation altogether.
None of the alterations to the analysis substantially changed the findings. However,
the absence of 1-year data may generally reduce the maximal neurological recovery
observed [12], despite the use of imputation. In summary, the IPD study allowed a
relationship of timing of surgery to neurological outcome to be assessed from multiple
studies of importance in the SCI field, even though this variable was not initially
reported from the NASCIS and Sygen studies. While some questions are raised regarding
this analysis, important information can be gleaned, and the application of these
advanced data analysis methods in the SCI field enhances the value of carefully collected
data.