Introduction

Albumin is the most abundant protein in blood plasma, primarily served as blood transporter and involved in maintenance of normal plasma oncotic pressure [1]. Hypoalbuminemia (i.e., concentration of blood albumin that is abnormally low) is attributed to a number of factors, including reduced synthesis associated with liver dysfunction, increased catabolism in the wake of infection, and altered distribution by an increase in transcapillary escape rates [2]. Decades of medical observation have demonstrated a relationship with hypoalbuminemia and increased risk of morbidity and mortality [3,4,5,6]. More recently, studies have identified hypoalbuminemia as a risk factor associated with faster disease progression and worse outcomes in ALS, Guillain-Barré syndrome, and stroke [7,8,9].

Given the extent of trauma associated with spinal cord injury (SCI), need for surgical interventions, and high incidence of infections, hypoalbuminemia is unsurprisingly common in the acute phases of injury [3, 5]. In addition to increasing the risk of mortality [2, 4], lower concentrations of serum albumin early after injury (≤1 month) are also associated with more severe SCI and predicted poor neurological recovery [10]. These seminal observations were demonstrated utilizing data from the GM-1 gangliosides acute SCI trial [11] demonstrating a unique potential for serum albumin as a prognostic biomarker. However, the generalizability of these observations, outside of a clinical trial, warrants further investigation in an independent cohort.

The overall goal of the current study was to further assess serum albumin and neurological outcomes after acute, traumatic SCI. The specific aim was to utilize contemporary data, reflecting more modern acute SCI management practices, collected outside of the scope of a clinical trial, in routine clinical management. Based on our previous findings, we hypothesized that serum albumin would be related to initial injury severity and independently predict long-term neurological outcomes after acute SCI. To address our hypothesis, we performed a secondary analysis of open source data from the Spinal Cord Injury Rehabilitation (SCIRehab) study (2007–2010) (https://www.icpsr.umich.edu/icpsrweb/ADDEP/studies/36567).

Methods

Study design and data source

The primary data used for this study were accessed from the SCIRehab study through Archive of Data on Disability to Enable Policy and research (ADDEP) [12]. Design, recruitment, inclusion criteria, and enrollment details have been previously described in detail [13]. Briefly, the SCIRehab study enrolled individuals aged ≥12 years with traumatic SCI rehabilitating at six participating centers in the United States from 2007 to 2009. Participant demographics and injury characteristics were extracted from the participant medical record (part of the National Institute on Disability and Rehabilitation Research SCI Model Systems Form I).

Inclusion criteria

For the purpose of our analyses, we included participants with cervical and thoracic injuries only to remain consistent with our previous study [10]. Additional inclusion criteria were applied for analysis of 1-year post-injury outcomes, such that participants who were non-ambulatory at admission to rehabilitation were included, and completed assessment of albumin and neurological outcomes (i.e., lower extremity motor scores (LEMS) and AIS grades).

Outcomes (dependent variable)

As a measure of residual function caudal to both thoracic and cervical lesions, LEMS and American Spinal Injury Association Impairment Scale (AIS) grades were examined as dependent variables at admission to rehabilitation. In brief, key muscles in LEMS were examined according to the International Standards of Neurological Classification of Spinal Cord Injury (ISNSCI) [14], with a maximum scores of 25 points for each left and right sides (for a maximum total scores of 50). AIS grades are an aggregate of sensory and motor function, ranging from AIS A (most severe) to D (least severe) [14]. To further differentiate AIS D injuries, we employed Functional Independence Measure (FIM) [15] to categorize participants based on their ambulatory status. Information from the ISNSCI (i.e., LEMS and AIS grades) and FIM were collected as part of the National Institute on Disability and Rehabilitation SCI Model Systems Form I.

To examine the relationship between serum albumin concentration and neurological recovery, we focused on change of LEMS and marked recovery at 1-year post-injury. Change of LEMS was calculated as LEMS 1-year post-injury minuses LEMS at admission. Marked recovery occurred for an individual converting 2 AIS grades or for individual regained walking function (i.e., AIS A to C, or AIS B to D/walking or from AIS C/D to walking). Similar to admission, walking was determined based on 1-year post-injury examination of FIM.

Predictor (independent) variables and covariates

The primary independent variable was lowest albumin concentration (CRLOWALBUMIN) (searchable in SCIRehab database). In the SCIRehab study, albumin concentrations were extracted from participant medical records. As an aggregate for each participant, mean, highest, and lowest albumin concentrations were determined through rehabilitation (i.e., raw values are not available). This means that albumin was not obtained at consistent timepoints in all individuals, but rather at variable time-points during their rehabilitation. Lowest recorded serum albumin concentrations were a priori selected for our analysis. This was done based on our previous study, where the lowest concentration, on average, yielded the highest predictive function of future neurological function [10]. The average number of elapsed days from injury date to rehabilitation admission was also reported. Potential confounders included LEMS and AIS grades at admission to rehabilitation.

Data availability, cleaning, and analysis statement

The data utilized in the current study was derived from the ADDEP database, which is a publicly accessible database https://www.icpsr.umich.edu/icpsrweb/ADDEP/studies/36567. Our cleaning data process and the dataset (excel file) that were used in this analysis are both publicly available at https://github.com/AnhKhoaVo/ADDEP. Briefly, the variables used in our data (searchable in the ADDEP database) were: “REVIEWASIAGRADEADM” (i.e., AIS grade at admission), “BASAImp” (i.e., AIS grade at 1-year), “AFLMODRB” (i.e., FIM-walking at admission to rehabilitation), and “BFIMLMod” (i.e., FIM-walking 1-year post-injury). We then updated our own AIS grades to include ambulatory statuses at admission, so that if individuals had AIS D and was able to walk at admission, they would be classified as “D/Walk” in our new variable, “ASIAGRADE_WALK”.

We also calculated a new variable representing marked recovery at 1 year, “Marked_Recovery_Annual_2”. Briefly, anyone who improved two AIS grades from admission and/or regained walking function, was then regarded as “Achieved Marked Recovery”. For example, for AIS A at admission, improvement to AIS C or D after 1 year is considered marked recovery; For AIS B at admission, 2 AIS grades improvement or walking regained at 1 year is considered marked recovery; whereas for AIS C and D, regaining ambulatory status at 1 year is marked recovery.

In terms of change scores between LEMS at baseline and 1-year post injury, we created “LOWER_MS_REHAB” (i.e., based on sum of left and right motor scores from L2 to S1 at admission to rehabilitation), and “LOWER_MS_ANNUAL” (i.e., based on sum of left and right motor scores from L2 to S1 at 1-year post-injury). We then subtracted “LOWER_MS_ANNUAL” to “LOWER_MS_REHAB” to create “Change_Scores”. Refer to https://github.com/AnhKhoaVo/ADDEP/blob/master/ADDEP_Clean_Analysis.R for details of our data cleaning process.

Descriptive statistics

Demographics, injury characteristics, and albumin concentrations were compared between included and excluded cohorts to assess selection bias. Specifically, chi-squares tests were used for categorical variables (i.e., demographics and injury characteristics), while t-tests were used for continuous variables (i.e., albumin concentrations). Our coding for descriptive statistics can be found at https://github.com/AnhKhoaVo/ADDEP/blob/master/ADDEP_Clean_Analysis.R.

Unbiased recursive partitioning: conditional inference tree

Our primary statistical analysis applied the unbiased recursive partitioning (URP) conditional inference tree (URP-CTREE). In brief, URP-CTREE is a tree-based regression model using sequential tests of independence between predictors (e.g., albumin concentration, and LEMS admission) and outcomes (e.g., change of LEMS and marked recovery) [16].

Fundamentally, URP-CTREE follows two steps, which are repeated after each split of the original heterogeneous population. First, the algorithm assesses whether any of the predictors is statistically associated with the outcome by multiple-testing corrected p value (i.e., Bonferroni correction). If the initial null hypothesis between predictors and outcome cannot be rejected (no statistically significant association between predictors and outcome), the algorithm stops producing any split. If the null hypothesis can be rejected (at least one predictor is significantly associated with the outcome), the algorithm splits by selecting the predictor with smallest P value. Second, after the tree has split, the algorithm then calculates another possible split based on the new subset created from previous significant predictor. Each split is evaluated by a two-sample linear statistic (e.g., Fisher’s exact test for nominal responses). If none of the variables included in a model is significantly associated with outcome, no splits are created for the data. One advantage of URP-CTREE is the ability to handle both regression and classification models without being subjected to their restrictive requirements. URP-CTREE can employ non-parametric approach, which allows itself exempt from normality assumption, and simultaneously perform variable selection known to be more stable than stepwise selection in linear and logistic regressions [17]. Most importantly, URP-CTREE provides cut-off values that are valuable application for clinical settings. Thus, URP-CTREE has been previously applied in a number of SCI studies [18,19,20].

Statistical analyses

For the baseline association analyses, bivariable URP associations were performed between lowest albumin concentrations and AIS grades at admission, and between lowest albumin concentrations and LEMS at admission. For the primary analysis of 1-year outcomes, bivariable URPs were done for lowest albumin concentration and each outcome (i.e., change of LEMS and marked recovery). Multivariable URPs for each outcome were also created adjusting for LEMS at admission (for change of LEMS outcome) or AIS grades admission (for marked recovery outcome). The performances of both bivariable URPs models (change of LEMS and marked recovery) at 1-year post-injury analyses were also assessed via cross-validation methods (10 folds). Details regarding the analyses are available online at https://rpubs.com/AnhKhoaVo/5861028.

Results

Cohort summary

A total of 430 and 439 participants were included in the association analyses of LEMS and AIS grades at admission, respectively (Table 1). 117 and 167 participants were analyzed in the change of LEMS and marked recovery between admission and 1-year post-injury, respectively (Table 2). The average elapsed days from injury date to rehabilitation admission date was 30 days (±27). The average elapsed days from injury date to lowest albumin recorded was 45 days (±36). The average number of times albumin concentration values were captured in each individual was 5 times (±6). The percentage of individuals whose albumin concentrations was recorded once was 22.7%. Most included and excluded participants shared similar demographics and injury characteristics.

Table 1 Cohort Summary at admission to rehabilitation.
Full size table
Table 2 Cohort Summary at 1-year post-injury.
Full size table

Admission to rehabilitation analysis

The relationship between LEMS and AIS grade and serum albumin concentrations at admission to rehabilitation are shown in Fig. 1a, c. At admission, URP of LEMS identified two participant cohorts based on a cut-off serum albumin concentration of 2.6 g/dL (p < 0.001) (Fig. 1b). For URP of AIS grades at admission to rehabilitation, two participant cohorts were also identified, also based on an albumin cut-off concentration of 2.6 g/dL (p < 0.001) (Fig. 1d). Both analyses, of LEMS and AIS grade at admission to rehabilitation, support that lower serum albumin concentrations are significantly associated with more severe injuries.

Fig. 1: Summary figure of Albumin concentration and outcomes at admission to rehabilitation.
figure 1

a Scatter plot between albumin concentration and lower extremity motor scores (LEMS) at rehab admission. b Unadjusted unbiased recursive partitioning (URP) between LEMS at admission and albumin concentrations. Albumin concentration was a significant variable at a cutoff value of 2.6[g/dL]. Cohort with albumin >2.6[g/dL] has mean LEMS = 11, but cohort with albumin ≤2.6[g/dL] has mean LEMS = 4. c Boxplot between albumin concentration and AIS grades at admission (i.e., AIS A, B, C, D, and D&Walk). Albumin concentration in AIS C and D&Walk is significantly different from albumin concentration in AIS A and AIS B. d URP between albumin concentration and AIS grades at admission. Albumin concentration was a significant variable with cutoff values 2.6[g/dL]. For cohort of albumin ≤2.6[g/dL], 60% was AIS A, and less than 10% was AIS D&Walk. However, in cohort of albumin >2.6[g/dL], only 40% was AIS A, and roughly 20% was AIS D&Walk.

Full size image

1-year post-injury Analysis

The prognostic value of serum albumin concentrations to predicting neurological recovery is shown in Fig. 2. For change in LEMS between rehabilitation admission and 1-year post-injury, URP yielded two participant cohorts based on a serum albumin cut-off of 2.8 g/dL (p < 0.001, Fig. 2a). A cross-validation method with 10 folds was used to assess the performance of this bivariable URP, and indicated no overfitting in URP (i.e., Root Mean Square Error remained consistent). Adjusting for LEMS at admission, URP yielded two participant cohorts, based solely on selected a cut-off value for LEMS (p < 0.001, Fig. 2b). The absence of serum albumin in the adjusted model suggests that concentrations did not provide any additional prognostic information beyond baseline LEMS.

Fig. 2: Summary figure of Albumin concentration and outcomes at 1-year post-injury.
figure 2

a The bivariable URP model between lowest albumin concentration and change of lower extremity motor scores revealed albumin to be significant (p < 0.001). In the albumin ≤2.8[g/dL] cohort (N = 66), change of LEMS’ median = 0, and mean = 2.36; however, in the albumin >2.8[g/dL] cohort (N = 51), change of LEMS’ median = 2, and mean = 9.35. b In the multivariable URP for change of LEMS scores adjusting for both lowest albumin concentration and LEMS at admission, only LEMS at admission showed up suggesting that lowest albumin concentration was not associated with outcome. c The bivariable URP model between lowest albumin concentration and marked recovery indicated albumin was significant (p = 0.001). In albumin ≤3.1[g/dL] cohort (N = 129), less than 20% achieved marked recovery; whereas in albumin >3.1[g/dL] cohort (N = 38), 50% achieved marked recovery. d In the multivariable URP model for marked recovery adjusting for AIS grades and serum albumin, only AIS grades at admission showed up suggesting that lowest albumin concentration was not associated with outcome.

Full size image

At a concentration of 3.1 g/dL, serum albumin yielded two-terminal categories (nodes) for marked recovery (p = 0.001, Fig. 2c). This bivariable URP’s performance was also assessed via cross-validation methods with 10 folds, which reported accuracy of 75%, AUC (Area Under the Curve) for ROC (Repeated Operating Characteristics) of 59.9%. After adjusting for admission AIS grades and albumin concentration, three cohorts of participants were identified (nodes 2, 4, and 5) (p < 0.001, Fig. 2d). The absence of serum albumin suggests that albumin concentration did not perform better than AIS grades at admission in terms of predicting marked recovery.

Discussion

The primary goal of this study was to examine the prognostic value of serum albumin concentrations and neurological recovery after SCI. In support of our hypothesis, injury severity at admission to rehabilitation was related to serum albumin concentrations. Without adjustment for baseline injury characteristics, serum albumin concentrations also predicted neurological recovery. However, this prognostic value with long-term neurological outcome did not persist after adjustment for baseline injury characteristics. These observations, based on a contemporary collection of data, partially agree with our previous findings that low serum albumin concentrations are predictive of poor neurological recovery.

In a previous study, serum albumin concentrations measured between 24 h and 1 month were associated with both baseline injury severity and predictive of long-term neurological outcome [10]. These seminal observations were made using data from a completed Sygen clinical trial, where serum albumin was collected to assess the therapeutic safety of GM-1 gangliosides in acute SCI [21]. An obvious limitation is that the Sygen trial ran from 1992 to 1997 [22], and as a result, was subject to outdated acute management practices (i.e., administration of methylprednisolone). As a result, methylprednisolone could affect blood chemistry values [23]. To address this concern, we performed an analysis on a contemporary data source, where serum albumin concentrations were extracted from medical records as part of a large observational study [12].

Consistent with previous studies [5, 10], lower serum albumin at admission to rehabilitation was associated with more severe SCI. This was evidenced for both AIS grade and LEMS, based on a cut-off serum albumin concentration identified by URP. By nearly all standards, albumin concentration of 2.6 g/dL up to 45 days post-injury represents a state of hypoalbuminemia [10]. Mechanistically, this may reflect a number of factors, including a higher degree of trauma and rates of complications (e.g., infections) in the initial days to weeks post SCI [24]. One potential underlying mechanism of hypoalbuminemia is to reflect the prolonged state of infections, thus limiting the extent of neurological recovery [10].

Our primary analysis aimed to determine if serum albumin concentrations could predict neurological recovery, to a similar degree or beyond that already attributable to baseline injury characteristics. To this end, serum albumin concentrations > 3.1 and 2.8 g/dL (as identified by the URP) were predictive of double the proportion of participants achieving marked recovery and significantly greater recovery of LEMS, respectively. However, serum albumin concentration did not improve the prediction of neurological recovery beyond that which could already be achieved from baseline outcomes derived from the ISNCSCI (see Fig. 2b, d). The exclusion from the “tree” ultimately suggests that serum albumin concentration is only useful for cases where neurological details at admission (i.e., LEMS or AIS grade at baseline) are not available. This may arise because of concomitant injury (e.g., fractures) or in cases where an individual is unable to follow the instructions required of them to test sensory and motor function.

As a point of interest, both the current analysis and previous study evaluated the impact of albumin on neurological outcomes using data collected for alternative purposes. Specifically, the Sygen trial was planned to evaluate the efficacy of GM-1 gangliosides on sensorimotor recovery from acute SCI [11], and the Spinal Cord Rehabilitation study was designed to identify which rehabilitation interventions are strongly associated with positive outcomes [13]. The investigators of the latter have made their data entirely open, which is accessible through the ADDEP [https://www.icpsr.umich.edu/icpsrweb/ADDEP/studies/36567]. The advantages of secondary analyses of existing data are reduction in time and costs associated with data collection. In addition, secondary analyses of existing data have demonstrated disease-modifying effects of various events and interventions associated with acute SCI, including the impact of infections [25] and acute pain medications [19, 26, 27].

Our slightly different outcome could be due to the differences between these two studies. First, in the SCIRehab, serum albumin may have only been recorded in a subset of participants, where monitoring serum albumin was deemed clinically important (albumin was only recorded in 63% of individuals in SCIRehab Study, see https://www.icpsr.umich.edu/icpsrweb/ADDEP/studies/36567); whereas, blood chemistry was recorded in the majority of participants in the Sygen trial. Second, serum albumin in the current study was extracted from medical records as a single value (i.e., lowest recorded albumin) and recorded at an arbitrary time-point (i.e., average 45 days after injury). In comparison, raw serum albumin concentrations in the Sygen trial were determined at fixed time-points after injury (e.g., 24–72 h, 1-month). Thirdly, given the differences in study design (i.e., observational study versus prospective clinical trial), the SCIRehab incurred a higher dropout at 1-year post injury in terms of neurological outcomes (only 30–40% individuals followed up at 1-year) than the Sygen trial, leading to various other types of selection bias (e.g., sampling and attrition).

Finally, the statistical approach varied between studies. Our analysis of the SCIRehab utilized a machine learning approach, URP [16], whereas the previous study applied more conventional linear and logistical regressions. From a technical perspective, URP manages various types of independent and dependent variables (e.g., neurological recovery as continuous and dichotomous outcomes), and is not subject to parametric assumption [17], unlike conventional linear and logistic regressions. URP can also process interaction from large number of predictors and provide variable importance better than stepwise regression in standard regression methods [17].

In summary, we have confirmed that albumin concentration is associated with baseline characteristics and long-term neurological recovery after SCI. At admission to rehabilitation, serum albumin concentrations provide a crude estimate of injury severity and future neurological recovery. The potential application of serum albumin to the clinic warrants prospective study.