Prognostic factors of Guillain-Barré syndrome after intravenous immunoglobulin or plasma exchange

Outline of the study.

All patients were entered into the Dutch GBS trial, a multicenter clinical trial comparing the effect of IVIg and PE, the background, design, and results of which have been published elsewhere.4

Clinical assessment at start of treatment.

Before treatment began, the following features were determined: age, sex, antecedent episodes in the 4 weeks before onset of weakness (a gastrointestinal or upper respiratory tract infection), time from onset of weakness until start of treatment, distribution of muscle weakness (predominantly distal, proximal, global, or mixed according to the method described elsewhere),15 functional (F) score,4 Medical Research Council (MRC) sum score,16 and presence of sensory loss and cranial nerve deficits.

Clinical assessment during follow-up.

During follow-up, the F score, MRC sum score, and presence of cranial nerve palsies were determined three times a week during weeks 1 and 2, once a week up to week 6, and in weeks 8, 10, 14, 18, 22, and 26. The method of assessment of severity of sensory loss (none, mild, or severe) during the first 2 weeks of follow-up is described elsewhere.17 The 147 patients with GBS were followed up in a prospective study, and the highest functional score, lowest MRC sum score, severity of sensory loss at nadir, and time until the lowest MRC sum score (time until nadir) were determined from the data. These factors were known within the first 2 weeks of follow-up.

Electrodiagnostic studies at start of treatment and during follow-up.

Electrodiagnostic testing (electromyography [EMG]) was performed using standardized conventional techniques at entry and at 1 and 4 weeks after randomization. The details of testing have already been reported.18 Because other studies emphasized the importance of the prognostic value of electrophysiologic data,6-8 these variables were tested extensively. We used the amplitude of the compound muscle action potential (CMAP) of the abductor pollicis brevis muscle after distal stimulation of the median nerve, the CMAP of the abductor digiti quinti muscle after distal stimulation of the ulnar nerve, and the sum of these CMAPs. The EMG data obtained at randomization and 1 week later were used.

Improvement of the distal CMAPs after stimulation of the ulnar and median nerves was also taken into account. CMAP improvement after 1 week was established when the increase in the CMAP of the individual nerves at the second investigation was >1 mV compared with the first EMG, and the summed CMAP improved by >2 mV. The peroneal nerve was tested in only 52 patients, and recruitment pattern on maximal voluntary effort, an important prognostic factor in univariate analysis,19 was available for only 101 patients. These data were therefore excluded from the analyses.

Laboratory studies before onset of treatment.

All available pretreatment samples were retested by ELISA for the presence of IgA, IgM, and IgG antibodies to Campylobacter jejuni, using an acid-glycine extract from C. jejuni as antigen.20 The ELISA to detect IgA, IgM, and IgG antibodies to ganglioside GM1 was also optimized, and thin-layer chromatography was used for confirmation.20 IgM antibodies against cytomegalovirus (CMV) in the initial serum samples were determined by ELISA (Vidas; bioMésieux, Mercy-l’Etoile, France).

Statistical methods.

End points were the ability to walk independently at 8 weeks and at 6 months. Eight weeks was chosen as one end point because half of the 147 patients were able to walk independently at this point. The final clinical assessment was performed at 6 months. Patients unable to walk independently at this stage are likely to have considerable permanent clinical deficits.

The identification of prognostic factors for each combination of binary outcome (ability to walk independently at 8 weeks and 6 months) and set of potential prognostic factors was performed using the following strategy.

• 1. All continuous variables were analyzed with respect to representation (continuous or categorized). An appropriate coding system was devised by creating four categories of continuous variables based on the quartiles of their distribution. With univariate logistic regression on the three dummy variables derived from the four categories, the log odds ratios (ORs) were calculated. Plots of these log ORs versus the categories were used as an empirical assessment of a linear (or quadratic) relation of the log ORs with the factor of concern. Factors with three categories were recoded to two dummy variables.

• 2. Univariate associations for each factor separately were assessed with the chi-square test (a two-tailed test) without correction for continuity.

• 3. Forward and backward stepwise logistic regressions were applied with two different sets of entry and drop criteria: p-enter = 0.25, p-remove = 0.35 and p-enter = 0.15, p-remove = 0.20. Only factors with significant (p < 0.05, Wald test) coefficients were finally maintained.

• 4. For the selected factors, the interaction with treatment was incorporated into the model when significant (p < 0.05).

• 5. The overall fit of the model was studied using the Hosmer-Lemeshow test.

• 6. The model’s accuracy was determined using a receiver operating characteristic (ROC) analysis and by calculation of the area under the ROC curve. The heuristic shrinkage estimator of van Houwelingen and le Cessie21 was used to quantify overfitting.

• 7. For each selected model, the predicted probabilities for each covariate pattern (i.e., a combination of outcomes of prognostic factors) was calculated, with and without correction for overfitting (shrinkage).

The available data were tested in rank order according to the sequence in which they became available in the clinical setting. The clinical features were analyzed first, followed by the EMG and laboratory findings.

To illustrate the effect of prognostic factors, time to reach the stage of independent locomotion between the different groups was analyzed by the method of Kaplan and Meier and the log-rank test.

Univariate analysis of factors influencing outcome at 8 weeks and 6 months.

Table 1 summarizes the results of the univariate analyses of baseline characteristics. Exploring the continuous prognostic factors did not reveal a linear or quadratic relation for the log ORs; therefore, we preferred dichotomization. Cutoff points were based on literature or median values. Of all potential prognostic clinical factors evaluated, age ≥50 years, MRC sums core <40, a preceding gastrointestinal illness, and being bedbound or on the ventilator were associated with a significantly increased risk of being unable to walk independently 8 weeks or 6 months after the start of therapy. An upper respiratory tract infection was related to an improved outcome 8 weeks after the start of treatment, but did not remain significant at the 6-month end point. Rapid onset of weakness (≤4 days) before the start of treatment resulted in a poor outcome only at 6 months, and was not of significance at 8 weeks.

Table 1 also summarizes the univariate analysis of prognostic factors obtained by ancillary investigations done at randomization. A CMAP of the abductor digiti quinti of ≤3 mV was associated with a poorer outcome at both 8 weeks and 6 months; a CMAP of the abductor pollicis brevis of ≤3 mV was not a prognostic factor; and a value for the sum of the CMAPs of ≤6 mV was a relevant prognostic factor only at 8 weeks. Presence of anti-GM1 antibodies reached statistical significance at 6 months, and a recent C. jejuni infection was strongly associated with a poor outcome at that stage.

Table 2 gives details of the univariate analysis of the clinical and EMG data obtained during the first 2 weeks of follow-up (at nadir). An MRC sum score of ≤30 and need for artificial ventilation during follow-up were associated with decreased recovery. A low CMAP after distal stimulation of the ulnar nerve and median nerve and a low sum of these potentials obtained 1 week after randomization were all strongly related to a poorer recovery. Improvement of the amplitude by >1 mV between the first and second EMG indicated improved outcome (table 2).

Multivariate analysis of factors influencing outcome at 8 weeks and 6 months.

Both the MRC sum score and the F score are measurements of the severity of muscle weakness. We found that the MRC sum score was a more powerful predictor of outcome, and therefore chose it for the multivariate logistic regression analyses. Using all clinical factors known at the time treatment started, multivariate analysis revealed that age ≥50 years, a preceding gastrointestinal illness, an MRC sum score of <40, and absence of cranial nerve involvement remained independent factors indicating a poor prognosis at 8 weeks, whereas a preceding upper respiratory tract infection was related to an improved outcome. These five factors were put in a final model, and three of them remained independently significant: age ≥50 years, a preceding gastrointestinal illness, and MRC sum score of <40. These factors were then evaluated for treatment interactions, and it was found that the effect of diarrhea was more striking in the PE-treated group. Table 3 shows the ORs for the different prognostic factors. After inclusion of the laboratory data, one additional independent prognostic factor could be found; a recent CMV infection indicated a poor outcome at 8 weeks (table 3). The EMG data did not give rise to additional independent risk factors.

Clinical factors that remained of prognostic importance at the 6-month end point were older age (≥50 years), a low MRC sum score (<40), diarrhea, and a rapid onset of weakness until start of treatment (≤4 days; table 3).

These factors in relation to the proportion of patients who did not achieve independent locomotion are shown in the Kaplan-Meier curves (figure). The effect of diarrhea is shown for each treatment separately (figure, A and B).

• Download figure
• Open in new tab
• Download powerpoint

Figure. Kaplan-Meier curves indicating the proportion of patients who did not recover to independent locomotion (F = 2) during 181 days of follow-up, according to prognostic factors: (A) presence (n = 16) or absence (n = 57) of diarrhea for plasma exchange (PE) treatment, p (log-rank) = 0.0004; (B) presence (n = 11) or absence (n = 63) of diarrhea for intravenous immune globulin (IVIg) treatment, p (log-rank) = 0.19; (C) age <50 years (n = 74) or ≥50 years (n = 73), p (log-rank) = 0.003; (D) MRC sum score <40 (n = 73) or ≥40 (n = 74), p (log-rank) < 0.0001; (E) time of onset of weakness ≤4 days (n = 53) or >4 days (n = 60), p (log-rank) = 0.08.

Addition of the ancillary laboratory or EMG factors revealed no new independent prognostic factors. The calculated areas under the ROC curve for the four selected models range from 0.73 to 0.81. The four selected models all had a good fit (Hosmer-Lemeshow test p > 0.25).

Prediction of probabilities.

For each of the four models in table 3, the predicted probabilities of being able to walk independently (after 8 weeks or 6 months) can be calculated using the logistic model. Because of its limited value due to problems of overfitting (the number of potential prognostic factors is large in relation to the total number of patients; see, e.g., Harrell et al.22), we present only the results of the first model: three clinical factors as predictors for improvement at 8 weeks. The prognostic index for an individual patient can be calculated as follows: index = 0.30 − 1.04 × age (age < 50 years = 0, age ≥50 = 1) + 0.93 × MRC sum score (<40 = 0, ≥40 = 1) − 2.42 × diarrhea (no = 0, yes = 1) + 2.05 × diarrhea × treatment (no diarrhea = 0, treatment with PE = 0, diarrhea and treatment with IVIg = 1). For example, a patient treated with PE, 50 years of age or older, with an MRC sum score at entry of <40, and diarrhea, has a prognostic index of 0.30 − 1.04 × 1 + 0.93 × 0 − 2.42 × 1 + 2.05 × 0 = −3.16.

The probability of being able to walk independently after 8 weeks is calculated by 1/(1 + exp[−index]), or 4% in this example.

Overfitting is corrected by multiplying the prognostic index with the heuristic shrinkage factor,21 which was 0.43 for this model. This results in a corrected probability of 1/(1 + exp[−0.43 × (−3.16)]) = 20%.

Table 4 shows the predicted probabilities for each treatment for a patient with a poor prognosis and a patient with a good prognosis. Note that the overfitting-corrected probabilities move in the direction of the overall outcome.