Prognostic factors of a favorable outcome following a supervised exercise program for soldiers with sub-acute and chronic low back pain

Attempts to treat low back pain (LBP) have typically had limited success [1‐4]. These difficulties are believed to be due, in part, to the fact that the cause of the pain is rarely known, leaving approximately 85% of individuals with LBP to be included in a “non-specific LBP” group [4, 5]. However, LBP is a complex condition influenced by a range of psychosocial, physical and activity-related factors that differ from person to person [6‐9], rendering the non-specific LBP group heterogeneous, and unlikely to respond to the same treatment approach. It has thus been suggested that classification of LBP subjects into subgroups, followed by specific treatment of each subgroup according to their needs, is necessary in order to efficiently treat this condition [9]. Different classification systems, in which LBP subjects are assigned according to personal and clinical characteristics, have therefore been created and modified over the past 20 years [10, 11]. However, more conclusions on how to classify LBP sufferers and how to treat each subgroup are needed [12, 13]. Recent literature reviews still examine studies implementing interventions in non-specific, unclassified LBP populations. The most recent reviews evaluating the efficiency of exercise programs in the treatment of chronic LBP, for example, did not discuss LBP subgroups but treated the subjects as a homogenous population [4, 14]. Searle et al.’s review [4] confirmed the previously-held belief that stabilization and strengthening exercise programs decrease chronic LBP, but with only small effects. It remains to be examined whether division of the LBP population into subgroups will allow researchers and clinicians to obtain larger treatment effects when using exercise program interventions [2, 4, 15].

In order to accurately classify LBP patients into subgroups to which a specific intervention is assigned, valid clinical prediction rules (CPR) need to be developed [16, 17]. These rules are sets of data obtained from the history and physical evaluation of each patient, as well as from auto-administered questionnaires, that indicate which subgroup of patients will likely respond to a specific intervention. In 2011, a multi-station full-body supervised exercise program was created for soldiers with LBP and implemented at a military base of the Canadian Armed Forces. Although the program has enjoyed some clinical success, as subjectively reported by a number of soldiers, it is of great importance to maximize the efficacy of the program by determining the LBP subgroup most susceptible to benefit from it.

The development of CPR typically follows a three-step process: derivation, validation and impact analysis [16, 18, 19]. Derivation is the early hypothesis generation step in which prognostic factors are identified from a set of clinical variables that are believed to have predictive value, based on clinical experience or previous research. Kent et al. [20] recommended subdividing the derivation step into two sub-steps: hypothesis-setting, which represents the exploratory phase in which data from cohort studies may generate hypotheses about potential treatment effect modifiers, and hypothesis-testing, which is the confirmation sub-step in which pre-specified (a priori) hypotheses are tested more rigorously on subgrouping effects in samples of people similar to those who participated in the hypothesis-setting study. Subsequently, a validation step is required to identify treatment effect modifiers by testing subgroups/treatment interaction effects in randomized controlled trials involving samples of people with characteristics that are similar to (narrow validation) or different from (broad validation) those in the hypothesis-testing study. Lastly, the impact analysis aims to verify the usefulness of the CPR in improving outcomes and patient satisfaction and in decreasing costs once it is implemented in clinical practice.

According to this development framework, the present study is at the early hypothesis-setting step. Our objective was therefore to identify variables associated with a favorable outcome in soldiers with sub-acute and chronic LBP participating in a multi-station full-body supervised exercise program. The results obtained may permit generation of potential treatment effect modifiers that will eventually have to be validated before being recommended for clinical practice.

Of the 104 individuals that were recruited, nineteen (18.3%) did not complete the 6-week program or did not take part in the final evaluation. Therefore, a constant sample of 85 participants (81.7%) was included in all analyses. The reasons given by the individuals for dropping out of the study were: difficulties in coming to the Valcartier Health Centre due to military duties (n = 2), medical reasons (n = 1), new employment outside of the military (n = 2), prolonged absence for a military exercise or mission (n = 3) and personal or unknown reasons (n = 11). The individuals who dropped out had higher scores on the numerical pain rating scale evaluating worst and mean LBP perceived in the last 48 h when compared to the participants included in the final analyses (p < 0.05). No other differences were found. Subjects retained for analyses participated in 14.4 ± 2.6 sessions on average during the 6-week program (a mean of 2.34 sessions/week).

The mean baseline ODI score for the whole sample was 32.1 ± 12.1, considered as a moderate disability level [26]. Forty participants (47%) were categorized as having favorable outcome based on the ODI criterion, and 45 participants (53%) as unfavorable. The mean ODI change was 23.9 ± 11.0 for the group with favorable outcome and 3.4 ± 10.3 for the group with unfavorable outcome.

Tables 1, 2 and 3 present the clinical variables at baseline for the whole sample, as well as for both groups with favorable/unfavorable outcome. Notably, 74 % of the participants had previous episodes of LBP, and the time elapsed since the last episode varied from 1 to 200 months (mean 16.3 ± 32.2 months). Thirty-two percent (n = 27) had referred pain to the lower limbs. Finally, 51.8% of the participants (n = 44) had work restrictions.

The univariate tests identified seven potential prognostic factors with a KMO index of .692. As all MSA were above .600, no potential prognostic factors needed to be removed. The first multiple logistic regression indicated that two prognostic factors (work restrictions because of LBP and worst LBP perceived in the last 48 h) brought no unique information, with p-values respectively of .843 and .813, and were thus removed from the model. The second set of five factors with the group variable had a KMO index of .635. Although two prognostic factors had a MSA value very slightly below .6, all were kept in the model as these MSA were above .6 when the seven prognostic factors were used. [40] Furthermore, when the prognostic factors were dichotomized (see below), these MSA rose respectively to .60 and .62, and all five MSA were above or equal to .6. The model of the second multiple logistic regression had a highly significant (p = .00011) moderate capacity (Nagelkerke R² = .412) to predict the program outcome (favorable outcome: 86% correct; unfavorable outcome: 70% correct; global: 78% correct). This set of five variables included: (1) “no pain in lying down” [dichotomous], (2) “no use of antidepressants” [dichotomous], (3) “FABQ Work subscale” [continuous], (4) “number of treatments received before the first evaluation” [continuous], and (5) “no work restriction of 6 months or more” [dichotomous]. Individually, only variables 1 and 4 had p-values below .05 (respectively of .017 and .030). Since we were at the hypothesis-setting step of CPR development and since it is possible that beyond our sample these variables contain unique information that can predict the favorable or unfavorable outcomes, we decided to continue with the full set of five variables.

The FABQ Work subscale was dichotomized with the help of a recursive partitioning technique that indicated an optimal cutoff point of 22.5. Then, the potential dichotomized prognostic factor of a favorable outcome was set at a score below 22.5 on the FABQ Work subscale. The recursive partitioning technique did not successfully dichotomize the number of previous treatments. Therefore, this variable was dichotomized based on the threshold that led to the model yielding the best accuracy in predicting outcome. A criterion of 4 previous treatments or fewer was set as a potential dichotomized prognostic factor of favorable outcome.

The results of the third and final multiple logistic regression are shown in Tables 4 and 5. The model has a significant (p = .00004) moderate capacity (Nagelkerke R² = .370) to predict the program outcome (favorable outcome: 78% correct; unfavorable outcome: 80% correct; global: 79% correct). Furthermore, as the five variables had an odds ratio above 2.0 in this final regression, we decided to retain all variables in the final model. Even though the Nagelkerke R² may be lower, the outcome classification appears almost identical if not very slightly better. There was no sign of multicollinearity between variables included in the model. As seen in Table 4, only the first variable has a p-value below .05 (p = .017; Odds ratio = 3.7 [1.3–10.6]). It is also noteworthy that two variables have p-values of .061: (a) “no use of anti-depressants” (Odds ratio: 5.2 [0.9–29.4]), and (b) “FABQ Work below 22.5” (Odds ratio: 2.9 [0.9–8.6]).

In sum, for a patient or client presenting with all five criteria, the prognostic factors have low sensitivity and high specificity. The sensitivity improves to 0.78 for individuals presenting at least four of the five criteria. However, this increase in sensitivity comes at the expense of a decrease in specificity. Table 6 reports the prediction rate of a favorable outcome according to the number of criteria fulfilled. The satisfaction of four out of five criteria appears to be the best compromise between sensitivity and specificity.

This study identified five variables of a favorable outcome in patients with LBP who participated in a multi-station full-body supervised exercise program. Clinicians may anticipate a favorable outcome when their initial assessment identifies 4 or 5 variables included in our model. On the other hand, a less favorable outcome is expected if 3 or fewer variables are present, since the expected failure rate would be 80%. It is to be noted that these findings were established for a 50% improvement on the ODI and different variables would likely have been obtained had a different ODI threshold or other outcomes been used. Furthermore, the LR+ (3.9) and LR- (0.28) were relatively low [35], suggesting that the predictive capability of the model is limited. In comparison, Hicks et al. [15] found similar LR+ (LR+ 4.0), but higher LR- (LR- 0.18), with their preliminary CPR. Rabin et al. could not, however, validate this preliminary CPR in a subsequent study [13]. In contrast, Stolze et al. [23] reported a LR+ of 10.6 (95% confidence interval [CI]: 3.52, 32.14) in a CPR deviation study created to generate potential treatment effect modifiers of pilates exercises.

Three of the five variables included in our model had already been identified as prognostic factors of favorable outcomes for individuals with LBP in previous studies. We found that a score below 22.5 on the FABQ work subscale predicted a favorable outcome, which is in agreement with previous studies in which fears and beliefs about work were associated with poor recovery in patients with work-related LBP [41]. The FABQ work score was not identified as a predictor in CPRs developed by Hicks et al. [15] and Stolze et al.[23]. However, a score of less than 19 on the FABQ work subscale was considered a treatment effect modifier in the CPR developed by Flynn et al. [22], later validated by Childs et al., [42] whose aim was to identify patients likely to benefit from lumbar manipulation. As in the current study, participants in Flynn and Child’s studies were recruited from care facilities within the military, suggesting that fear-avoidance beliefs about military tasks may be of particular concern in soldiers with LBP. We also found that patients who do not use antidepressant medication were more likely to have a better outcome. This finding suggests that patients requiring antidepressants had less favorable outcomes, an observation which concords with the literature showing that patients with LBP in a depressive state are predisposed to longer recovery time and to the development of chronicity [43, 44]. A third prognostic factor of outcome in our model was “no work restriction of six months or more”. In the military context, a work restriction of fewer than six months should be interpreted as a temporary modification of regular duties for a non-chronic health condition, in contrast to a six-month medical category which is attributed to people for whom slow recovery or permanent health problems are anticipated. This finding is in agreement with previous studies that identified prolonged absence from work and the number of days of reduced activity as predictors of slower recovery in patients with LBP [41, 43]. “No pain when lying down” and “having had fewer than five physiotherapy treatments before the baseline evaluation”, that were both identified as prognostic factors in our model, have, to our knowledge, never been associated with the outcome of LBP. It is to be noted that the cutoff criteria for two of the prognostic factors (FABQ Work subscale and number of previous treatments) are actually tied to our sample and must be confirmed by future studies.

This single-arm study was the preliminary step of a process aiming to develop a CPR that can be used to identify patients most likely to benefit from the proposed exercise program. Three of the five variables identified in our final model are generally recognized as prognostic factors in people with LBP and only one of these has been validated as a treatment effect modifier within a CPR aiming to predict patients likely to benefit from lumbar manipulation. The methodological approach used in our study was deliberately liberal. To minimize the possibility of missing potential prognostic factors, we set the p value for retaining variables at 0.1 and, as suggested in hypothesis-setting studies, [20] we did not perform corrections for multiple comparisons. Investigating a high number of variables (more than 50) in a relatively small sample, as we did in the present study, increases the likelihood of finding significant associations by chance (type 1 error), which can be considered a limitation. However, only one prognostic factor was below the generally accepted level of significance of 0.05 and our final model has only a limited predictive capability. Thus, before continuing to the validation step, our results need to be confirmed in a hypothesis-testing study in which a limited number of a priori hypotheses will be tested and appropriate adjustments for multiple comparisons will be made. Any subsequent validation study should also be conducted with non-military groups (broad validation), as the present results may not be generalizable to the greater population. On the other hand, the well-standardized multivariate protocol and the acceptable dropout rate represent strengths of this study. Finally, the targeted sample size of 90 participants to be included in the statistical analyses was not met, as only 85 of the 104 participants took part in both evaluation sessions.

The present study established five variables to identify patients most likely to have a favorable outcome, regardless of their participation in the exercise program. Careful use of these variables is mandatory for clinical purposes as this study is at the early stage of CPR development. Future validation studies should be carried out with other populations to confirm this CPR and subsequently, to verify whether some of these factors may be considered treatment effect modifiers.