Model Performance
Predictive models provide individualized risk-adjusted postoperative outcome estimations, which in the absence of nonsurgical outcome data have tremendous potential to assist providers during the preoperative assessment of patients and to improve patient engagement in shared decision-making concerning treatment planning. Using a prospective, multicenter, longitudinal registry, we developed 4 predictive models for 12-month PROs after elective surgery for degenerative lumbar spine pathology. In descending order of importance, the most important predictors of overall disability, QOL, and pain outcomes following lumbar spine surgery were patient employment status, baseline NRS-BP scores, psychological distress, baseline ODI scores, level of education, workers’ compensation status, symptom duration, race, baseline NRS-LP scores, ASA score, age, predominant symptom, smoking status, and insurance status.
Predictors of OutcomeOccupation-related factors, including employment status at the time of surgery and type of occupation, were the most important predictors of the overall outcomes. Patients not actively working (disability claim or retired) had significantly lower odds of having better outcomes at 12 months postoperatively than those who were actively employed. Furthermore, patients employed in medium- and heavy-labor jobs had lower odds of achieving better outcomes than those with a sedentary-type job. While these findings potentially can be explained in part by the presence of more severe back pathologies in patients with more laborious jobs and in those receiving disability funds, these factors cannot fully account for such findings. As such, the association between occupation and outcomes is likely multifactorial. Previous sociological models have revealed that patients’ capacity for recovery and return to function also depends on psychological state, job satisfaction, physical demands at work, income, and social support.
5,20,31,47,52
A patient’s baseline functional status was also found to be an important predictor of overall outcomes. The impact of baseline functional status on postoperative outcomes following lumbar spine surgery is a consistent finding in the literature.11,12,19,28,36–38,44,46,48,65 Analogous to our results, numerous previous studies have demonstrated that the patients whose baseline disability, pain, and QOL scores are, on average, worse than those of others are more likely to have poorer outcomes at 12 months.3,14,48 Conversely, we found that patients with higher baseline NRS-LP scores had greater odds of having better outcomes. Pearson et al. reported that surgery in patients with higher preoperative LP scores yielded significantly better pain relief than in those with higher preoperative BP scores.58 This may be because the patients with higher LP scores often have associated radiculopathy or neurogenic claudication, which is typically more responsive to surgical decompression than BP. In a cross-sectional study assessing associations between patients’ expectations for lumbar spine surgery and baseline characteristics, Mancuso and colleagues demonstrated that patients with greater preoperative disability—for whom the literature consistently demonstrates a lower odds of achieving better outcomes—have the greatest expectations for postoperative functional recovery.45 This finding highlights the importance of adjusting for varying degrees of baseline disability, pain, and QOL status when assessing 12-month outcomes. This also underscores the importance of shared preoperative decision-making, facilitated with evidence-based decision support tools, such as the predictive models presented here.
Patients’ preoperative psychological distress was an important predictor of overall outcomes and it was the one of the most important predictors of postoperative 12-month QOL (EQ-5D). A number of previous studies have reported that preoperative psychological distress is associated with worse outcomes following lumbar spine surgery.2,4,22,35,39,56,63 Sinikallio et al., in a prospective analysis of 96 patients, demonstrated that the patients with preoperative depression and those who had continuous depression postoperatively experienced poor outcomes. At the 2-year follow-up evaluation, the patients who recovered from depression demonstrated postoperative improvement similar to the patients who had normal mood preoperatively and postoperatively.60 In a recent randomized control trial, Archer et al. demonstrated that incorporating targeted cognitive behavioral therapy in postoperative care results in improved outcome 3 months after lumbar spine surgery.6 This suggests that medical and behavioral interventions for concomitant psychological disease during the preoperative and postoperative periods may help to improve outcomes in patients with psychological distress. We observed that a higher level of education was an important predictor for better overall PROs. Authors of previous studies have also indicated that higher education has a positive effect on patient outcomes.40,55,61 Further support of the importance of education on clinical outcomes is the observation of some authors that a correlation exists between lower levels of education and treatment noncompliance, and other health-compromising behaviors.55
With respect to surgical procedures, the surgery-specific factors analyzed in the present study, including fusion, surgical approach, and number of vertebral levels involved, was observed to have a lower overall predictive importance compared with other variables (and, therefore, a lower impact on predicting PROs). These findings reinforce numerous and consistent previous observations that patient-specific factors are primary drivers of outcomes following spine surgery.1,16,19,38,48,52
Based on our data, we found that other baseline patient-specific factors—workers’ compensation, symptom duration, race, smoking status, preoperative comorbidities, and insurance status—also strongly influence outcomes following lumbar surgery, a finding consistent with those reported in a number of studies.3,11,13,15,18,19,48
Application to Real-World PracticeThe 4 predictive models described provide individualized risk-adjusted postoperative outcome projections. The cases presented represent hypothetical case scenarios comparing patient characteristics that predict better or worse outcomes after surgical therapy for lumbar degenerative diseases. Although this information may potentially be used to tailor the application of invasive therapies (particularly to help avoid care that is highly likely to be ineffective) the authors caution that restriction of surgical therapy based solely on patient characteristics such as race, socioeconomic status, and/or age is inappropriate, potentially discriminatory, and is not the intended use of these models. In that regard, we would like to emphasize that no single variable is likely to influence ultimate outcomes, that all variables have differential effects in individual patients, and that all of the predictors used in this model have an additive effect on predicting surgical outcomes. The predictive models presented here are primarily intended to engage patients in shared decision-making and facilitate true patient-centered care. Furthermore, the decision to offer surgery to any given patient requires consideration of numerous factors including, but certainly not restricted to, likely longer-term (e.g., 12-month) outcomes. As long-term expected clinical outcomes are only one element of the surgical decision process, results of this model would not necessarily deter clinicians from operating on patients with negative predictors of outcome, such as low baseline function.
Probabilistic discussions regarding postoperative outcomes during preoperative assessment can assist providers in setting realistic presurgical expectations for patients and families and help improve patient satisfaction with care. Clinicians can use these tools to adjust modifiable patient characteristics preoperatively to help modify postoperative outcomes. They can also use predictive models to help identify patient populations to recognize potentially ineffective care before it is given, thus facilitating greater surgical effectiveness and increasing the overall value (cost-benefit ratio) of spine surgery. The tools can also be used to provide meaningful comparisons of performance between service providers, by allowing for the generation of risk-stratified benchmarks for care. By comparing actual and expected (i.e., risk-adjusted) outcomes, providers who typically take care of the sickest and highest-risk patients (e.g., those in tertiary care centers) can do so with diminished concern of financial penalty in an increasingly value-based reimbursement environment.
As predictive models can be challenging to apply in “real-world” clinic settings, we have created a user-friendly, online application that can be more easily used in the spine clinic to predict PROs (http://statcomp2.vanderbilt.edu:37212/app_0/). By enabling patient-specific probabilistic counseling at the point of care, we seek to facilitate involvement of all stakeholders in true, shared decision-making. This activity holds promise with respect to improving patient outcomes and increasing health care savings.
Study Limitations and StrengthsThe limitations and weaknesses inherent in the current study have implications for the interpretation of its findings. Without controls, outcomes of patients can be compared but not outcomes of competing therapies. An intrinsic limitation of predictive models is the discrete number of variables that are inputted into its creation. In the present analysis, we included 30 patient-specific and surgery-specific variables collected as a part of the multicenter QOD registry. It is possible that variables not collected and therefore not accounted for in this analysis will play a significant role in a patient’s disability, QOL, and pain status after surgery. Such missing data would affect the performance of the models in discriminating accurately between observed and predicted outcomes. With exponentially increasing numbers of patients enrolled in the QOD registry, and with further refinement of variables collected, we will be able to update the predictive model and thereby increase its performance and accuracy.
Nonetheless, our risk-adjusted predictive models provide value as a starting point for patient-level assessment to guide shared decision-making and optimize outcomes at the individual patient and population levels. The c-index for our models was in the range of 0.64–0.69, which reflects a “good” discrimination index.32,33 Finally, the patients included in the QOD registry are enrolled from 74 centers across the US representing all practice types, i.e., academic, community, small, large, rural, and urban.7,29,53 The diverse mix of practices included and patients enrolled generate a representative sample of patients undergoing elective lumbar spine surgery, allowing these results to be generalizable (i.e., applicable) to most spinal surgery practices.
