Exploring Associations of Preoperative Physical Performance With Postoperative Outcomes After Lumbar Spinal Fusion: A Machine Learning Approach

Abstract

To explore the association between preoperative physical performance with short-and long-term postoperative outcomes in patients undergoing lumbar spinal fusion (LSF). Design: Retrospective cohort. Setting: University hospital. Participants: Seventy-seven patients (N=77) undergoing elective LSF were preoperatively screened on patient demographics, patient-reported outcome measures, and physical performance measures (movement control, back muscle endurance strength and extensor strength, aerobic capacity, flexibility). Interventions: Not applicable. Main Outcome Measures: Associations between preoperative variables and inpatient functional recovery, hospital length of stay (LOS), and 1-to 2-year postoperative pain reduction were explored using random forest analyses assessing the relative influence of the variable on the outcome. Results: Aerobic capacity was associated with fast functional recovery <4 days and prolonged functional recovery >5 days (median z scores=7.1 and 12.0). Flexibility (median z score=4.3) and back muscle endurance strength (median z score=7.8) were associated with fast functional recovery <4 days. Maximum back extensor strength was associated with prolonged functional recovery >5 days (median z score=8.6). Flexibility (median z score=5.1) and back muscle endurance strength (median z score=13.5) were associated with short LOS <5 days. Aerobic capacity (median z score=8.7) was associated with prolonged LOS >7 days. Maximum back extensor strength (median z score=3.8) was associated with 1-to 2-year postoperative pain reduction and aerobic capacity (median z score=2.8) was tentative. Conclusions: Physical performance measures were associated with both short-and long-term outcomes after LSF. Adding these measures to prediction models predicting outcomes after LSF may increase their accuracy.

Predicting patient-reported outcomes following lumbar spine surgery: development and external validation of multivariable prediction models

“…Thus, the missingness of the smoking status could potentially be related to other unmeasured variables. However, the significant prognostic value of smoking is consistent with several previous prediction models for spinal surgery outcomes (Janssen et al, 2021; Khor et al, 2018; McGirt et al, 2017; Staub et al, 2020).…”

“…However, the significant prognostic value of smoking is consistent with several previous prediction models for spinal surgery outcomes [14,[16][17][18].…”

“…These metrics are comparable or better than in similar externally validated models predicting painrelated outcomes (AUC 0.52-0.83, 6-19% [17][18][19]). COMI models showed poorer discrimination (AUC 0.63, 16% explained variance), although comparable with external validity of another model relying on the same measure (17% [18]), suggesting that composite outcomes like COMI may be more difficult to predict than, for instance, specific disability measures (AUC 0.71 [19]).…”

“…Similar studies relying on internal validation generally reported better discrimination (AUC 0.64-0.84; 23-49%; Abbott, 2011; Cobo Soriano et al, 2010; Khor et al, 2018; McGirt et al, 2015, 2017), highlighting a potential degree of over-optimism when model performance is only assessed on resampled or randomly-split data. Furthermore, model calibration was not always assessed (Abbott et al, 2011; Janssen et al, 2021; McGirt et al, 2017), but accurate prediction of outcomes can be particularly problematic in external data (Quddusi et al, 2020; Staub et al, 2020). Our models showed good calibration in the validation data across all outcomes, although there was a mild tendency to underestimate back pain MCID and leg pain reduction.…”

“…However, performance of the existing prediction models has been rarely quantified in patient data other than that used to develop them. In the few available examples, external validation revealed poorer discrimination ability in the new data (Janssen et al, 2021; Staub et al, 2020), or poor calibration leading to under- and overestimation of outcome probabilities (Quddusi et al, 2020).…”