Validating the predictive precision of the dialogue support tool on Danish patient cohorts

BACKGROUND: Despite advances in surgical techniques and diagnostics, some patients remain unsatisfied with the result following spine surgery. One way to improve patient satisfaction may be found in better alignment of expectations. Prognostic tools might prove useful in strengthening surgeon-patient communication prior to surgery. The purpose of this study is to assess the predictive capabilities of the Swedish based Dialogue Support (DS) tool for spine surgery on a Danish population.

METHODS: The study included the diagnoses lumbar disc herniation, lumbar spinal stenosis, and lumbar degenerative disc disease. A total of 5.954 patients were retrieved from the Danish national spine registry (DaneSpine). For each group, 200 random cases with complete preoperative and 1 year follow-up data were selected. Two outcome measures were used: Global assessment of pain (GA pain) and satisfaction with outcome. Predictions were produced by manual entry in the DS application. Goodness of fit tests were used to compare the predicted distribution of proportions with successful outcomes (GA pain) to the actual distribution in the three samples. Binomial tests were performed to evaluate the predicted proportion of satisfied patients. Furthermore, ROC-curves, calibration plots, and metrics were calculated to assess the predictive performance.

RESULTS: ROC curves showed comparable AUC values with the values reported by the developing authors of the DS from 0.62 to 0.73 (GA pain) and 0.64 to 0.70 (satisfaction with outcome). The calibration plots, however, revealed a low degree of concordance. For GA pain sensitivity varied from 92.4% to 99.3%, and specificity from 1.5% to 13.4%. For satisfaction, sensitivity varied from 97.1% to 99.2% and specificity from 0.0% to 2.9%.

CONCLUSIONS: The predictive capabilities of the DS tool could not be generalized to the Danish sample cohorts. Further research on larger samples, provided full access to the underlying algorithms can be obtained, could produce a different result.