Mathematics in Prediction Modeling: Predicting Quality of Life of Patients Treated for Spinal Metastatic Disease.

Abstract

Background: Metastatic spine disease significantly impacts the quality of life (QoL) in oncological patients. As most of these patients are treated in a palliative setting, maintaining or enhancing QoL emerges as the primary therapeutic objective. QoL can be enhanced by improving neurological impairment, preserving mobility, and ameliorating pain. Metastatic spine disease is unique in the spine subspecialty because treatment plans must consider QoL benefits in the context of expected survival. Many models have been developed to predict survival, but only one model predicts QoL and has several limitations.

Objective: To develop a model that predicts patient-reported QoL three months after treatment in patients with metastatic spine disease receiving radiotherapy and/or surgery.

Methods: We included 548 patients undergoing radiotherapy (324; 59%), surgery (129; 24%), both (38; 7%), or no treatment (57; 10%) through retrospective chart review at a tertiary spine center in the Netherlands, using data from three existing prospective registries between January 2016 and December 2021. We assessed QoL using prospectively collected EQ-5D-3L questionnaires at baseline and follow-up three months post-treatment. From each EQ-5D-3L questionnaire, we calculated a health state index score ranging from less than 0 (where 0 is a health state equivalent to death; negative values are valued as worse than death) to 1 (perfect health). An improvement of 0.06 was considered the Minimal Clinical Important Difference (MCID) and was achieved in 253 (47%) patients. The 548 patients were divided into a training (80%) and testing cohort (20%), ensuring equal proportions of the MCID in each set. Four models - random forest, stochastic gradient boosting, penalized logistic regression, and support vector machine - predicted MCID. The variable selection, model building, and predictive performance assessment were conducted on the training set. The best-performing models in the training set were evaluated on the testing set for internal validation.

Results: The majority of patients were male (57%), with a median age of 67 years (interquartile range [IQR] 59 - 73). The three-month survival rate was 81%. On the training data, the penalized logistic regression algorithm had the best performance on the training data and achieved excellent calibration (intercept of 0.00, slope of 1.01), acceptable discrimination (area under the receiver operating characteristic curve [AUC]: 0.77 [95% confidence interval [CI] 0.60 – 0.88]), and a Brier score of 0.20 (null-model Brier score of 0.24). On the independent test set, the model achieved excellent calibration (intercept of 0.05, slope of 0.98), acceptable discrimination (AUC: 0.73 [95% CI 0.63 – 0.82]), and a Brier score of 0.21 (null-model Brier score of 0.24). On decision curve analysis, the model outperformed the treat-all and treat-none decision strategies between 0.1 and 0.6, and between 0.65 and 0.8. The penalized logistic regression algorithm was chosen as the final model considering performance and ease-of-use since it only uses the variables of the primary histology group by Katagiri et al., and the baseline EQ-5D-3L health state index.

Conclusion: We developed and internally validated a model predicting post-treatment estimation of improvement of QoL using prospectively collected institutional data. The algorithm shows promise and requires further external validation before implementation in clinical practice. Adding more laboratory values, such as albumin and lymphocytes, as possible predictors may improve algorithm performance. Integrating models that predict outcomes, such as survival and QoL, into electronic health records may guide clinical decision-making.