Predicting quality of life of patients after treatment for spinal metastatic disease: development and internal evaluation.

BACKGROUND CONTEXT: When treating spinal metastases in a palliative setting, maintaining or enhancing quality of life (QoL) is the primary therapeutic objective. Clinicians tailor their treatment strategy by weighing the QoL benefits against expected survival. To date, no available model exists that predicts QoL in patients after treatment for spinal metastases.

PURPOSE: To develop and internally evaluate a model predicting QoL for patients after treatment for spinal metastases, across the spectrum of (local) treatment modalities.

STUDY DESIGN/

SETTING

Cohort study of prospectively collected data.

PATIENT SAMPLE: Patients with spinal metastases referred to a single tertiary referral center in the Netherlands between January 1st, 2016, and December 31st, 2021.

OUTCOME MEASURES: The primary outcome was achieving a minimal clinically important difference (MCID) on QoL using the EQ-5D-3L index score 3 months after the referral visit (at the outpatient clinic or emergency department).

METHODS

Five prediction models using machine learning were developed: random forest, stochastic gradient boosting, support vector machine, penalized logistic regression, and neural network. Performance was assessed using cross-validation during development and bootstrapping for internal evaluation with discrimination (area under the curve (AUC)), calibration, and decision curve analysis. This study was funded by the AOSpine under the Discovery & Innovation award (AOS-DIA-22-012-TUM). A total amount of CHF 40,000 ($45,000) was received.

RESULTS

In total, 953 patients were included in the study, of which 308 (32%) achieved the MCID at 3 months. Discrimination was fair and comparable between the models, but the random forest model outperformed the other models on calibration and was therefore chosen as the final model (AUC 0.78; confidence interval (CI): 0.71 to 0.85; calibration intercept: -0.06; CI: -0.31 to 0.25; calibration slope: 1.05; CI: 0.70 to 1.44), with the following predictors ranked by importance: baseline EQ-5D-3L index score, Karnofsky Performance Scale, primary tumor histology, opioid use, and presence of brain metastases.

CONCLUSIONS

We developed and internally evaluated a random forest model that predicts clinically meaningful improvement of QoL 3 months after the baseline visit at the outpatient clinic for patients with spinal metastases. Future studies should externally evaluate the random forest model to confirm its robustness and generalizability in daily practice.