AUTHOR=Mekhael Elio , El Rachkidi Rami , Saliby Renee Maria , Nassim Nabil , Semaan Karl , Massaad Abir , Karam Mohamad , Saade Maria , Ayoub Elma , Rteil Ali , Jaber Elena , Chaaya Celine , Abi Nahed Julien , Ghanem Ismat , Assi Ayman TITLE=Functional assessment using 3D movement analysis can better predict health-related quality of life outcomes in patients with adult spinal deformity: a machine learning approach JOURNAL=Frontiers in Surgery VOLUME=10 YEAR=2023 URL=https://www.frontiersin.org/journals/surgery/articles/10.3389/fsurg.2023.1166734 DOI=10.3389/fsurg.2023.1166734 ISSN=2296-875X ABSTRACT=Introduction

Adult spinal deformity (ASD) is classically evaluated by health-related quality of life (HRQoL) questionnaires and static radiographic spino-pelvic and global alignment parameters. Recently, 3D movement analysis (3DMA) was used for functional assessment of ASD to objectively quantify patient's independence during daily life activities. The aim of this study was to determine the role of both static and functional assessments in the prediction of HRQoL outcomes using machine learning methods.

Methods

ASD patients and controls underwent full-body biplanar low-dose x-rays with 3D reconstruction of skeletal segment as well as 3DMA of gait and filled HRQoL questionnaires: SF-36 physical and mental components (PCS&MCS), Oswestry Disability Index (ODI), Beck's Depression Inventory (BDI), and visual analog scale (VAS) for pain. A random forest machine learning (ML) model was used to predict HRQoL outcomes based on three simulations: (1) radiographic, (2) kinematic, (3) both radiographic and kinematic parameters. Accuracy of prediction and RMSE of the model were evaluated using 10-fold cross validation in each simulation and compared between simulations. The model was also used to investigate the possibility of predicting HRQoL outcomes in ASD after treatment.

Results

In total, 173 primary ASD and 57 controls were enrolled; 30 ASD were followed-up after surgical or medical treatment. The first ML simulation had a median accuracy of 83.4%. The second simulation had a median accuracy of 84.7%. The third simulation had a median accuracy of 87%. Simulations 2 and 3 had comparable accuracies of prediction for all HRQoL outcomes and higher predictions compared to Simulation 1 (i.e., accuracy for PCS = 85 ± 5 vs. 88.4 ± 4 and 89.7% ± 4%, for MCS = 83.7 ± 8.3 vs. 86.3 ± 5.6 and 87.7% ± 6.8% for simulations 1, 2 and 3 resp., p < 0.05). Similar results were reported when the 3 simulations were tested on ASD after treatment.

Discussion

This study showed that kinematic parameters can better predict HRQoL outcomes than stand-alone classical radiographic parameters, not only for physical but also for mental scores. Moreover, 3DMA was shown to be a good predictive of HRQoL outcomes for ASD follow-up after medical or surgical treatment. Thus, the assessment of ASD patients should no longer rely on radiographs alone but on movement analysis as well.