Video-based finger kinematics for degenerative cervical myelopathy: a smartphone-based computer vision approach for screening.

purposeDegenerative cervical myelopathy (DCM) is a progressive spinal cord disorder that is frequently underdiagnosed, with diagnostic delays averaging 1 to 4 years. A key limitation in current clinical practice is the lack of objective and accessible screening tools. The 10-second grip-and-release test is commonly used to assess hand dysfunction in DCM, but its diagnostic performance is limited, particularly in older individuals with comorbid hand conditions such as osteoarthritis or peripheral neuropathy. To address this limitation, we developed and evaluated a smartphone-based computer vision tool that quantifies finger kinematics during the grip-and-release test. Our primary objective was to determine whether video-derived finger kinematics can provide superior diagnostic performance compared to grip count alone. A secondary objective was to assess how these video features correlate with cervical spinal cord compression on Magnetic Resonance Imaging (MRI).

methodsWe collected smartphone videos of 58 participants with DCM and 65 age-matched controls (including healthy individuals and those with non-DCM hand dysfunction) performing the 10-second grip-and-release test. Finger landmarks were extracted using MediaPipe, and 250 kinematic features per finger were computed and combined across both hands. Feature selection was performed using ANOVA (p < 0.05) and mutual information scores (> 0.01). A CatBoost classifier was trained on selected features using an 80/20 train-test split and five-fold cross-validation. A logistic regression model was trained using grip count alone. Model performance was evaluated using AUC, F1-score, sensitivity, and specificity.For the secondary analysis, we used linear regression models to evaluate associations between video-derived kinematics and cervical spinal cord compression, quantified on MRI, in 56 DCM participants.

resultsMean grip count was significantly lower in the DCM group (7.92 ± 3.27) compared to controls (10.26 ± 3.78; p < 0.001). The CatBoost model trained on 66 selected kinematic features achieved an AUC of 0.90, F1-score of 0.83, sensitivity of 83.3%, and specificity of 84.7%. The grip count-only model achieved lower performance (AUC 0.69, F1-score 0.67, sensitivity 75.0%, specificity 46.2%). Video-derived features were associated with MRI-derived measures of spinal cord compression including transverse diameter (R² = 0.43, p= 0.002), sagittal diameter (R² = 0.45, p = 0.001), compression ratio (R² = 0.42, p = 0.003), and maximum spinal cord compression ratio (R² = 0.36, p = 0.018).

conclusionWe demonstrated that a smartphone-based computer vision tool can accurately detect hand motor impairment specific to DCM. Finger kinematic analysis demonstrated significantly higher diagnostic accuracy than grip count alone and was associated with spinal cord compression on MRI. This approach offers a promising tool for early and scalable screening for DCM in both clinical and community settings.