PREDICTING TIME-TO-PAIN PROGRESSION IN KNEE OSTEOARTHRITIS: A CLINICAL AND RADIOMICS FEATURE ANALYSIS

Knee osteoarthritis (KOA) is a prevalent joint disorder causing significant physical disability worldwide. Estimating time-to-pain progression could help identify high-risk patients and guide treatment decisions. While prior studies have explored binary progression prediction, predicting the specific year of pain onset within a multi-year timeframe remains underexplored. Pain is inherently subjective and multifactorial, and it is unclear whether baseline clinical assessments or structural imaging features carry sufficient signal to predict when pain progression will occur.

To investigate whether baseline clinical variables or MRI-derived radiomics features can predict the year of knee pain progression (Years 1-8) in KOA patients.

We utilized data from the Osteoarthritis Initiative (OAI) database. Subjects with a baseline WOMAC pain score increase of ≥9 (scaled 0-100) sustained for at least two years were classified as having knee pain progression. Only progression cases were included. From 1,224 baseline clinical variables in OAI, Lasso regularization optimized via Optuna selected 18 features. A self-supervised pretrained TabNet classifier was trained for 8-class time-to-progression prediction (Years 1-8) using masked feature reconstruction followed by supervised fine-tuning with cross-entropy loss. For comparison, 3D radiomics features extracted from bones and soft tissue segmentation masks of baseline DESS MRI were combined with Lasso feature selection and XGBoost classification (Figure 1). Data was split into training/validation/test sets (n=2,064/306/589). Performance was evaluated using accuracy within ±1 year and mean absolute error (MAE) in years.

The TabNet model using clinical variables achieved 57.8% accuracy within ±1 year and MAE of 1.72 years. Feature importance analysis identified KOOS Quality of Life, SF-12 Physical Summary Scale, WOMAC Pain Scores (for affected and contralateral knee), CES-D Depression Score, and knee pain severity as the most attended features. However, the model predominantly predicted early progression years (Year 1 recall=0.83) while failing to identify later progression time points (Years 5-8 recall≈0), indicating that it largely defaulted to predicting the majority class rather than learning meaningful temporal patterns. The radiomics-based approach (Lasso + XGBoost on DESS features) performed worse, achieving 48.2% accuracy within ±1 year. Feature selection yielded 20 radiomic features spanning six anatomical regions (femur, femur cartilage, tibia, tibia cartilage, patella cartilage, and meniscus), including first-order, texture, and shape features extracted using wavelet, logarithm, exponential, LBP-3D, and gradient filters. Neither approach demonstrated sufficient discriminative ability for fine-grained temporal prediction of pain progression.

Neither baseline clinical variables nor MRI-derived radiomics features could meaningfully predict the year of knee pain progression. These findings highlight the inherent difficulty of temporal pain prediction: pain is a subjective experience influenced by psychological, social, and neurological factors beyond what baseline clinical scores or structural joint features can capture. The severe class imbalance across progression years further compounds this challenge. These results suggest that predicting time-to-pain progression in KOA may require richer data, deep learning-based imaging features, or multimodal integration, and underscore that this remains a fundamentally difficult prediction task.