MicroRNA signature for early prediction of knee osteoarthritis structural progression using integrated machine and deep learning approaches

Objective: Conventional methodologies are ineffective in predicting the rapid progression of knee osteoarthritis (OA). MicroRNAs (miRNAs) show promise as biomarkers for patient stratification. We aimed to develop a miRNA prognosis model for identifying knee OA structural progressors/non-progressors using integrated machine/deep learning tools.

Methods: Baseline serum miRNAs from Osteoarthritis Initiative (OAI) participants were isolated and sequenced. Participants were categorized based on their likelihood of knee structural progression/non-progression using magnetic resonance imaging and X-ray data. For prediction model development, 152 OAI participants (91 progressors, 61 non-progressors) were used. MiRNA features were reduced through VarClusHi clustering. Key miRNAs and OA determinants (age, sex, body mass index, race) were identified using seven machine learning tools. The final prediction model was developed using advanced machine/deep learning techniques. Model performance was assessed with area under the curve (AUC) (95% confidence intervals) and accuracy. Monte Carlo cross-validation ensured robustness. Model validation used 30 OAI baseline plasma samples from an independent set of participants (14 progressors, 16 non-progressors).

Results: Feature clustering selected 107 miRNAs. Elastic Net was chosen for feature selection. An optimized prediction model based on an Artificial Neural Network comprising age and four miRNAs (hsa-miR-556-3p, hsa-miR-3157-5p, hsa-miR-200a-5p, hsa-miR-141-3p) exhibited excellent performance (AUC, 0.94 [0.89, 0.97]; accuracy, 0.84 [0.77, 0.89]). Model validation performance (AUC, 0.81 [0.63, 0.92]; accuracy, 0.83 [0.66, 0.93]) demonstrated the potential for generalization.

Conclusion: This study introduces a novel miRNA prognosis model for knee OA patients at risk of structural progression. It requires five baseline features, demonstrates excellent performance, is validated with an independent set, and holds promise for future personalized therapeutic monitoring.