Artificial Intelligence (AI)

An Artificial Intelligence (AI)-driven system will be developed for OA diagnosis and monitoring, utilizing features from diverse data domains, including demographic (e.g., age, gender, body-mass index), clinical (e.g., pain level, Kellgren-Lawrence grade, joint stiffness), imaging (e.g., joint space width, cartilage thickness, synovitis) biomechanics (e.g., smoothness, symmetric balance, impulse and range of motion), and omics (e.g., chemical concentrations of proteins and metabolites of the synovial fluid, blood and serum) information. The input data will consist of various feature types, including numerical, categorical, ordinal, and continuous variables.

A data harmonization processwill be employed to ensure data homogeneity across domains. Data will be standardized at the feature level, applying a unified naming convention, unit normalization, and consistent feature encoding. Missing values will be handled using multiple imputation methods such as k-nearest neighbors and multiple imputation by chained equations. Outliers will be detected and corrected using domain knowledge-driven thresholds and multi-variate Machine Learning (ML)-based anomaly detection.

A metadata document will be created to ensure data consistency, interoperability, and integration across clinical, imaging, biomechanics, and omics registries. It will serve as a structured data dictionary, including each feature’s name, description, alternative names, data type, unit of measurement, format, domain, and relevant standards (e.g., SNOMED CT, ICD-10, DICOM, GA4GH). Additionally, it will specify handling procedures for missing values, feature harmonization rules, outlier detection, privacy considerations, and versioning. The metadata will be maintained in a structured JSON/SQL format and will guide data preprocessing, ensuring AI models receive standardized and validated multi-modal inputs.

Feature fusion strategies will be employed to integrate information from multiple modalities. Early fusion will combine all features at the input level, enabling ML models to learn cross-domain relationships, while late fusion will aggregate predictions from separate domain-specific ML models to produce a final decision (e.g., using stacking or weighting of outputs from individual models).

For ML model implementation, eXtreme Gradient Boosting (XGBoost) models will be designed, optimized and validated. XGBoost can efficiently handle multi-modal data, including numerical, categorical, ordinal, and continuous features. It provides high predictive performance by capturing non-linear relationships and feature interactions, and offers good interpretability through feature importance, crucial for clinical decision support. Additionally, it is robust, scalable, and prevents overfitting, while its flexibility with hyperparameters allows optimization for both accuracy and generalization, ensuring reliable results for patient stratification and disease monitoring.

The validation process will follow a multi-tiered strategy that incorporates internal cross-validation, external testing on independent datasets, and performance assessment across diverse patient populations. Internally, models will undergo k-fold cross-validation (5-fold and 10-fold) to evaluate robustness. Performance metrics such as AUC-ROC, F1-score, sensitivity, specificity, Gini index, and calibration and precision-recall curves will be used to assess effectiveness. External validation will be conducted on datasets from multiple clinical centres across different regions and countries, ensuring that models perform consistently across different demographic groups.

To mitigate bias, subgroup analysis will be conducted based on age, gender, and ethnicity. Adversarial debiasing models will be applied to identify and quantify biases, ensuring that predictions are not skewed by confounding variables. Synthetic data augmentation techniques, including Tabular Generative Adversarial Networks (TGANs) and Synthetic Minority Over-Sampling Technique (SMOTE), will be employed to balance datasets and improve model fairness.

Fairness will be measured using demographic parity, ensuring that predictions are equally distributed across groups, and equal opportunity, which guarantees similar true positive rates among different subgroups. Counterfactual fairness will ensure that model predictions remain consistent when sensitive attributes are altered.

To enhance explainability and interpretability, the model will incorporate methods such as Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), which will provide insights into how individual features contribute to the model’s predictions, ensuring compliance with regulatory standards. Additionally, the model will be continuously tested for bias control using fairness metrics and counterfactual fairness checks, ensuring that the AI system delivers equitable and reliable predictions.

The integration of multi-source data, rigorous bias mitigation, and robust validation will ensure that the AI platform delivers fair, explainable, and clinically actionable insights, aiding OA diagnosis and monitoring.

The federated learning framework will enable AI model training across multiple institutions without centralizing raw patient data. Each institution will train a local model, and only model updates (weights/gradients) will be shared. To preserve privacy, secure aggregation techniques like homomorphic encryption will ensure that the server only receives aggregated updates, while differential privacy will add noise to local updates, preventing data leakage. The model updates will be aggregated using federated averaging to create a global model, ensuring collaboration while maintaining patient confidentiality.

The Academic Computing Center HPC@Polito infrastructure will support data harmonization, AI model training, validation, and federated learning development. It features three integrated InfiniBand clusters—CASPER (16 nodes, 32 AMD Bulldozer cores, 128GB RAM each), HACTAR (29 nodes, 24 Intel XEON v3 cores, 128GB RAM each), and LEGION (57 nodes, 32 Intel Scalable Processors Gold cores, 384GB RAM each). The center provides 228 TB of total storage, with users allocated up to 1TB per storage system, ensuring scalable and secure computational resources for AI development.