A fuzzy-intelligence framework integrating tissue and liquid biopsy methylation profiling to empower early-risk stratification in osteosarcoma

Osteosarcoma is a highly aggressive malignant bone tumor with poor survival outcomes, particularly in patients with metastatic or recurrent disease. Current prognostic assessment primarily relies on clinicopathological characteristics and imaging findings, which lack sufficient molecular information for individualized risk stratification. Aberrant DNA methylation plays an important role in osteosarcoma development and can be detected noninvasively through circulating tumor DNA (ctDNA). However, constructing reliable and interpretable predictive models remains challenging due to the high dimensionality, technical variability, and biological heterogeneity of methylation sequencing data. To address these challenges, we developed Fuzzy-OS, a prognostic framework that integrates methylation profiles from tumor tissue and ctDNA using a fuzzy deep learning approach with enhanced interpretability. DNA methylation profiles were retrospectively analyzed in two independent patient cohorts from a single center: tumor tissue samples ( n = 89, Illumina MethylationEPIC array) and pre-treatment plasma ctDNA samples ( n = 11, whole-genome bisulfite sequencing). Because the two cohorts shared no patients (no paired tissue–ctDNA samples), integration was performed at the feature level on the 189,177 CpG sites common to both platforms. A Dirichlet distribution–based multi-view Takagi–Sugeno–Kang (TSK) fuzzy system was constructed to handle high-dimensional data while preserving interpretability, incorporating variational inference for uncertainty estimation, tensor factorization for dimensionality reduction, and attention-based fusion of multi-view information. Model performance was evaluated using 30 repeated, leakage-free stratified train–test splits and compared with clinical expert assessments and four conventional machine learning models. On 300 held-out test predictions accumulated over 30 repeated splits, Fuzzy-OS achieved a pooled prognostic accuracy of 0.63 (95% CI 0.57–0.68), comparable to the tissue-only TSK model (0.65; 95% CI 0.60–0.69; P = 0.66) and significantly better than clinical expert evaluation (0.46 ± 0.08) and four conventional machine learning models (0.40–0.42; P < 0.0001). In one illustrative retrospective case, Fuzzy-OS assigned a high-risk label more than 10 months before radiological confirmation of metastasis. This preliminary observation does not establish prospective clinical utility and requires external validation. Fuzzy-OS provides an interpretable framework for integrating tissue and liquid biopsy methylation data for osteosarcoma risk stratification. The proposed approach demonstrates the potential of combining multi-source epigenetic data with fuzzy learning methods to improve molecular prognostic assessment in osteosarcoma.