MOSAIC: An Artificial Intelligence-Based Framework for Multimodal Analysis, Classification, and Personalized...

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Rare cancers represent over 20% of neoplasms and present significant unmet medical needs.
• Effective classification and prognostication are essential for improving patient outcomes and guiding treatment strategies.
• Myelodysplastic syndrome (MDS) serves as a model for rare hematologic cancers with complex clinical and genomic heterogeneity.

Purpose of the Study:

• To introduce MOSAIC, an artificial intelligence (AI)-based framework for multimodal analysis, classification, and personalized prognostic assessment in rare cancers.
• To clinically validate MOSAIC using myelodysplastic syndrome (MDS) as a representative rare cancer.
• To compare AI-driven approaches with conventional methods for improved patient stratification and survival prediction.

Main Methods:

• Analysis of 4,427 MDS patients using training and validation cohorts.
• Integration and imputation of clinical/genomic features via deep learning.
• Clustering using Uniform Manifold Approximation and Projection for Dimension Reduction + Hierarchical Density-Based Spatial Clustering of Applications with Noise (UMAP + HDBSCAN), compared to Hierarchical Dirichlet Process (HDP).
• Survival prediction using linear and AI-based nonlinear models, with Explainable AI (Shapley Additive Explanations [SHAP]) and federated learning for interpretation and performance enhancement.

Main Results:

• UMAP + HDBSCAN achieved more granular patient stratification with a higher average silhouette coefficient (0.16 vs. 0.01 for HDP) and improved Random Forest classification accuracy (92.7% ± 1.3%).
• AI survival prediction models, particularly Nonlinear Gradient Boosting Survival, outperformed conventional methods, showing high Concordance-Index (C-Index) in internal (0.77) and external (0.74) validation.
• SHAP analysis confirmed consistent feature importance across cohorts, and federated implementation enhanced model accuracy.

Conclusions:

• MOSAIC offers an explainable and robust AI framework for optimizing rare cancer classification and prognosis.
• AI approaches demonstrate superior accuracy in identifying genomic similarities and providing individual prognostic information.
• Federated implementation of MOSAIC facilitates broad clinical adoption while ensuring data privacy and high performance.