Decentralized Clinical Data Evaluation via Local Model Aggregation

The system segments the clinical data evaluation process into decentralized local model training at each medical entity and centralized aggregate evaluation. Each entity maintains local datasets and trains local models independently, then shares only model parameters而非raw data. This segmentation enables data quality improvement through diverse data sources while preserving anonymity by never exposing the actual clinical records or their sources.

The patent introduces an intermediary evaluation system that receives local models from multiple medical entities, aggregates them, and identifies outlier datasets without revealing which specific entity contributed which data. The evaluation system acts as a mediator that assesses data quality through model performance metrics while maintaining the anonymity chain, allowing high-quality data identification without compromising source privacy.

The system implements a universal evaluation framework that can handle multiple types of clinical datasets from different medical entities through a common process. The evaluation system performs multiple functions: it aggregates diverse local models, evaluates their performance, identifies outliers, and provides feedback - all through a single decentralized architecture. This multi-functionality improves research efficiency across different data types while managing system complexity through a unified approach.

Each medical entity independently trains its own local models and contributes them to the aggregate evaluation without requiring direct coordination or manual intervention for data sharing. The system enables self-service by allowing entities to autonomously participate in the decentralized evaluation process, automatically receiving feedback on their data quality. This reduces coordination overhead while maintaining research efficiency through autonomous participation.

The system performs preliminary local model training at each medical entity before the aggregate evaluation phase. By pre-training local models on local datasets, the system preserves privacy (since raw data never leaves the entity) while reducing the time required for centralized processing. The preliminary action of local training allows the centralized system to work with compact model parameters rather than voluminous raw data, significantly reducing evaluation time.

The patent implements partial action by having each medical entity train only a local model on its own data rather than requiring full centralized training on all data. This partial training approach balances privacy preservation (local processing) with efficiency (smaller model sizes for aggregation). The excessive action component involves evaluating all local models against the aggregate, ensuring thorough privacy protection while the system optimizes the overall time investment through efficient aggregation algorithms.