Aggregate Algorithm for Medical Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in developing a method for processing medical data, particularly medical image data, that allows for local training while ensuring privacy concerns are respected, given the limitations of existing technologies in assembling large and well-annotated datasets, privacy issues, and the need for validation of algorithms before release.
Innovation Solution
A method involving the training of multiple local algorithms on respective local systems using local data, with each algorithm's dataset transferred to an aggregating system to generate an aggregate algorithm, ensuring privacy by not transferring sensitive data and allowing for continuous improvement through user feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If local machine learning is performed on end user systems using local training data, then immediate improvement of the algorithm depending on user feedback is possible, but the pool of locally available training data is typically rather small, allowing only slow improvement of the algorithm
Solution Approach 1:
The patent combines multiple local algorithms from different end-user systems into a single aggregate algorithm. Each local algorithm is trained on small local datasets, but their combination creates a more robust model that benefits from the collective training data across multiple systems, thereby accelerating algorithm improvement despite limited individual data pools.
Solution Approach 2:
The patent creates copies of the algorithm at each end-user system, allowing each instance to be trained locally on available data. These local algorithm copies are then aggregated to form the improved algorithm, enabling parallel training across multiple systems and faster overall improvement.
2Reliability
If a sufficiently large and sufficiently well annotated set of training data is assembled, then the performance of the algorithm improves, but major privacy concerns arise since medical image data concerns highly private information
Solution Approach 1:
The patent extracts only the essential algorithmic parameters and performance characteristics from the training process, rather than transferring the actual sensitive medical data. The local algorithms are trained on private local data, and only the resulting model parameters are aggregated, thereby maintaining algorithm performance while protecting patient privacy.
Solution Approach 2:
The patent introduces an intermediate representation (the local algorithm models) that mediates between the need for large training datasets and privacy protection. Instead of directly sharing sensitive medical data, the system shares processed algorithmic models that have learned from the data without exposing the data itself.
3Quantity of substance
If extensive anonymisation protocol is implemented and access to large datasets is gained, then training data availability improves, but the process becomes elaborate and slow
Solution Approach 1:
The patent enables each end-user system to independently train its own local algorithm using its own local data without requiring external data collection or anonymization processes. This self-service approach eliminates the time-consuming data preparation and sharing protocols while still accumulating training data across multiple systems.
Solution Approach 2:
The patent segments the centralized data collection process into multiple independent local training processes. Instead of gathering all data centrally and then anonymizing it, the system divides the training task across multiple local systems, each processing its own data independently, thereby eliminating the elaborate centralized anonymization process.
4Manufacturing precision
If manual annotation of extensive dataset is performed by in-house experts or contractors, then data quality improves, but the process requires acquisition of relevant knowledge and increases complexity
Solution Approach 1:
The patent enables automatic local training at each end-user system without requiring manual annotation by experts. The local algorithms automatically learn from the available data, eliminating the need for complex manual annotation processes while still achieving good performance through the aggregation of multiple local models.
Data Source
AI summary
A method is for providing an aggregate algorithm for processing medical data. In an embodiment, a multitude of local algorithms are trained by machine learning. The training of each respective local algorithm is performed on a respective local system using respective local training data. A respective algorithm dataset concerning the respective local algorithm is transferred to an aggregating system that generates the aggregate algorithm based on the algorithm datasets.