Model Deconstruction and Transfer Platform for Healthcare Prediction Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current healthcare prediction models face challenges in collecting, processing, and analyzing large datasets from multiple sources due to resource constraints and the logistical, technical, and financial burdens of protecting personal identifiers, limiting their utility and accessibility to entities with modest resources.
Innovation Solution
The method involves using a Model Deconstruction and Transfer (MDT) platform that allows participating healthcare centers to generate prediction models on-site without transferring personal identifiers, utilizing a Variable Library (VL) to select variables, and creating a Model Component Library (MCL) that can be used by third parties to build diverse prediction models for broader patient populations without de-identifying data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large datasets from multiple healthcare facilities are collected and processed to build meaningful prediction models, then the clinical utility and accuracy of prediction models are improved, but the resource expenditure (human and monetary) increases significantly
Solution Approach 1:
The patent segments the data collection and model building process by allowing each healthcare facility to generate its own local prediction model using its own data. These local models are then combined through a meta-modeling approach to create a comprehensive prediction model, eliminating the need to centrally collect and process all raw data from multiple facilities.
Solution Approach 2:
The patent introduces an intermediary meta-model that synthesizes predictions from multiple local models. This meta-model acts as a mediator that combines the strengths of individual facility models without requiring direct access to or transfer of sensitive patient data between facilities, reducing the resources needed for data management and security.
2Quantity of substance
If personal identifiers are transferred and processed from multiple healthcare facilities to build prediction models, then the comprehensiveness of the data is improved, but the administrative, logistical, and contractual hurdles increase
Solution Approach 1:
The patent extracts only the necessary predictive features and model parameters from each facility's data while leaving the personal identifiers and sensitive patient information at the source facilities. This extraction approach maintains data comprehensiveness for model building while eliminating the need to transfer or manage personal identifiers across facilities.
Solution Approach 2:
Instead of transferring original data with personal identifiers, the patent creates local copies of prediction models at each facility. These model copies can be shared and combined without involving the original sensitive data, thereby maintaining model comprehensiveness while avoiding the administrative burden of managing personal identifier transfers.
3Object-affected harmful factors
If personal identifiers are removed to protect privacy, then the liability risks are reduced, but the logistical, technical, and financial costs of producing de-identified data sets increase
Solution Approach 1:
The patent performs de-identification and model generation as preliminary actions at each individual facility before any data or models are shared. By completing the de-identification process locally and extracting only anonymized model parameters, the system reduces liability risks while minimizing the subsequent processing costs that would be required to handle sensitive data centrally.
4Measurement precision
If data is transferred beyond the physical and network boundaries of each healthcare facility, then the ability to build comprehensive prediction models is improved, but the security costs and liability protections increase
Solution Approach 1:
The patent uses local prediction models as intermediaries that reside at each facility. These model intermediaries enable comprehensive predictive analysis by combining insights from multiple facilities without requiring the transfer of sensitive patient data beyond facility boundaries, thereby maintaining model capability while reducing security costs.
Solution Approach 2:
The patent creates and shares copies of prediction models rather than transferring original sensitive data. These model copies can be distributed and combined to build comprehensive predictive capabilities while keeping the original patient data confined to secure facility boundaries, thus improving model capability without proportionally increasing security costs.
Data Source
Figure 1~2
Figure 3~4
AI summary
Provided is a method for generating prediction models from multiple healthcare centers. The method allows a third party to use data sets from multiple sources to build prediction models. By entering the data sets in a Model Deconstruction and Transfer (MDT) platform, a healthcare center may provide data to a third party without the need to de-identify data or to physically transfer any identifying or de-identified data from the healthcare center. The MDT platform includes a variable library, which allows the healthcare center to select variables that will be used to generate and validate the prediction model. Also provided is a method for compensating sources that contribute data sets based upon the percentage of clinical data that is used to generate a prediction model.