Partially Customized ML Models for Data De-identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated de-identification systems face challenges in achieving high performance on unstructured text data, such as clinical notes, due to variability in formatting and jargon, and require large labeled datasets for customization, which is costly and time-consuming.
Innovation Solution
A partially customized de-identification system is developed using pre-trained machine learning models that can be configured with a small set of labeled or unlabeled data specific to an organization, employing neural networks to generate embeddings and predict protected data for de-identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fully customized de-identification system is developed using large labeled datasets, then performance (recall and precision) is improved, but cost and time required for data preparation increase significantly
Solution Approach 1:
The system applies preliminary action by using pre-trained machine learning models that have already been trained on large datasets before deployment. This allows the system to achieve high performance without requiring organizations to spend time preparing large labeled datasets from scratch, thus resolving the contradiction between performance and data preparation time
Solution Approach 2:
The system uses copying by leveraging pre-trained models that have learned patterns from large datasets, effectively copying the knowledge and performance characteristics of models trained on extensive data without requiring the target organization to create their own large labeled datasets, thereby achieving high performance with minimal data preparation time
2Reliability
If a fully customized de-identification system is developed using large labeled datasets, then performance (recall and precision) is improved, but cost increases significantly
Solution Approach 1:
The system performs preliminary action by pre-training machine learning models on large datasets before deployment to target organizations. This eliminates the need for each organization to incur high costs of collecting, labeling, and preparing large datasets, thus achieving high performance at lower cost
Solution Approach 2:
The system applies copying by using pre-trained models that encapsulate knowledge from large datasets, allowing organizations to achieve high de-identification performance without bearing the substantial costs of creating their own large labeled datasets, thereby resolving the contradiction between performance and development cost
3Ease of operation
If traditional de-identification methods are used on unstructured text data, then processing is simpler, but recall and precision deteriorate due to variability in formatting and jargon
Solution Approach 1:
The system replaces mechanical system by substituting traditional rule-based de-identification methods with machine learning models. These models automatically learn to handle variability in formatting and jargon in unstructured text, achieving high recall and precision without requiring complex manual processing rules, thus resolving the contradiction between processing simplicity and identification accuracy
Data Source
AI summary
Apparatus and methods related to de-identifying data are provided. An example method includes receiving, by a computing device, input data comprising text. The method further includes applying a neural network to a tokenized representation of the input text, to generate an embedding based on contextual information associated with an entity. The method also includes predicting, by the neural network and based on the embedding, whether the input data comprises protected data in the text, wherein the neural network has been trained on a training dataset that has been partially customized based on the entity. The method further includes de-identifying the protected data in the text upon a determination that the input data comprises protected data in the text.