Prompt-Based Multi-Modal EHR Generation for Privacy and Realism
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating synthetic electronic healthcare records (EHRs) are limited to single-modal data and result in unrealistic records due to low flexibility and are vulnerable to re-identification, despite efforts to de-identify and anonymize them.
Innovation Solution
The use of prompt-based learning with Bidirectional and Auto-Regressive Transformers (BART) for EHR generation, allowing for flexible, multi-modal EHR synthesis, evaluated by longitudinal and cross-modality perplexity measures, and incorporating adversarial approaches for privacy preservation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If de-identification and anonymization processes are applied to EHRs, then privacy protection is improved, but the data becomes vulnerable to re-identification attacks and loses realism
Solution Approach 1:
The patent generates synthetic copies of EHR data that replicate the statistical properties, patterns, and relationships of real EHRs without containing actual patient information. The synthetic data maintains longitudinal consistency and cross-modality correlations while being inherently anonymized, thus protecting privacy without sacrificing realism or vulnerability to re-identification attacks
2Ease of manufacture
If existing synthetic EHR generation methods are used, then data generation is simplified, but the generated records are unrealistic and limited to single-modal data
Solution Approach 1:
The patent employs a composite generative architecture combining variational autoencoders (VAEs) for latent space transformation and generative adversarial networks (GANs) for realistic data synthesis. This composite approach integrates multiple generation techniques to produce multi-modal synthetic EHRs that capture complex longitudinal patterns and cross-modality relationships, significantly improving realism while maintaining automated generation
Solution Approach 2:
The patent extends synthetic EHR generation from single-modal to multi-modal data by incorporating additional dimensions such as longitudinal temporal patterns, cross-modality correlations, and hierarchical data structures. This dimensional expansion enables the generation of realistic multi-modal EHRs that capture the complexity of real healthcare records
3Object-affected harmful factors
If strict de-identification processes are implemented, then privacy security is improved, but the sharing and research utility of EHRs deteriorates
Solution Approach 1:
The patent creates synthetic copies of EHR data that preserve all research utility including longitudinal patterns, cross-modality relationships, and statistical properties while being inherently privacy-protected. These synthetic copies enable unrestricted sharing and research applications without exposing real patient information, thus maintaining both security and versatility
Data Source
AI summary
An example embodiment may involve obtaining text-based, ground truth electronic health records (EHRs), wherein the ground truth EHRs specify a sequence of medical visits involving a plurality of modalities, and wherein each of the medical visits specifies tokens representing at least one of the modalities; generating a training data set by perturbing the ground truth EHRs, wherein perturbing the ground truth EHRs involves deleting or shuffling some of the tokens in the ground truth EHRs; and iteratively applying a machine learning trainer application to the training data set, wherein the machine learning trainer application includes: (i) a bidirectional language model encoder that takes EHRs within the training data set and produces vector embeddings therefrom, (ii) an autoregressive language model decoder that takes the vector embeddings and infers predicted EHRs therefrom, and (iii) a loss function that compares the predicted EHRs to their corresponding ground truth EHRs.