Elastic Shape Analysis for Synthetic ECG Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ECG analysis using AI faces challenges in accurately classifying arrhythmias due to insufficient annotated data, particularly for certain heart conditions, as simple geometric transformations can create unrealistic ECG waveforms and fail to improve inference accuracy beyond a certain point.
Innovation Solution
The method employs Elastic Shape Analysis (ESA) to transform and interpolate ECG waveforms, maintaining key element waveforms while adjusting time axes to generate natural, annotated training data, thereby increasing the quantity and quality of training data for AI models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If simple geometric transformations are applied to ECG waveforms, then the quantity of training data increases, but the realism and quality of generated waveforms deteriorates
Solution Approach 1:
The patent applies elastic deformation transformations that modify waveform parameters (time axis, amplitude) in a controlled manner while preserving the physiological characteristics of ECG waveforms. This allows generating diverse training data through parameter variations without creating unrealistic waveforms, directly resolving the contradiction between data quantity and waveform quality
2Measurement precision
If more annotated data is collected for rare heart conditions, then the accuracy of AI classification improves, but the time and cost for data collection increases
Solution Approach 1:
The patent performs preliminary data preparation by generating synthetic annotated ECG data through elastic transformations before AI model training. This pre-generated data serves as additional training samples, reducing the need for extensive manual data collection and annotation while improving classification accuracy for rare conditions
Solution Approach 2:
The patent creates copies of existing annotated ECG waveforms through elastic deformation transformations, generating synthetic data that preserves the essential characteristics of rare heart conditions. This copying approach with transformation generates diverse training samples without requiring additional patient data collection
3Measurement precision
If AI models are trained with more diverse training data, then the inference accuracy for boundary cases improves, but the complexity of data processing increases
Solution Approach 1:
The patent systematically varies waveform parameters through elastic transformations to generate diverse training data that specifically targets boundary cases between normal and abnormal heart conditions. This parameter-based generation approach provides controlled diversity without requiring complex data processing pipelines
Data Source
AI summary
A data generation method by a computer is disclosed. First waveform data including marking information at a first position on a waveform, and acquiring second waveform data are acquired. A transformation function is specified that transforms the first waveform data to reduce the difference between a first value of a time axis for a first characteristic point in the first waveform data and a second value of the time axis for a second characteristic point, in the second waveform data, corresponding to the first characteristic point. Third waveform data are generated, in which the marking information is applied at a second position corresponding to the first position in the first waveform data, the second position being determined by using the transformation function.