Blood Pressure Data Splitting for Machine Learning Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data splitting methods for machine learning models, particularly in blood pressure estimation, fail to maintain data independence and distribution balance across training, validation, and test sets due to non-independent data points from the same subject and skewed distributions of systolic and diastolic blood pressure data.
Innovation Solution
A data splitting method that divides the measurement range of systolic and diastolic blood pressure data into intervals, generates classes based on these intervals, determines match conditions for each subject, and distributes blood pressure data into subsets to maintain similarity and balance across sets, ensuring that data from the same subject remains together and distributions are consistent across training, validation, and test sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If random data partitioning is used to split BP datasets, then the data splitting process is simple and fast, but data independence between sets is compromised and distribution balance is lost
Solution Approach 1:
The patent segments the BP dataset by subject ID first, creating subject-specific groups. Within each subject group, data is further segmented by BP value ranges (intervals). This hierarchical segmentation ensures that when splitting into training/validation/test sets, entire subject groups remain together, maintaining data independence while preserving distribution characteristics across sets.
Solution Approach 2:
The patent changes the partitioning parameter from random assignment to structured assignment based on subject ID and BP value intervals. By defining specific intervals for SBP and DBP values and assigning subjects to intervals based on their measurements, the method transforms the splitting process from random to controlled, ensuring both independence and distribution balance.
2Adaptability or versatility
If data from the same subject is distributed across different sets, then more diverse data is covered in each set, but model evaluation becomes over-optimistic due to data leakage
Solution Approach 1:
The patent extracts the subject ID as a key identifying feature and uses it to group related data points. By extracting and utilizing this identifying information, the method separates subject-specific data clusters from the overall dataset, ensuring that all measurements from a single subject end up in the same evaluation set, thereby preventing data leakage and ensuring valid model evaluation.
3Ease of manufacture
If standard cross-validation methods are applied to BP estimation problems, then general ML validation procedures are followed, but the specific characteristics of BP data (non-independence, skewness, multiple targets) are not addressed
Solution Approach 1:
The patent applies local quality by treating different subjects and their BP measurements with specialized handling rather than uniform random assignment. Each subject's data is processed locally based on their specific measurement characteristics (SBP and DBP values), assigning them to appropriate intervals and ensuring their complete group remains together during splitting. This localized approach accommodates the specific characteristics of BP data while maintaining a systematic validation process.
Data Source
AI summary
A data splitting method for validating machine learning is adapted to a BP (blood pressure) dataset having a plurality of subjects and includes: dividing measurement ranges of the SBP (systolic blood pressure) data and the DBP (diastolic blood pressure) data into first intervals and second intervals; generating a plurality of classes according to the first and second intervals, wherein each class includes one of the first intervals and one of the second intervals; determining and recording a match condition of the BP data of each subject and classes, and thereby generating a plurality of match conditions corresponding the plurality of subjects, wherein each match condition includes a plurality of labels corresponding to the classes, each label has a first state or a second state; and performing a distribution procedure according to the matching conditions to distribute the BP data of subjects into a plurality of subsets.