Classifier Selection for Missing Medical Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical decision support systems face challenges in handling missing data, particularly when dealing with heterogeneous patient information sources, as most machine learning algorithms are not designed to handle incomplete records effectively, leading to potential incorrect conclusions.
Innovation Solution
The system selects and applies a classifier based on the available features in a patient record, using pre-computed classifiers or constructing a new one from a subset of training data that matches the available features, thereby addressing missing data by optimizing the feature set for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a global value replaces missing data, then the classifier can operate with complete input, but the accuracy of classification deteriorates due to incorrect conclusions
Solution Approach 1:
The patent segments the classification process into multiple classifiers, each trained on a specific subset of features. Instead of using a single classifier that requires all features, the system divides the problem into multiple specialized classifiers that can handle different combinations of available features, thereby avoiding the need to impute missing data while maintaining accuracy.
Solution Approach 2:
The patent changes the parameter of the classification system from a fixed single classifier to a dynamic selection of classifiers based on the available features. The system adapts the classifier selection based on which features are present in the input data, allowing the system to operate correctly with varying degrees of data completeness without requiring data imputation.
2Ease of operation
If the most probable value replaces missing data using Bayesian network, then the classification can proceed, but the system complexity increases due to solving joint probability
Solution Approach 1:
The patent segments the complex Bayesian network into multiple simpler classifiers, each trained on a specific feature subset. This avoids the need to solve the full joint probability distribution while still capturing the relationships between features and the target class through multiple specialized models.
Solution Approach 2:
The patent performs preliminary action by pre-training multiple classifiers on different feature subsets during the training phase. This allows the system to quickly select and apply the appropriate classifier during inference without having to perform complex probability calculations at runtime, thereby reducing system complexity while maintaining operational capability.
3Reliability
If a learning-based approach estimates joint probability, then the missing data can be handled, but the amount of training data required becomes extremely large
Solution Approach 1:
The patent segments the training data requirement by training multiple classifiers on different feature subsets rather than requiring all classifiers to be trained on the complete feature set. This reduces the effective training data volume needed for each individual classifier while maintaining the ability to handle missing data through the ensemble of specialized classifiers.
Data Source
AI summary
Missing data is addressed in a medical decision support system. The classifier applied to the patient record with missing data is obtained as a function of the available data. For example, one of a plurality of different classifiers is selected based on the features available in the patient record to be classified. The different classifiers are developed using different feature sets. The classifier developed using a feature set closest to or a sub-set of the features available in the patient record is selected for classifying the patient record. As another example, features in a training set corresponding to features available in the patient record are used to build a classifier. The classifier is applied to the patient record by inputting the available features of the patient record.