Neural Network Survival Analysis Apparatus for Interpretable Feature Importance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing analysis methods, particularly those using deep learning models like Deep Belief Networks and DeepHit, fail to output factors contributing to predictions in survival analysis, especially when data is censored, and assume linear independence, which is often not true, leading to inaccurate results.
Innovation Solution
An analyzing apparatus that processes data including censored information using a neural network capable of generating a linear model for each patient, allowing the output of factors contributing to predictions by optimizing learning parameters through a statistical gradient method and calculating feature importance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models (e.g., Deep Belief Network, DeepHit) are used for survival analysis, then prediction accuracy is improved, but the ability to output factors contributing to predictions is lost
Solution Approach 1:
The patent segments the deep learning model into multiple functional layers: an encoding layer for feature extraction, a decoding layer for reconstructing feature importance, and a hazard function estimation layer. This segmentation allows the model to maintain high prediction accuracy while separately outputting interpretable factors contributing to predictions, thus resolving the contradiction between accuracy and interpretability.
2Loss of information
If linear models (e.g., Cox proportional hazard model) are used to output prediction factors, then interpretability is improved, but the ability to handle non-linear relationships and censored data is reduced
Solution Approach 1:
The patent creates a composite model structure that combines the interpretability of linear models with the predictive power of deep learning. The encoding layer uses non-linear neural network operations to handle complex relationships and censored data, while the decoding layer uses linear operations to produce interpretable feature importance weights. This composite approach simultaneously achieves both interpretability and reliability.
3Device complexity
If assumption of linear independence is made in analysis models, then model simplicity is improved, but prediction accuracy deteriorates when the assumption is violated
Solution Approach 1:
The patent introduces dynamic adaptive mechanisms where the model automatically adjusts feature representations and relationships based on the data. The neural network learns non-linear transformations and interactions between features without requiring explicit assumptions of linear independence. This dynamic approach maintains model tractability while accurately capturing complex dependencies in the data.
Data Source
AI summary
The analyzing apparatus: generates first internal data; converts a position of first feature data in a feature space, based on the first internal data and a second learning parameter; reallocates, based on a result of first conversion and the first feature data, the first feature data to a position obtained through the conversion in the feature space; calculates a predicted value of a hazard function of analysis time in a case where the first feature data is given, based on a result of reallocation and a third learning parameter; optimizes the first to third learning parameters, based on a response variable and a first predicted value; generates second internal data, based on second feature data and the optimized first learning parameter; converts a position of the second feature data in the feature space, based on the second internal data and the optimized second learning parameter; and calculates importance data.