Large Kernel CNN for Disease Severity Indexing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for analyzing time series of medical events, such as binned encounter counters, fail to properly weight and distinguish between the severity and recency of health conditions, and do not effectively capture the evolution of chronic diseases due to their reliance on small stride symmetric kernels that do not account for decaying properties and casual structures.
Innovation Solution
The implementation of a one-dimensional Convolutional Neural Network (CNN) with large kernels based on Gamma distributions, which weights recent medical data more than older data, effectively capturing temporal relations and evolving characteristics of health conditions by using a weighted large kernel function that assigns higher importance to recent occurrences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If binned encounter counters are used for analyzing time series of medical events, then the analysis method is simple, but it fails to properly weight and distinguish between severity and recency of health conditions
Solution Approach 1:
The patent transforms the simple binned encounter counter approach into a sophisticated temporal analysis method by introducing time-varying weights based on Gamma distributions. The parameter changes include: (1) introducing a temporal weight parameter that decays over time according to Gamma distribution, (2) transforming static encounter counts into dynamic weighted sums where each encounter is weighted by its recency and severity, (3) using learnable parameters in the neural network to adapt weights to different disease patterns. This resolves the contradiction by maintaining computational simplicity while achieving precise differentiation of severity and recency through parameter-based weighting.
Solution Approach 2:
The patent replaces the mechanical, static binned counting system with a neural network-based computational system. Instead of simply counting encounters in fixed time bins, the system uses a one-dimensional CNN with large kernels to compute weighted sums where weights are determined by Gamma distributions and learned through training. This substitution enables the system to automatically learn optimal weighting schemes for different disease patterns, achieving both precision in distinguishing severity/recency and adaptability to various medical conditions.
2Manufacturing precision
If small stride symmetric kernels are used in CNN, then the model can capture local patterns, but it fails to capture the evolution of chronic diseases and does not account for decaying properties and casual structures
Solution Approach 1:
The patent applies asymmetry by using asymmetric Gamma distribution kernels instead of symmetric kernels. The Gamma distribution naturally produces asymmetric weight patterns where recent events receive higher weights and older events receive progressively lower weights, reflecting the decaying property of medical events. This asymmetric weighting scheme enables the model to capture the temporal evolution of chronic diseases where recent symptoms are more indicative of current disease state than distant history, while still maintaining the ability to capture local patterns through the kernel convolution operation.
Solution Approach 2:
The patent introduces dynamics by making the kernel weights time-dependent through Gamma distributions. Instead of using static, fixed-size kernels, the system employs dynamic kernels where the weight assigned to each time point varies according to the Gamma distribution parameters. This allows the model to adaptively capture different temporal patterns: for acute conditions, the Gamma distribution can be configured to emphasize very recent events, while for chronic conditions, it can be configured to consider longer time windows with appropriate decay. The neural network training process further dynamically adjusts these parameters to fit different disease patterns.
3Measurement precision
If large kernels based on Gamma distributions are used, then the model captures temporal relations and evolving characteristics effectively, but the computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the temporal analysis into manageable convolutional operations. Instead of computing a single large-weighted sum over the entire time series, the system uses a one-dimensional CNN that processes the time series through a series of overlapping local windows (kernels). Each kernel position computes a weighted sum over a local segment, and these local computations are aggregated to produce the final temporal representation. This segmentation reduces computational complexity from O(T) for a full-time-series weighted sum to O(T-k+1) for k-sized kernels, while still capturing temporal relations through the sliding window approach and Gamma distribution weighting.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for generating a disease severity index based on a machine learning system executing a one-dimensional convolutional neural network (CNN) using a large kernel. An embodiment operates by collecting training data sets, extracting a training data feature set from the training data sets, weighting features in the training data feature set to generate a weighted training data feature set, selecting a large kernel comprising the weighted training data feature set, executing a one-dimensional convolution of the weighted training data feature set based on the large kernel to generate an aggregation of weighted occurrences and generating an index based on the aggregation of weighted occurrences.