Weighted Tensor Learning Data Generation for Medical Prediction Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep tensor technologies face challenges in accurately predicting medical treatment needs due to similar partial patterns in attendance record data being processed as the same common pattern, leading to noise and deterioration in prediction accuracy, especially for individuals with medical history.
Innovation Solution
A learning data generation method that generates weighted tensors based on attendance record data, where the weight of the tensor is adjusted according to the period of time, with higher weight for the period before medical treatment and lower weight for the period after, allowing for more accurate differentiation between individuals with and without medical history.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep tensor processes partial common patterns in the core tensor, then processing efficiency is improved, but data with different features processed as same pattern causes noise and deteriorates prediction accuracy
Solution Approach 1:
The patent applies local quality by assigning different weights to different time periods within the tensor data. Specifically, data periods before medical treatment receive higher weights while periods after treatment receive lower weights. This allows the model to focus on clinically relevant periods while still processing the entire time series, thereby maintaining processing efficiency without sacrificing prediction accuracy.
Solution Approach 2:
The patent changes the parameter of data importance by introducing a weighting mechanism that modifies the significance of different time periods. By dynamically adjusting weights based on temporal relationships with medical treatment events, the model can differentiate between relevant and irrelevant patterns, resolving the contradiction between processing efficiency and prediction accuracy.
2Quantity of substance
If attendance record data of persons with medical history is included in learning data, then data volume is increased, but similar partial patterns become noise and deteriorate prediction accuracy for new medical treatment
Solution Approach 1:
The patent applies local quality by differentiating the importance of different time periods within the attendance record data. Data from periods before medical treatment is assigned higher weights as it contains predictive signals, while data from periods after treatment is assigned lower weights as it represents post-event behavior. This allows the model to utilize large volumes of historical data without allowing post-treatment noise to degrade prediction accuracy for new treatment events.
3Device complexity
If irregular attendance patterns before and after medical treatment are treated as same common pattern, then pattern recognition simplicity is improved, but differentiation capability between pre and post treatment situations deteriorates
Solution Approach 1:
The patent resolves this contradiction by applying local quality through temporal weighting. The model maintains a unified pattern recognition framework for simplicity, but introduces local differentiation by assigning higher weights to pre-treatment periods and lower weights to post-treatment periods. This allows the same pattern recognition mechanism to adaptively differentiate between pre and post treatment situations based on temporal context, maintaining both simplicity and differentiation capability.
Data Source
AI summary
A learning device receives, for each target, learning data that represents the source of generation of a tensor including a plurality of elements which multi-dimensionally represent the features of the target over a period of time set in advance. When the target satisfies a condition set in advance, the learning device identifies the period of time corresponding to the condition in the learning data. Subsequently, the learning device generates a weighted tensor corresponding to the learning data that is at least either before or after the concerned period of time.