Geographic Adherence Risk Index Using Reduced SDoH Dimensions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-dimensional social determinants of health (SDoH) data are challenging to analyze computationally, making it difficult to predict patient behaviors such as adherence to treatment regimens due to complexity and resource constraints.
Innovation Solution
An ensemble of AI/ML models is employed to preprocess and reduce the dimensionality of SDoH data, followed by training predictive models to generate a geographic-based index of patient adherence risk, utilizing dimensionality reduction models, predictive models, patient classification, and regional similarity models, with visualization for insights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-dimensional SDoH data are analyzed directly, then prediction accuracy of patient behaviors is improved, but computational complexity and processing resources increase significantly
Solution Approach 1:
The patent segments the high-dimensional SDoH data analysis process into multiple stages: dimensionality reduction phase followed by prediction phase. This segmentation allows the system to first reduce computational complexity through dimensionality reduction techniques, then apply prediction models to the reduced data, thereby maintaining prediction accuracy while reducing overall computational burden.
Solution Approach 2:
The patent applies dimensionality reduction as a preliminary action before performing patient behavior prediction. By pre-processing the high-dimensional SDoH data to reduce its dimensionality, the system prepares the data in a form that is computationally more efficient for subsequent prediction operations, thus resolving the contradiction between accuracy and complexity.
2Measurement precision
If high-dimensional SDoH data are analyzed directly, then comprehensive patient behavior prediction is achieved, but processing time increases
Solution Approach 1:
The patent performs dimensionality reduction as a preliminary step before patient behavior prediction. This pre-processing action reduces the data volume and complexity, enabling faster subsequent prediction operations while preserving the essential information needed for comprehensive patient behavior analysis.
Solution Approach 2:
The analysis process is segmented into dimensionality reduction and prediction stages. This segmentation allows the system to optimize each stage independently, with the reduction stage focusing on efficiency and the prediction stage focusing on comprehensiveness, thereby reducing overall processing time while maintaining prediction quality.
3Device complexity
If dimensionality reduction is applied to SDoH data, then computational complexity is reduced, but data information may be lost
Solution Approach 1:
The patent employs dimensionality reduction techniques that transform the data representation while preserving essential information. By changing the parameters and structure of the data through mathematical transformations (such as PCA or other reduction methods), the system reduces computational complexity while maintaining the critical information needed for accurate patient behavior prediction.
Data Source
AI summary
Methods and systems to train and use an ensemble of artificial intelligence/machine learning (AI/ML) models to extract information from social determinants of health (SDoH), including training each of multiple dimensionality reduction models to reduce dimensionality of socio-demographic variables associated with a respective one of multiple SDoH categories, training a predictive model to predict a patient behavior for a geographic region (e.g., risk of non-adherence to treatment regimens) based on dimensionally reduced SDoH (alone or in combination with selected socio-demographic variables and/or other data), training a patient classification model to classify patients based on prescription transactions, and/or training a regional similarity model to determine a measure of similarity between geographic regions based on SDoH and/or dimensionally reduced SDoH. Also disclosed are techniques to visually represent outputs of the models on a user-interactive display.