ML Model Ecosystem for Low-Incidence Risk Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle with accurately predicting low-incidence events, such as opioid overdoses or opioid use disorder, due to imbalanced class distributions and high-dimensional data, leading to inefficiencies and inaccuracies in identifying minority classes.
Innovation Solution
A system of machine learning models that includes a generative model to create synthetic minority-class records, a predictive model to make predictions, an explanatory model to provide feature explanations, and a certainty model to estimate confidence, addressing the imbalanced class problem and improving prediction accuracy and certainty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If oversampling is used to balance class distributions, then the model can learn minority class patterns, but the model overfits and fails to generalize to new data
Solution Approach 1:
The patent uses SMOTE (Synthetic Minority Oversampling Technique) to generate synthetic copies of minority class samples by interpolating between existing minority class instances. This creates artificial training examples that preserve the underlying distribution patterns without simply duplicating existing data, thereby improving minority class representation while maintaining generalization capability.
Solution Approach 2:
The patent applies random noise addition and feature perturbation to modify the parameters of synthetic samples. By introducing controlled variations in feature values and adding Gaussian noise, the synthetic data maintains the essential characteristics of the minority class while exhibiting sufficient diversity to prevent overfitting and improve generalization.
2Productivity
If undersampling is used to balance class distributions, then the model can train more efficiently, but valuable data is discarded
Solution Approach 1:
Rather than discarding majority class samples through undersampling, the patent creates synthetic copies of minority class samples to achieve class balance. This approach preserves all original training data while generating additional minority class examples, thereby maintaining information integrity and avoiding data loss.
3Reliability
If SMOTE is used for data augmentation, then synthetic minority samples are generated, but the technique fails in high-dimensional spaces
Solution Approach 1:
The patent segments the high-dimensional feature space by selecting a subset of relevant features for SMOTE synthesis. By identifying and focusing on the most discriminative features through feature selection methods, the patent reduces the effective dimensionality of the synthesis process, making it computationally tractable while preserving the essential patterns needed for accurate prediction.
Solution Approach 2:
The patent applies different processing strategies to different features during synthetic data generation. By identifying locally important features and applying targeted transformations or noise addition only to those features, the patent maintains high-quality synthetic samples in critical dimensions while reducing computational burden in less important dimensions.
Data Source
AI summary
A system of machine learning (“ML”) models for making actionable predictions regarding a low-incidence event, including a predictive ML model that has been trained on an augmented training data set comprising synthetic minority-class records to produce a prediction and a certainty ML model that produces a certainty estimate. A method of applying ML models to make an actionable prediction, including training a predictive ML model to make a prediction regarding a low-incidence event from a medical record, applying the predictive ML model to medical records to produce a prediction, and applying a certainty model to generate a certainty estimate. The low-incidence event may comprise risk of opioid use disorder.