Clinical Outcome Prediction From Temporal EHR Text Sequences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting clinical outcomes from electronic health records face challenges due to differing data structures and ontologies across healthcare providers, leading to noise, bias, and overfitting, especially in patients with comorbidities, and lack clinical interpretability.
Innovation Solution
A method that converts structured electronic health record data into a text sequence, using pre-trained language models to encode temporal information, and employs data augmentation and masking techniques to train a machine learning model for improved prediction and clustering of clinical outcomes, minimizing loss and enhancing interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If manually curated mappings between ontologies are used to combine different electronic health record data sources, then the data can be integrated into a single model, but noise and bias are added to the already existing sources of noise, error and missing values in EHRs
Solution Approach 1:
The patent introduces an intermediary mapping layer that translates between different clinical code ontologies (ICD-9, ICD-10, Read codes, CPT codes) and a unified internal ontology. This intermediary layer uses automated computational methods rather than manual curation to reduce the introduction of noise and bias while maintaining the ability to integrate diverse EHR data sources from multiple providers
Solution Approach 2:
The patent transforms the approach by changing the parameter of ontology mapping from manual curation to automated computational mapping. This parameter change enables the system to process multiple ontology types while reducing human-introduced errors and maintaining prediction accuracy through algorithmic consistency
2Productivity
If existing machine learning techniques are applied to EHR data, then predictions can be made about health conditions, but the models overfit to prevalent diseases particularly with respect to patients having comorbidities
Solution Approach 1:
The patent applies preliminary data augmentation techniques before model training, including synthetic data generation and resampling methods, to balance the dataset and reduce overfitting to prevalent diseases. This preliminary action ensures that rare diseases and comorbidities are adequately represented in the training data, improving prediction accuracy across all disease prevalences
Solution Approach 2:
The patent uses excessive sampling of rare disease cases and minority classes through data augmentation to counterbalance the natural prevalence distribution. This partial or excessive action on underrepresented classes ensures they receive sufficient training attention without compromising the overall model performance on common diseases
3Adaptability or versatility
If different providers use different data structures with differing ontologies to describe clinical data, then each provider can maintain their own digital record, but combining multiple modalities in a single model becomes a significant technical challenge
Solution Approach 1:
The patent segments the data integration process into distinct modular components: individual provider ontology mapping layers, a unified internal ontology, and prediction model layers. Each provider's data structure is handled independently through its own mapping module, reducing integration complexity while maintaining flexibility for diverse data sources
Solution Approach 2:
The patent creates a universal internal ontology that serves as a common interface for all provider-specific ontologies. This universal layer enables the system to handle multiple data structures and modalities from different providers through a single unified framework, reducing overall integration complexity
4Adaptability or versatility
If manual mapping techniques are used to combine ontologies, then data integration can be achieved, but information and granularity of the original data are lost during mapping
Solution Approach 1:
The patent changes the mapping parameter from lossy manual aggregation to lossless automated translation. The system maintains detailed clinical granularity by using computational methods that preserve the specificity of original clinical codes and observations while enabling cross-ontology integration, preventing information loss during the mapping process
Data Source
AI summary
A computer-implemented method of training a machine learning model to predict a clinical outcome or characteristic based on a patient's clinical history is disclosed. The method comprises: providing training data comprising structured electronic health record data for a plurality of patients, the structured electronic health record data comprising a plurality of clinical observations, each clinical observation having a text description and an associated time stamp, wherein the training data for each patient is labelled with one or more labels, each representing a clinical outcome or characteristic; converting each patient's electronic health record data into a text sequence comprising the text descriptions concatenated in sequence of the time stamps; inputting the text sequence into a machine learning model; and training the machine learning model to predict a clinical outcome or characteristic based on the input text sequence.