ICU Mortality Prediction Model for Documentation Bias Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ICU mortality prediction models suffer from bias due to variations in documentation practices across institutions, leading to inaccurate mortality risk estimates.
Innovation Solution
A method and system using a trained ICU mortality prediction model that extracts and analyzes defined ICU prediction features, minimizing outlier bias through manual curation and grouping admission diagnoses, and employs a generalized additive model to generate accurate mortality likelihoods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated extraction of EHR data is used for risk modeling, then efficiency and inter-rater reliability are improved, but bias through variation in documentation patterns is introduced
Solution Approach 1:
The patent introduces an intermediary layer of manual curation between automated EHR data extraction and the final risk model. Clinical experts review and curate the extracted data to identify and correct documentation variations, acting as a mediator that filters out bias while preserving the efficiency of automated extraction. This resolves the contradiction by allowing automated processing to maintain productivity while human expertise eliminates documentation-related bias.
Solution Approach 2:
The patent implements feedback mechanisms where curators review extracted data and provide corrections, which are then fed back into the model training process. This iterative feedback loop allows the system to learn from documentation variations and adjust its risk estimates, eliminating bias while maintaining the high-throughput capability of automated extraction.
2Measurement precision
If manual curation of features is performed to minimize bias, then prediction accuracy is improved, but time consumption increases
Solution Approach 1:
The patent applies partial curation by focusing manual review on only the most critical and bias-prone features rather than all features uniformly. This selective approach maintains high prediction accuracy for the most important variables while reducing the overall time investment required for curation. The system identifies which features need manual verification based on their susceptibility to documentation variation.
Solution Approach 2:
The patent performs preliminary automated extraction and preliminary curatorial screening before final model training. This staged approach allows bulk processing of data initially, then applies detailed manual curation only where necessary, reducing total time consumption while maintaining accuracy. The preliminary actions filter out obvious errors and prepare data for more targeted human review.
3Measurement precision
If admission diagnosis is used as a prediction feature, then mortality prediction capability is improved, but misclassification due to subjective variation occurs
Solution Approach 1:
The patent transforms the admission diagnosis parameter from its original subjective, text-based form into standardized, coded categories through manual curation. By changing the parameter representation from free-text diagnoses to controlled vocabularies and standardized codes, the system maintains predictive capability while eliminating misclassification caused by subjective variation in diagnosis coding across different providers and institutions.
Data Source
AI summary
The present disclosure relates to methods and systems for predicting intensive care unit (ICU) mortality. More specifically, the methods and systems for predicting a likelihood of ICU mortality described herein enable robust modeling of ICU mortality that addresses biases in automated data collection, including variations in documentation practices across different units, different hospital systems, and across time. In certain embodiments, the methods described herein include: providing an ICU mortality prediction system; obtaining a plurality of records for a patient in an ICU covering at least a first time period; extracting a plurality of different defined ICU prediction features for the patient; analyzing the extracted plurality of different defined ICU prediction features using a trained ICU mortality prediction model; generating a likelihood ICU mortality for the patient based on the analysis; and presenting the generated likelihood of ICU mortality for the patient via a user interface.