ICU Risk Prediction Function for Cross-Site Dataset Shift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive computational models for medical complications in clinical settings face challenges due to inhomogeneities within datasets and dataset shift caused by differences in sampling frequencies, equipment, demographics, and clinical environments, affecting model performance.
Innovation Solution
A method using a machine learning model, such as a deep-learning neural network, trained on external data with varied sampling frequencies, combined with local assessment data to monitor and mitigate dataset shift, enabling performance adjustments through updates or replacements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning model is trained on external data from multiple clinical sites with different sampling frequencies and equipment, then the model's adaptability to various clinical environments is improved, but the model performance deteriorates due to dataset inhomogeneities and dataset shift
Solution Approach 1:
The patent applies parameter changes by transforming the clinical data through resampling to uniform time intervals, normalizing feature scales, and standardizing sampling frequencies across different clinical sites. This preprocessing aligns the heterogeneous data parameters, enabling the model to learn from diverse sources while maintaining performance reliability.
Solution Approach 2:
The patent introduces an intermediary data processing layer that acts as a mediator between the heterogeneous clinical data from multiple sites and the machine learning model. This intermediary performs domain adaptation techniques, including feature alignment and distribution normalization, to bridge the gap between different data sources and the model's learning requirements.
2Ease of manufacture
If the prediction function is initially assessed only using external assessment data, then the initial model development is simplified, but the assessment of model performance in local clinical settings becomes inaccurate
Solution Approach 1:
The patent applies preliminary action by performing initial model training and assessment using external data from multiple clinical sites before deployment. This preliminary phase establishes a baseline model that can then be fine-tuned or validated against local assessment data, ensuring both efficient development and accurate local performance evaluation.
Solution Approach 2:
The patent implements feedback mechanisms where local assessment results are used to evaluate and potentially retrain the prediction function. The performance metrics from local clinical settings feed back into the model development process, allowing for continuous improvement and adaptation to specific local conditions while maintaining the benefits of multi-site training data.
Data Source
Figure 1~3
Figure 4~6
AI summary
The disclosure concerns, inter alia, a medical data processing device (100, 110, 120), configured to receive or retrieve a pre-generated prediction function, the prediction function configured to determine a risk indicator (800) for a medical complication event associated with a patient during a hospitalization in an intensive care unit and/or an intermediate care unit and/or a telemetry unit of a first clinical site (601), in particular following a surgery, the pre-generated prediction function generated by training a machine learning model (701) using external training data (210) comprising patient data originating from one or more second clinical sites (602), the patient data comprising manually and/or automatically collected data of a set of measurement-based clinical parameters associated with a plurality of patients, the data comprising at least a subset of the clinical parameters collected with a plurality of different respective sampling frequencies; and assess (S420) the performance of the prediction function (S420) by generating at least one performance metric by using, as the input of the prediction function, local assessment data (320) comprising patient data originating from the first clinical site (601).