Patient Health State Estimation With Interpretable GAM Time Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive models for patient health states, particularly those using neural networks or decision trees, lack interpretability and suffer from undesirable jumpiness in predictive scores, making them ineffective for reliable and smooth health state estimation.
Innovation Solution
A device and method utilizing a generalized additive model (GAM) for estimating health states, which applies a non-linear univariate function on time-series physiological data and uses a sigmoid function to provide a smooth and interpretable probability estimation, enhanced by an interaction component to increase expressiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks or deep learning techniques are used for temporal-based models, then predictive power is improved, but interpretability deteriorates
Solution Approach 1:
The patent segments the complex predictive modeling task into two distinct components: an interpretable base model (generalized additive model with decision trees) that maintains transparency, and a neural network component that captures complex temporal patterns. This segmentation allows each component to specialize - the GAM provides interpretability while the neural network enhances predictive power through learning temporal dependencies in the data.
2Loss of information
If decision tree models are used, then interpretability is improved, but smoothness with respect to feature space deteriorates, causing jumpiness in predictive scores
Solution Approach 1:
The patent merges decision tree models with smoothing techniques to create a hybrid approach. The decision trees provide the interpretable structure for feature selection and initial predictions, while the smoothing component (such as Gaussian processes or kernel methods) is applied to the predictions to eliminate jumpiness and produce continuous, smooth score trajectories over time, combining the strengths of both approaches.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
The present invention relates to device for estimating a probability of a health state of a patient, particularly a probability of an infection of the patient, the device comprising an input unit configured to obtain physiological data of the patient, the physiological data comprising information about a plurality of time series of physiological parameters of the patient, a processor configured to extract the plurality of time series from the physiological data and to apply a machine learning algorithm on the plurality of time series to estimate the probability of the health state of the patient, and an output unit configured to provide the probability of the health state, wherein the machine learning algorithm is based on a generalized additive model, in which results of an application of a learned non-linear univariate function on the parameter time series are summed up and the probability of the health state is obtained by applying the standard sigmoid function on said sum.