AI Medical Condition Probability Prediction via Feature Vectors
Solution Overview
Problem
Current AI models in cardiac ultrasound exams are inflexible and require specific sets of views to detect heart conditions, making them time-consuming and prone to errors due to the need for manual selection of relevant input data.
Innovation Solution
A method using two machine-learning models to predict the probability of a medical condition in a subject by first extracting feature vectors from input data, which can include a variety of medical images and views, and then using these vectors to generate a medical condition score.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual selection of relevant input data is performed, then the AI model can focus on relevant information, but the process becomes time-consuming and error-prone
Solution Approach 1:
The AI model automatically identifies and selects relevant sequences and views from ultrasound exams without requiring manual curation by experts. The system self-services by autonomously filtering input data, eliminating the time-consuming manual review process while maintaining detection accuracy through intelligent algorithmic selection of diagnostically relevant information.
Solution Approach 2:
The system performs preliminary automated selection and filtering of relevant input data before it is processed by the AI model for disease detection. This preliminary action of automatically identifying pertinent sequences and views prepares the data in advance, saving expert time while ensuring the model receives appropriate input for accurate detection.
2Device complexity
If a fixed set of views is required for AI models, then the model structure is simplified, but the system becomes inflexible and situation-dependent
Solution Approach 1:
The system dynamically adapts the set of input views and sequences based on the specific clinical situation and diagnostic requirements. Rather than using a fixed rigid structure, the AI model flexibly selects which ultrasound views and sequences are relevant for each particular exam and suspected condition, allowing the system to adjust its input requirements dynamically while maintaining a manageable model structure.
Solution Approach 2:
The AI model is designed to handle multiple types of ultrasound views and sequences universally, accommodating various cardiac conditions and examination protocols. The system can process different input configurations (different numbers of views, different sequence types) without requiring separate specialized models, achieving versatility while maintaining a unified model architecture.
3Reliability
If experts manually curate input data for AI models, then relevant information can be selected, but the process is prone to errors and may leave out relevant information
Solution Approach 1:
The system eliminates manual curation by implementing automated intelligent selection of relevant input data. The AI model autonomously identifies pertinent sequences and views based on diagnostic criteria, removing human error from the selection process while improving operational simplicity. The system self-services by automatically filtering and preparing appropriate input data without requiring expert intervention.
Solution Approach 2:
The system incorporates feedback mechanisms where the AI model evaluates the diagnostic value of different input sequences and views, continuously optimizing its selection process. Based on feedback from detection results and relevance assessment, the system refines its automatic selection criteria, improving input selection accuracy over time while maintaining ease of operation through automated decision-making.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Proposed concepts thus aim to provide schemes, solutions, concept, designs, methods and systems pertaining to predicting a probability of the presence of a medical condition in a subject. In particular, embodiments aim to provide a method for predicting a probability of the presence of a medical condition in a subject by providing input data comprising at least one medical image of the subject to machine-learning models which first predict at least one feature vector of the input data and then, based on the feature vector(s), predict a probability of the presence of an indicator of a medical condition in the input data.