Clinically-Driven Virtual Patient Cohorts Generation: An Application to Aorta.

Area of Science:

• Computational biology
• Medical imaging
• Machine learning

Background:

• Machine learning and computational modeling offer valuable insights for therapies and devices via in-silico experiments.
• Acquiring large patient cohorts for machine learning is data-intensive and often requires manual intervention.
• Generating synthetic patient cohorts automates data acquisition but can be computationally expensive for clinical relevance and inter-patient variability.

Purpose of the Study:

• To address the challenge of generating virtual patient cohorts of thoracic aorta geometries for in-silico trials.
• To develop methods for generating cohorts that meet specific clinical criteria without needing a reference sample of that phenotype.
• To formalize clinically-driven sampling and evaluate strategies for efficiency and statistical control.

Main Methods:

• Formalization of clinically-driven sampling for virtual cohort generation.
• Assessment of various sampling strategies focusing on efficiency and statistical property control.
• Utilizing generative adversarial networks (GANs) for creating synthetic thoracic aorta geometries.
• Employing non-linear predictors as efficient alternatives for evaluating anatomical or functional parameters.

Main Results:

• Generative adversarial networks (GANs) demonstrated the ability to produce reliable, clinically-driven cohorts of thoracic aortas.
• The proposed methods achieved good sampling efficiency, ensuring generated individuals belong to the target population.
• The statistical properties of the generated cohorts could be effectively controlled.
• Non-linear predictors proved to be an efficient alternative to computationally expensive parameter evaluations.

Conclusions:

• Generative adversarial networks are effective for creating clinically relevant virtual patient cohorts of thoracic aortas for in-silico trials.
• Clinically-driven sampling strategies enhance the efficiency and control of virtual cohort generation.
• This approach facilitates the development and testing of new therapies and clinical devices through automated, in-silico experimentation.