Development of a Latent Space Representing State Transitions and Transition Prediction Methods Using Health check-up Data Linked to Information on the Presence or Absence of Colorectal Polyps

Reference No.	2025a026
Type/Category	Grant for Young Researchers and Students-Short-term Joint Research
Title of Research Project	Development of a Latent Space Representing State Transitions and Transition Prediction Methods Using Health check-up Data Linked to Information on the Presence or Absence of Colorectal Polyps
Principal Investigator	Raiki Yoshimura（Nagoya University・Student (D2)）
Research Period	January 5,2026. - January 9,2026.
Keyword(s) of Research Fields	Machine Learning, disease progression, colorectal polyp
Abstract for Research Report	This study analyzes health check-up data linked to colorectal polyp information. We adopt an approach combining latent space construction based on Variational Autoencoder (VAE) with transition prediction methods to quantitatively represent individual time-series state transitions based on health examination data. Specifically, we construct a VAE model that reconstructs a patient's check-up data after a specific time period from their data at an earlier point. Unlike conventional VAE models, our approach utilizes data from two time points rather than one. We handle the encoding and decoding latent space vectors separately and incorporate time difference information into the connecting perceptron, enabling the construction of a latent space that can represent state transitions after any time period. Additionally, our loss function considers the binary classification loss for colorectal polyp presence/absence. The theoretical features of this research include: - Establishing a framework that extends conventional VAE to represent state transitions between two time points - Mathematical modeling of state transitions using perceptrons incorporating time difference information - Designing a loss function with dual constraints: colorectal polyp status and patient identification Through this research, we expect to gain quantitative insights into state transition mechanisms from health examination data based on colorectal polyp status. Since the transition prediction method using latent space will utilize relatively simple blood test data, it can be implemented in medical institutions nationwide, potentially serving as a support tool for monitoring health conditions and preventive medicine.
Organizing Committee Members (Workshop) Participants (Short-term Joint Usage)	Raiki Yoshimura（Nagoya University・Student） Shizuo Kaji（Kyushu University・Professor）