各校計畫成果

活動簡介

Prof. Liu’s current research focuses on developing Machine Learning models in HealthTech. The complex healthcare system presents a wide range of meaningful opportunities, from risk assessments, early predictions, to treatment selection and drug discovery.

The research focuses on understanding cancer heterogeneity and enhancing cancer treatment through three key research areas: image segmentation, drug property prediction, and multi-modal data integration. These areas aim to advance our knowledge of how cancer varies between patients and even within a single tumor, which is crucial for developing more personalized treatment approaches.

The first research area involves the development of advanced image segmentation methods specifically for cell images, including circulating tumor cells (CTCs) and immune cells. CTCs play a critical role in cancer metastasis, and by accurately identifying and analyzing them, researchers can better understand how cancer spreads throughout the body. Additionally, immune cells provide valuable insights into how the body’s immune system responds to cancer. By improving image segmentation techniques, the research aims to provide more precise data on these cell types, helping to tailor cancer therapies
more effectively.

The second area of focus is on creating pretrained models that can transform a drug compound into a latent representation. This representation can be used for downstream tasks such as predicting drug properties, including efficacy and potential side effects. By generating these models, researchers aim to streamline the drug development process by providing more accurate predictions of how drug compounds will behave in the body. This approach not only speeds up the development of new treatments but also allows for more targeted therapies, potentially reducing the time and cost associated with bringing new cancer drugs to market.

The third area emphasizes integrating multiple data modalities, such as genomic, proteomic, and molecular data, to predict a patient’s drug sensitivity. This approach enables more personalized treatment plans by identifying which therapies are most likely to be effective for individual patients. The integration of different types of data provides a more holistic view of cancer, leading to more accurate predictions of how a tumor will respond to treatment.

To support long term monitor, information extracted above should be expanded to a longitudinal monitoring system. Our team is working on building manifold learning over time by analyzing vital signs collected from patients in the Emergency Department. They are developing models to predict adverse events, such as sudden deterioration in a patient’s condition. Once these models are proven useful, the goal is to build latent representations that describe a patient’s overall health status. These representations will then be integrated with molecular data, such as multiomics data, to provide a
comprehensive view of a patient’s health and how it changes over time. This could lead to early detection of health issues and more timely interventions.

Figure 1: The oncology patient management from risk assessment, early diagnosis,
treatment selection to minimal residual disease detection. Each step involves the
integration of various data modalities. In this illustration, we focus on treatment
selection, and integrate model development for drug compounds, cell images, molecular
signals.