We are building the data and computing infrastructure and developing pioneering methods to link data science, laboratory technologies, clinical studies and trials, genomics, and other methodologies to help researchers identify pathways for perioperative complications and develop corresponding interventions to improve perioperative outcomes.
- Leverage electronic health records (EHR) resources and well-proven AI models for real-time perioperative risk management. Over the past years, our AI team has accumulated extensive experience in machine learning based analytics for real-world EHR data. For example, we have demonstrated the real-world applicability, robust transferability, and potential clinical utility of Gated Recurrent Unit with Decay (GRU-D) 1 based longitudinal deep learning architecture in real-time risk assessment of post-surgical complications (PSCs) in a series of studies. We also successfully reengineered GRU-D to estimate the Weibull probability density function, which can provide probability and point estimates across various prediction horizons at multiple time points of follow-up. The proven ability to handle data missingness, asynchronicity, and provide real-time risk update makes GRU-D based architecture an ideal candidate for tackling challenges associated with perioperative risk assessment.
- Foster collaboration between clinical and healthcare AI teams on developing novel AI architectures. A highly relevant example would be temporal models with built-in missing parameterization as well as RETAIN 7 like self-explainability. The currently available model agnostic explainability mechanisms (e.g. Sharpley Value, partial dependence plot) do not suit temporal models which incorporate newly arriving measurements and expect on-the-fly feature contribution analysis. In contrast, temporal models with automated missing parameterization and self-explainability are significantly underexplored, highlighting a critical area for advancement.
- Advance digital health and AI through addressing information overload problems which can help patients, clinicians, and health systems to optimize healthcare decisions. We plan to combine advanced digital health and AI techniques with a deep understanding of healthcare workflows and user needs to address information overload problems in perioperative medicine. We will use careful data integration, intelligent information filtering, user-centric design, and rigorous validation.
- Leverage AI technology with local and national data resources, imaging facilities, and sequencing technologies to optimize diagnostic and therapeutic development. The research team leverages big data sets such as IQVIA or Epic Cosmos and AI foundation models for novel diagnostic and therapeutic development.