Objective: Understand and characterize clinical, genetic, and temporal data interactions in health
Rationale: Health data consists of clinical, genetic, imaging data sources, which can be combined to i) predict patient outcomes, ii) detect conditions and diseases, and iii) determine which treatments are effective. For example, early prediction of psychosis disorders like schizophrenia can help allocate resources to needy patients and can improve patient outcomes. However, we do not clearly understand which modality helps more in predicting psychiatric conditions (clinical versus genetic). Further the statistical patterns of interactions between these data sources are unclear. Clinical data is complex and consists of labs, procedures, imaging, and other data. However, brute-force combinations reduce our ability to weigh different features. Current AI models can provide biased results if not designed with care to correctly model how much information each data source carries. This poses a problem in our ability to identify which data to collect for patients, especially when some data, like genetic data, are expensive. Considering these challenges, we propose to address the following questions:
Research questions: How should clinical data, which consists of tabular, imaging, and clinical text data over time, be combined with unchanging genomic features to understand their interactions across major disorders, starting with mental health disorders?
Which AI models improve prediction but also capture which data modalities are the most informative. We aim to develop new AI models architectures that improve biomedical discovery. We will apply these models to predict major conditions, starting with mental health disorders.