We propose to advance precision health in pediatrics by developing a database that addresses racial and ethnic disparities, combined with an automated classification system, for bone age assessment in children. Bone age assessment is central to the detection of growth disorders and the better understanding of growth models and their link to obesity, a health epidemic that affects 17% of Michigan youth age 10-17. Our approach develops groundbreaking diagnostic capability, by combining the data available through U-M Precision Health, information theory, artificial intelligence, clinical expertise, and a web-based interface for broad dissemination. In addition, our work will serve as an example of the successful broadening of outdated standards for better clinical accuracy and to better represent our diverse population. As such, the broader impacts go far beyond pediatrics and bone age assessment.

Molecular genetic classification has transformed the practice of brain tumor diagnosis, treatment, and prognostication. Specific genetic mutations drive diffuse gliomas, the most common and deadly primary brain tumor type, to behave more or less aggressively. Discrete, mutually exclusive genetic glioma subtypes have been identified that stratify the patient’s prognosis and survival better than any other classification scheme to date. Critically, the effectiveness of surgery (e.g., biopsy versus subtotal versus aggressive resection) varies depending on the specific genetic subtype. A major problem in precision health is that the genetic subtype is unknown at the time of surgery, potentially resulting in imprecise surgical treatment, early tumor recurrence, increased morbidity and decreased survival. In this proposal, we aim to address this problem by combining stimulated Raman histology (SRH), a rapid, label-free, optical imaging method, and deep neural networks, a type of artificial intelligence (AI) model, to rapidly and accurately diagnose genetic subtypes at the bedside, point-of-care, in a fully automated fashion, which we call precision molecular histology. Upon completion, we will have made (1) a scientific advancement by demonstrating that genetic mutations can be reproducibly identified using AI and SRH data alone and (2) a clinical advancement by providing point-of-care molecular classification for optimal surgical planning and personalized patient care.

Pan-cancer analysis has the potential to identify common driver genes and biomarkers with greater statistical power and accuracy by taking advantage of the increased sample size when integrating datasets from different cancers. In this project, we propose to develop novel statistical and computational methods for asymmetric integration of datasets from different cancers for the identification of cancer-related genes and biomarkers in case-control analyses. We will also develop software tools implementing the proposed methods and apply them to analyze data collected from cancer patients in U-M MGI.

Cardiac surgery is performed annually on more than 300,000 US patients and is commonly complicated by acute kidney injury (AKI). Treatment strategies are available to reduce the risk and severity of AKI in this population, but most AKI-reducing strategies have tradeoffs that adversely affect other organ systems or incur substantial costs and thus should be reserved for high-risk patients. Generalizable models to predict cardiac surgery-associated AKI (CSA-AKI) do not exist, and developing such models has been difficult due to low case numbers of both cardiac surgery and CSA-AKI at individual medical centers and lack of granular perioperative data. While combining data can overcome limitations of individual health centers, this practice runs counter to patients’ privacy expectations. Federated learning refers to a family of approaches that enable the development of multicenter models without the need to share data. Federated learning, while immensely promising and widely used in mobile devices, has been underutilized in healthcare settings. In this proposal, we will use a newly developed approach known as federated stacked learning to develop generalizable models for predicting CSA-AKI using granular high-dimensional perioperative data from 33 hospitals participating in the Multicenter Perioperative Outcomes Group (Aim 1). We will temporally validate these models using registry data (Aim 2), and develop an R software package to ease the training and deployment of such models by other U-M researchers (Aim 3). These models, along with our R software package, will provide key preliminary data and infrastructure for our follow-on R01 proposal focused on learning and testing of personalized treatment strategies in high-risk patients.

Patients with cancer are commonly prescribed fluoropyrimidines (FP) or irinotecan, as well as supportive care medications to manage adverse effects (AE) of their cancer treatment. These medications have known risks for treatment inefficacy or severe toxicity based on pharmacogenetic (PGx) variation.1–5 Despite increasing evidence that genotype-guided strategies can mitigate risks of poorer medication outcomes,6,7 routine use of clinical PGx testing is uncommon. This study will aim to evaluate the impact of clinical PGx test results on clinician prescribing decisions and patient outcomes in a cohort of patients predicted to be at risk for AEs to their chemotherapy regimens based on Michigan Genomics Initiative (MGI) genotypes. Many PGx associations relevant to patients with cancer, such as dronabinol and CYP2C9 genotype are still considered investigational; this study will also leverage MGI to discover or validate additional PGx associations to establish workflows for future translation into prescribing decisions.

Alzheimer’s disease and associated dementias are major public health challenges with a multifold increase expected in the coming decades. Alzheimer’s disease is increasingly recognized as having network-level effects and interactions and requires precision methods to characterize and quantify these changes (as well as intervention-induced change). In this proposal, we will use a deep learning model for improved data representation, subtype classification and prediction of clinical outcomes and apply it to the domain of mild cognitive impairment (MCI) and dementia of the Alzheimer’s Type (DAT).

Deep learning approaches have been increasingly applied to functional magnetic resonance imaging (fMRI) data but face a common challenge on generalizability and explainability. When models learn to optimize the performance for a specific goal or dataset, they often fail to generalize to other or future goals and datasets. The models also tend to be too complex to understand or act as a “black box” unexplainable in terms of neural circuits and dynamics. Despite its increasing application to fMRI, deep learning remains inadequate to advance mechanistic understanding of brain functions or inform treatment of brain diseases. To address these issues, we propose to use an explainable and generalizable system of artificial intelligence. This system represents, model and predict fMRI activity for decoding and modulating neural circuits linked to behavior. To be explainable, the system learns representations that can be decoded and interpreted as spatial patterns and temporal dynamics of brain networks. The system is generalizable to different patients, brain states, behavioral tasks, and disease conditions, including MCI and DAT to be studied in the project.

Aim 1:Characterize latent representations of brain activity with task-free resting state fMRI. We will train the model with resting state fMRI data from the Human Connectome Project (HCP) and Alzheimer’s Disease Network Initiative (ADNI). The model will learn to disentangle the generative factors of spontaneous brain activity, to represent activity and connectivity patterns in terms of independent latent variables, and to decode the representations back to brain networks and dynamics. We hypothesize that the model-learned representations will be separable across subjects in different cognitive conditions (e.g. cognitively normal, MCI, DAT) and be usable to explain and predict behavioral measures.

Aim 2: Characterize latent representations of brain activity with pathology-specific task fMRI.
We will transfer the learned model from the large-scale resting state data to local data acquiring tasks relevant to Alzheimer’s Disease (spatial navigation, object location). We will apply the general linear model analysis to the time series of the latent representation, localize the task-related functions to specific regions in the latent space, and map the so-localized functions from the latent space back to the brain. We hypothesize that the latent space learned from the resting state data is generalizable to task conditions and it can be divided into subspaces or regions associated with specific task-related functions. By focusing on tasks related to MCI- and DAT-pathology, we can characterize individual responses and behavioral phenotypes.

The goal of the proposed study is to develop statistical methods and computational tools for automated curation and harmonization of multi-institutional electronic health records data (multi-EHR), with an application to transferring knowledge between Michigan Genomics Initiative (MGI) and UK Biobank. Although the adoption of manually curated ontologies and common data models has improved data interoperability and exchangeability, failure to reproduce and transfer research findings across healthcare systems has also become more common, largely due to the inherent heterogeneity in medical coding. Manual curation is not only immensely laborious but also error-prone. We propose to develop a series of methods for automated grouping and mapping of medical codes to unify the idiosyncratic “languages” of multi-EHR. Specifically, we will develop statistical methods to group detailed medical codes to clinically interpretable code concepts and to map heterogeneous medical codes between different institutions and across time, based on privacy-preserving summary data. We will validate the results against manually curated labels and perform a case study of transporting a phenotyping algorithm from MGI to UK Biobank. We will also develop software packages to facilitate dissemination.