2020 University of Michigan Precision Health Symposium

There has been a surge in neonatal abstinence syndrome (NAS) due to dramatic increases in prenatal opioid exposure, a consequence of the ongoing opioid epidemic in the United States. In-utero drug exposure is associated with increased risk of NAS and multiple adverse outcomes. Independent of socioeconomic status, children exposed to opioids in-utero and in the neonatal period are at greater risk for later cognitive, language, attention, and visual problems and poorer academic performance. The variability in outcomes, for comparable exposures, implicates a role of gene-by-environment interactions for these neurological phenotypes. Deconstructing the interplay between environment exposures, genetic variation, and cell-specific molecular features is a principle challenge to understanding personalized opioid responses. To address this, we are differentiating human neural precursors and cerebral organoids (with and without opioids exposure) from iPSC lines reprogrammed from 50 donors with high-depth phased whole genome sequence. We will integrate genetic variation with diverse molecular (scRNA-seq, scATAC-seq, BruUV-seq) and cellular (live imaging and micro-electrode assays) features to produce a dense catalog of quantitative trait loci (QTL) that can illuminate the cell types, genes, and regulatory elements associated with personalized opioid-linked neuronal outcomes. Using association summary statistics from these opioid-response QTL maps, we will perform scans across large cohorts (e.g. UK Biobank) to identify cell types and genes associated with diverse adult biology. Collectively, these results will provide a genetic catalog that maps the cell-specific molecular landscape of neuronal effects from developmental opioid exposure to adult disease outcomes and will be a platform for novel gene discovery.

Stephanie Bielas

Department of Human Genetics, University of Michigan Medical School1, Department of Biostatistics, University of Michigan School of Public Health2, Department of Pediatrics and Communicable Diseases3, Department of Radiation Oncology4, Department of Environmental Health Sciences5, Department of Computational Medicine and Bioinformatics6, University of Michigan Medical School, Ann Arbor, MI

Survival from head and neck squamous cell carcinoma (HNSCC) has remained stubbornly low at ~50%, the relapse rate is high and treatment options are frequently limited and debilitating. For late-stage HNSCC tumors, immunotherapy offers a novel and promising treatment option for ~20-30% of HNSCC cases. Lymphocytes that leave the bloodstream and infiltrate the tumor (tumor infiltrating lymphocytes, or TILs) induce cell death upon activation, and the density of TILs in a tumor has been shown to be an important prognostic indicator for many malignancies, including HNSCC. Our overall goal is to predict the benefit of immunotherapy options for individual HNSCC patients by integrating TIL imaging of readily available H&E slides with clinical information and easily obtained genetic results. We have optimized and completed image analysis for 104 cases (of a total of 500 tumors). Most cases were male (62.5%) and current smokers (43.3%) and stage III or IV disease (70.2%). Median age of diagnosis was 61.4 years, and median overall survival time was 32.4 months. Immune cell density was highest in HPV-positive oropharynx tumors compared to HPV-negative (p=0.003), and overall immune cell density was highest for oropharynx compared to the other sites (p = 0.09). While we do not have appreciable numbers to assess a prognostic benefit, there appears to be a benefit associated with higher immune cell density. We will present the machine learning algorithm and our preliminary findings, as well as our future directions.

Laura Rozek <rozekl@umich.edu>

1. Department of Environmental Health Sciences, University of Michigan 2. Department of Computational Medicine and Bioinformatics 3. Department of Pathology, University of Michigan 4. Department of Biostatistics, University of Michigan 5. Department of Otolaryngology, University of Michigan

Solid tumors consist of a significant population of non-malignant cells such as adipocytes, fibroblasts and immune cells that are re-educated by cancer cells to support disease progression, immune evasion, drug resistance and metastasis. The cancer-stroma interactions lie at the heart of several tumorigenic characteristics, which can help cancer cells thrive in harsh microenvironments devoid of nutrients and oxygen. Intratumoral interactions are not limited to signaling axes via cytokines, growth factors and other secreted factors, as our lab has discovered several mechanisms of metabolic cross-talk in pancreatic, prostate, and ovarian cancers. Dissecting metabolic interactions in human tumors is inherently challenging due to a lack of technology that can capture the dynamic nature of metabolic activity. Most intratumoral metabolic mechanisms have been elucidated in cell culture or mouse models, and most recently, in ex vivo human tumor organoids in our lab. However, technologies like next-generation sequencing and single-cell transcriptomics can present unique insight into the transcriptional regulation of tumor metabolism. Here we demonstrate the novel integration of machine learning-based multiomics analysis, along with stable-isotope based flux analysis to reveal metabolic cross-talk mediated by branched-chain ketoacids in cancers. The benefit of analyzing molecular profiles of human tumors also presents an opportunity to stratify patients according to the “metabolic footprints” of their tumors, thereby allowing targeted metabolic therapy.

Abhinav Achreja

1Laboratory for Systems Biology of Human Diseases, University of Michigan, Ann Arbor, MI, USA 2Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI, USA 3Biointerfaces Institute, University of Michigan, Ann Arbor, MI, USA 4Department of Chemical Engineering, University of Michigan, Ann Arbor, MI, USA 5Rogel Cancer Center, University of Michigan, Ann Arbor, MI, USA

Concussions and other types of traumatic brain injuries (TBIs) are difficult for physicians to diagnose because of the available assessment techniques. Physicians can physically examine a patient or take a CT scan of the brain, but these methods are only useful for assessing the primary trauma and tissue damage. Physical exams and CTs are incapable of monitoring the cell death, blood brain barrier breakdown, and swelling that stem from the primary injury. These secondary injuries caused by the primary trauma often determine patient outcomes. Blood-based biomarkers have been proposed to track the progression of secondary TBI injuries. Measurement of TBI biomarkers could result in further stratification of TBI injuries, precision treatment, and better prediction of long-term patient outcomes. However, TBI biomarker data is not actionable as laboratory testing takes hours. There is a need for a point-of-care device that can provide TBI biomarker concentrations within fifteen minutes, making the data actionable. Towards a standalone point-of-care system to quantify TBI biomarkers, we developed a glass microfluidic device consisting of a single channel with a variable height profile as well as bead-based QLISAs (quantum dot-linked immunosorbent assays) to quantify GFAP, IL-6, and IL-8. We used the variable height device to capture the assay beads and analyze the fluorescence intensity of buffer samples spiked with varying concentrations of the biomarkers of interest. Our microfluidic device will ultimately be able to keep pace with TBI biomarker discovery as additional biomarkers can be multiplexed simply by adding in antibody conjugated beads of different diameters.

Alyse Krausz

University of Michigan, Department of Biomedical Engineering, Department of Chemical Engineering, Department of Emergency Medicine

Introduction: A critical proportion of patients undergoing surgery go on to use opioids long-term postoperatively. Interventions to prevent persistent opioid use may prove beneficial but should be prioritized for patients at highest risk. Therefore, we sought to develop a model from multiple data sources to optimize prediction of persistent opioid use in patients undergoing surgery.

Methods: This study included adults from the Michigan Genomics Initiative cohort undergoing surgery from January 5, 2015 to July 3, 2018, divided into a derivation cohort (surgery prior to July 4, 2017) and a validation cohort. Preoperative variables to predict persistent opioid use using logistic regression included demographics, comorbidities, surgical service, preoperative opioid fills (from Michigan’s Prescription Drug Monitoring Program), patient-reported measures (pain, mental health), and natural language processing-derived variables (risky alcohol and drug use). Persistent opioid use was defined as any perioperative opioid fill followed by any opioid fill 91-180 days postoperatively.

Results: The analysis included 19,352 eligible patients (derivation cohort - 14,671; validation cohort - 4,591) with a median age of 54 (IQR 40-64), 54% female, and 90% Caucasian. Overall, the model had an area under the receiver-operating-characteristic curve (AUC) of 0.80. Using standardized coefficients, the five most important variables were any preoperative opioid use, internal medicine-cardiology service, hydrocodone/acetaminophen use, acute care surgery, and chronic pain. Risky alcohol use was among the least important indicators.

Conclusions: A regression model derived from multiple data sources predicted persistent opioid use following surgery with more accuracy than other published models. These results highlight risk factors relevant for future prevention interventions.

Anne Fernandez

University of Michigan

The process of identifying the locations of short substrings of a read in a candidate reference genome, known as seeding, is one of the most time-consuming steps in whole genome secondary analysis. For instance, in the widely used read alignment software BWA-MEM seeding consumes 56% of the overall runtime (measured on 50x coverage reads from Illumina Platinum Genomes). In addition, seeding is also a performance bottleneck in several metagenomic classification tools such as Centrifuge. Seeding is commonly performed using a highly-compressed index structure called the FM-index. While FM-index is highly space efficient (4.3 GB for a human genome), standard seeding algorithms using FM-index are memory bandwidth limited. Our key observation is that this trade-off is not optimal for modern systems which have availability of much larger main memory (192 GB on Amazon c5n.18xlarge instance). Our proposed seeding algorithm is based on a new data structure called Enumerated Radix Tree (ERT) designed to address this bandwidth bottleneck: it trades-off a larger memory footprint for a significant reduction in memory bandwidth. Overall, ERT when integrated into BWA-MEM2 speeds up overall read alignment by 1.28× and provides up to 2.1× higher seeding performance while guaranteeing identical output to the original software. Furthermore, we prototype an FPGA implementation of ERT on Amazon EC2 F1 cloud and observe 1.6× higher seeding throughput over a 48-thread optimized CPU-ERT implementation.

Software: https://github.com/arun-sub/bwa-mem2

Arun Subramaniyan

University of Michigan

Tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), is the deadliest infectious disease worldwide. Current TB treatment regimens have not changed for over 50 years. Multidrug resistance, diversity of pathogen strains, several candidate drugs and pharmaco-kinetics/dynamics (PK/PD) greatly complicate treatment design strategies. To study the interplay of these factors, we adopt a systems biology approach integrating drug kinetics, host immune environment and pathogen genotype. Aim 1. Predicting in vivo interactions of TB drugs: We present a multi-scale pipeline linking drug transcriptomics with pharmacokinetics (Cicchese et.al, BioRxiv 2020.09.03.281550). We have integrated a machine learning model, INDIGO-MTB, which predicts in vitro drug interactions using drug transcriptomics, with a multi-scale model of drug PK/PD and pathogen-immune interactions called GranSim, in collaboration with the Kirschner and Linderman labs. An in vivo drug interaction score (iDIS) was calculated from dynamics of drug diffusion, spatial distribution, and activity within lesions against various Mtb sub-populations. iDIS of drugs significantly correlates with efficacy metrics from clinical trials. Our approach identifies mechanisms that can amplify synergistic or mitigate antagonistic drug interactions in vivo by modulating the relative distribution of drugs. Aim 2. Predicting clinical outcomes of TB infection: We analyze electronic health records of TB patients from the NIH TB portal and Michigan Data direct, in collaboration with Prof. Zhenhua Yang. Clinical phenotype data including multi-drug regimens are available for several thousand patients infected with Mtb strains with different antibiotic susceptibilities. Using decision trees and INDIGO-MTB scores, we aim to identify features most predictive of clinical outcome for individual patients.

Awanti Sambarey and Sriram Chandrasekaran <asambare@umich.edu>

University of Michigan

The ongoing COVID-19 pandemic has raised concerns for many regarding personal and public health implications, financial security, and economic stability. Alongside many other unprecedented challenges, there are increasing concerns over social isolation and mental health. We introduce Expressive Interviewing--an interview-style conversational system that draws on ideas from motivational interviewing and expressive writing. Expressive Interviewing seeks to encourage users to express their thoughts and feelings through writing by asking them questions about how COVID-19 has impacted their lives. We present relevant aspects of the system's design and implementation as well as quantitative and qualitative analyses of user interactions with the system. In addition, we conduct a comparative evaluation with a general-purpose dialogue system for mental health that shows our system's potential in helping users to cope with COVID-19 issues.

Charles Welch

University of Michigan

Recently there has been rising interest in investigating the dynamic nature of functional connectivity (FC) in the brain, particularly at rest. Currently, though, the study of dynamic FC is limited by the sliding window method, which is very sensitive to window size and therefore has poor test-retest reliability. Furthermore, the overlapping nature of the windows both precludes definitive segmentation of the time course into states and poses scalability issues for long/granular time series. Here we propose a new statistical paradigm for detecting dynamicity in a time series with data-driven, graph-aware techniques that take maximal advantage of the temporal granularity of fMRI. Our method improves upon the standard approach in three major ways. First, we generate tailored segments of stable FC in each time series, bounded by peaks in our Global Temporal Derivative, a univariate summarization of instantaneous activation changes in fMRI time series. Next, we obtain a connectivity graph for each tailored segment using “top-K” thresholding on connectomes, which allows for direct comparison of FC between variably-sized segments. Finally, we use state-of-the-art graph embedding methods to generate a lower-dimensional latent representation for each connectivity graph, used as feature vectors for standard k-means clustering. Our method reliably recovered known transitions between task/rest blocks in structured fMRI (p-value < 2.2e-16) and the resulting clustering significantly outperformed the sliding window in accurately separating known task and rest states (homogeneity statistic 0.23 vs. 0.02, respectively). In rest, we reliably find 5 states across four fMRI sessions (I2C2 = 0.89).

Danai Koutra

University of Michigan

Cardiovascular disease (CVD) is the leading cause of death worldwide. Among the well-known risk factors for CVD are elevated circulating blood lipids, hypertension, and lifestyle factors such as obesity and smoking. While therapies lowering the levels of low-density lipoprotein cholesterol (LDL-C) have shown to be effective for patients with symptoms of cardiometabolic events, it is difficult to reverse the damage that is already presented in patient’s arteries. Unlike lifestyle factors, genetic risk factors are constant from birth to death and are therefore powerful predictors for future disease risk. As CVDs have been under high interest of genetic studies, the number of known genetic variations altering the risk of coronary artery disease (CAD) is already over 160. With this information, it is possible to construct a genetic risk score (GRS). We utilize existing optimized GRSs for CAD (Inouye et al. 2018, JACC) and for lipid traits (Transethnic meta-analysis of 1.3M samples, Graham et al. 2020, unpublished). We optimize the prediction models in the Nord-Trøndelag Health (HUNT) study and apply the models to the Michigan Genomics Initiative (MGI) cohort. The effect of the CAD GRS for the risk of future CAD event is as high as the effect of smoking. We are currently integrating more risk factor GRSs to better understand the overall effect of genome information in the disease risk prediction and aim to implement the optimized models into a web application that could be used in clinical practice for earlier risk assessment to fulfill the aim of precision medicine.

Ida Surakka

University of Michigan

Purpose: CKD is one of the leading health issues that causes more deaths than breast cancer or prostate cancer.(nccd.cdc.gov/CKD) In this study, we employed unsupervised machine learning to discover novel patterns and features that associate with the level of kidney function in CKD patients.

Methods: Thirty-eight trichrome digital slides from kidney biopsies were obtained from the C-PROBE Digital Pathology Image Repository at U-M. We performed unsupervised machine learning using a bag-of-words method to simplify representation used in natural language processing (NLP) using biopsy images tiled into 256x256 pixel patches for feature extraction. Features were clustered through K-means clustering and a histogram representation for each biopsy sample was created. We used a random forest model as a classifier to predict association with clinical patient outcomes such as eGFR.

Results: We estimated the optimal number of data cluster as 9 centroids (or phenotypes) represented as a cluster map. The out-of-back error from random forest is 0.23, sensitivity 0.74, and specificity 0.79. The area under ROC curve (AUC) is 0.77 and 95% confidence interval 0.613 – 0.927. F-score 0.76.

Conclusions: The results from our study showed that the visual dictionary (phenotypes) obtained from unsupervised machine learning could be novel features useful for discriminating levels of kidney function and could help in decision making during follow-up.

Jeffrey Hodgin <jhodgin@med.umich.edu>

1Department of Computational Medicine and Bioinformatics1, Pathology2, and Nephrology3, University of Michigan

Introduction: Cardiac rehabilitation (CR) is a medically supervised risk reduction program for patients recovering from a cardiovascular event and reduces hospital readmissions and mortality. As a result of the COVID-19 pandemic, access to CR has become limited. It is unknown whether a mobile device-facilitated telehealth program can improve functional capacity and quality of life for patients enrollment in CR.

Methods: We are performing a prospective, randomized-controlled trial of 220 participants using a virtual study design. We will approach all low- and moderate-risk patients who enroll in CR at Michigan Medicine and who own a compatible smartphone. Participants will be randomized to a control or telehealth arm and followed for 6-months. Participants in both arms will receive usual care and will be provided with a smartwatch (an Apple Watch Series 4 or Fitbit Versa 2). Those in the telehealth arm will additionally receive the following interventions: (1) micro-randomized notifications which encourage activity and are tailored on 4 dimensions of context (weather, time of day, day of week, phase of behavior change); (2) weekly activity summaries via email tailored to phase of behavior change; and (3) activity tracking and goal setting through a mobile application. The primary outcome will be 6-minute walk distance assessed using a mobile application and smartwatch and, secondarily, step count.

Results: 5 participants have consented and enrolled virtually as part of a study pilot. The study will anticipate to launch on September 28, 2020.

Conclusions: Virtual consent and enrollment into a mobile device-facilitated telehealth program is feasible. Further study is needed to determine whether participation improves functional capacity after a cardiovascular event.

Jessica Golbus

University of Michigan

Age-related decline in balance impacts quality of life and presents long-term care challenges for older adults (OAs) over 65 years of age. Balance exercise training programs, while effective for the recovery or development of new sensorimotor strategies for functional mobility, require visits to a clinic and supervision by a physical therapist (PT), for which the costs and associated travel are often prohibitive for the average OA. Home-based training systems are a potential solution, but one-on-one guided training with a PT “in-the-loop” is not scalable. Further, there is a lack of consensus regarding the appropriate progression of balance exercises to optimize sensorimotor integration in OAs. To overcome these limitations, we have developed models for automatically evaluating balance and delivering personalized training guidance for OAs. Current methods for evaluating balance rely on potentially subjective visual assessments from PTs or on self-assessments that do not correspond to actual performance. We have developed a novel machine learning and uncertainty quantification framework with which we are able to accurately characterize mappings between objective, inertial measurement unit-based kinematic recordings of postural sway and ratings of balance performance. We have used these mappings to develop an exercise recommender system that provides personalized exercise progression plans for OAs. Mirroring the process by which PTs evaluate OAs and make exercise progression decisions, this system evaluates OA balance performance and recommends future exercises that strike an optimal balance between ease and difficulty. These models are the first step in the development of an at-home, scalable balance training system for OAs.

Kathleen Sienko <sienko@umich.edu>

1 Department of Mechanical Engineering 2 Department of Electrical Engineering and Computer Science

To facilitate scientific collaboration on polygenic risk scores (PRS) research, we created an extensive PRS online repository for 35 common cancer traits integrating freely available genome-wide association studies (GWAS) summary statistics from three sources: published GWAS, the NHGRI-EBI GWAS Catalog, and UK Biobank-based GWAS. Our framework condenses these summary statistics into PRS using various approaches such as linkage disequilibrium pruning / p-value thresholding (fixed or data-adaptively optimized thresholds) and penalized, genome-wide effect size weighting. We evaluated the PRS in two biobanks: the Michigan Genomics Initiative, a longitudinal biorepository effort at Michigan Medicine, and the population-based UK Biobank. For each PRS construct, we provide measures on predictive performance and discrimination. Besides PRS evaluation, the Cancer-PRSweb platform (https://prsweb.sph.umich.edu) features construct downloads and phenome-wide PRS association study results for predictive PRS. We expect this integrated platform to accelerate PRS-related cancer research.

Contact: Lars Fritsche (larsf@umich.edu)

Lars Fritsche <larsf@umich.edu>

University of Michigan School of Public Health, Department of Biostatistics

Objective assessments of the level of physical activity of an individual can significantly improve clinical decision making for multifarious health issues. In particular, walking speed is a powerful metric for clinicians as it captures performance of a common physical activity, walking. It can easily be obtained in the clinical setting, and most importantly, is highly correlated with various health issues and outcomes. Wearable sensing technologies offer new opportunities to complement clinical assessments of walking. However, the continuous and unobtrusive monitoring of individuals in their daily life requires low-power and compact sensors, which often do not enable direct measurement or calculation of walking speed. We propose an innovative approach to accurately estimate walking speed in the free-living environment using a low-power, lightweight wearable sensor secured on the thigh. During a designed short walking task, highresolution measurements of stride speed enable the derivation of subject-specific models to map stride frequencies extracted from the thigh-worn sensor to stride speeds. The performance of the subject-specific models in the real-world was evaluated using a long unsupervised walk in the freeliving environment with stride speed calculated from a foot-worn inertial measurement unit used as the gold standard. Our results demonstrate that stride speed can be accurately estimated using low-power and low-resolution sensors. We believe that our approach enables new exciting opportunities for objectively measuring physical activity in the real-world, which will ultimately provide clinicians with a more complete picture of patient health.

Loubna Baroudi and Kenneth Shorter <lbaroudi@umich.edu>

1University of Michigan, 2University of Alabama

Type 2 diabetes is one of the earliest comorbidities identified for psoriasis, and epidemiological study suggests they are associated independent of body mass index (BMI) and other comorbidities (~1.5 Odds Ratio), yet there has been no large scale screening to understand the shared their genetic architecture. Our study conducted the first trans-disease meta-analysis between psoriasis and type 2 diabetes, utilizing summary statistics derived from >11,000 psoriatic, >75,000 T2D (BMI-adjusted), and >800,000 control individuals. We identified 4 genome-wide significant (p<5x10-8) shared genetic loci (in proximity to SPRED2, CHUK, PRDX5 and STAT3), each with p-value more significant than the association results for both diseases. These loci have significantly higher evidence of colocalization (p=5.7x10-4) than previously reported loci for psoriasis, and we confirmed our findings in 40,056 Caucasian patients from the Michigan Genomics Initiative, where the risk allele frequency for each locus was higher (by 8% on average) for patients with psoriasis and T2D than controls. Significantly, all four genetic loci overlap H3K27ac marks for primary T and B cell active enhancers, and have a potential role in NF-kB signaling. Mendelian randomization demonstrates that while most of the causal relationship between psoriasis and type 2 diabetes can be explained by BMI, there is a subtle but significant causal effect for psoriasis on T2D. Our work represents a breakthrough in understanding the shared genetic component in immune system between autoimmune and metabolic diseases, and can have significant impact in revealing new therapeutic drug targets for patients suffering from both conditions.

Matthew Patrick <mattpat@umich.edu>

Department of Dermatology, University of Michigan Medical School

Acute dyspnea (acute shortness of breath) is one of the most common presenting symptoms in hospitalized patients, and pneumonia, heart failure, and chronic obstructive pulmonary disease (COPD) are the most common underlying diagnoses. Determining the diagnosis is fundamental to ensuring patients receive the right treatments, but patients frequently receive incorrect initial treatment or experience treatment delays. Artificial intelligence tools have the potential to augment the diagnostic process in acute dyspnea and improve patient care. Previous work has focused on analyzing structured clinical data or chest radiographic data separately when training models to detect these conditions. Here, we propose a multi-modal framework for combining features from the chest radiograph with clinical data from the electronic health record (EHR), which more closely mimics how clinicians diagnose these conditions in clinical practice. We find that models can potentially latch onto spurious correlations between the input features and output label when trained on imaging data alone. For example, when training on a dataset where females patients are disproportionately more or less likely to have the disease of interest than males, a model can learn chest radiograph features that discriminate sex rather than features specific to the disease. We find that a simple transfer learning approach in which features learned from a related task are used to predict the disease of interest can mitigate much of this bias, resulting in a more robust model.

Michael Sjoding

University of Michigan, Internal Medicine, Medical School, College of Engineering

Background: Radical prostatectomy is the most common definitive treatment for men with intermediate risk prostate cancer (CaP). It has a number of side effects, including erectile dysfunction. A tool that would allow prediction of erectile function following radical prostatectomy may help patients make informed decisions and reduce decision regret. Objective: The purpose of this study is to predict 12- and 24-month erectile function following radical prostatectomy, using patient and treatment characteristics. Design, Setting, and Participants: The Michigan Urological Surgery Improvement Collaborative (MUSIC) is a consortium of 46 diverse urology practices that maintains a prospective registry of men with CaP. A random forest (RF) approach was used to train and test models that predict EPIC-26 sexual domain scores and patient reported quality of erections at 12- and 24-months following radical prostatectomy. Model factors included pretreatment patient and treatment characteristics and pretreatment assessments of Interest in Sexual Activity, Satisfaction with Sex Life, and EPIC-26 sexual domain scores. We evaluated model discrimination (area under the curve (AUC)) for binary outcomes and root mean squared error (RMSE) for continuous outcomes using 5-fold cross-validation. Results and Limitations: We identified 9,127 men who underwent a radical prostatectomy between May 2014 and August 2020 in the Michigan Urological Surgery Improvement Collaborative (MUSIC). Of these, 4,115 patients were included in the 12-month analysis and 2,313 were included in the 24-month analysis. RF models predicted EPIC-26 sexual domain scores with an RMSE of 23 at 12 months and 25 at 24 months. Corresponding AUC for the binary outcome of EPIC-26 sexual domain scores above and below 73 were 0.83 and 0.77 respectively, at 12 and 24 months. RF models predicted quality of erections with an AUC of 0.85 at 12 months and 0.80 at 24 months. Baseline sexual domain scores, erectile function, and age were the best predictors based on impurity-corrected random forest variable importance. The predictions for EPIC-26 domain score and quality of erection were well-calibrated. Conclusions: Prediction of EPIC-26 sexual domain scores had a clinically acceptable accuracy and prediction of the quality of erection had excellent discrimination and calibration at 12 and 24 months. This tool has the potential to augment prostate cancer shared decision making.

Nnenaya Agochukwu

University of Michigan

Short tandem repeats (STRs), consisting of repeated 1-6 nucleotide motifs, make up over 3% of the genome. Instability and expansion of ~fifty different STRs are known to cause human disease and these repeat expansions represent the most common monogenic causes of ALS, Frontotemporal Dementia, Ataxia, and Autism. However, these highly penetrant mendelian conditions likely to signify the tip of a much larger iceberg, as STRs represent a major source of genetic variance within humans that may explain missing heritability in complex diseases. Despite their potential importance, we know very little about the normal functions of these repetitive elements or their variation within the human population. Our hypothesis is that STRs throughout the genome are functional, highly variant, and contribute to human disease on a scale that is currently significantly underappreciated. In the emerging era of precision medicine, this blind-spot is deleterious to both our understanding of disease pathogenesis and prognostication and ignores the biological functions of a sizable portion of the human genome. To evaluate our hypothesis, we are applying emerging Oxford Nanopore long read sequencing coupled with bioinformatics approaches to define the human repeat variome (the “repeatome”) and detect both novel and known repeat expansions in human disease. We are also exploring the normal functions of these repetitive elements in gene function and behavior. Together, our preliminary findings describe novel approaches to defining the contributions of genomic STRs to human disease.

Peter Todd

University of Michigan

Alzheimer’s disease and associated dementias are major public health challenges with a multifold increase expected in the coming decades. A significant challenge is that clinical phenotypes (e.g., mild cognitive impairment, Alzheimer’s type dementia) can manifest as a result of multiple distinct or interacting etiologies. The proposed study will utilize cutting edge magnetic resonance imaging (MRI) methods to evaluate two relatively understudied mechanisms that hold potential for enhancing early detection of cognitive decline, clarifying etiology, and facilitating treatment at the patient specific level. First,vascular disease is known to both contribute to cognitive impairment and is frequently found in those with Alzheimer’s disease, yet sensitive and reliable quantitative methods are not currently available. We will overcome this limitation using myelin mapping and “fingerprint” arterial spin labeling (ASL). Second,Alzheimer’s disease is increasingly recognized as having network-level effects and interactionsand requires precision methods to characterize and quantify these changes. Therefore, we will employ multivariate models for improved subtype classification and prediction of clinical and behavioral outcomes. The proposed project leverages the unique cohort of the Michigan Alzheimer’s Disease Center’s ongoing longitudinal study of participants across the dementia spectrum. The overall goal of this project is to develop improved acquisition and analysis methods for the neuroimaging of subjects susceptible to Alzheimer’s dementia in order to improve subtype classification and early prediction of clinical outcome using quantitative measures.

Scott Peltier <spelt@umich.edu>

University of Michigan

Introduction: Postpartum hemorrhage (PPH) is a leading preventable cause of maternal death, although 62% of PPH occurs without identifiable risk factors. A recent study (Venkatesh et al.) using a research dataset found that PPH could be predicted with an area under the receiver operating characteristic curve (AUC) of 0.93 (highly predictable) using information available at the time of admission. Whether this generalizes to settings using electronic health record (EHR) data is unknown.

Methods: Using EHR data between 2/1/2019 and 5/11/2020, we identified delivery hospitalizations for women with an estimated gestational age of ≥ 22 weeks. After setting aside the most recent deliveries into a test set, we randomly divided the remaining into training (67%) and validation sets (33%). We mapped the 55 predictors used in the Venkatesh et al. model to EHR elements. We then fit logistic regression (LR), random forest (RF), and gradient-boosted machine (GBM) models (computer models) using the training and validation sets and compared their performance on the test set using an AUC.

Results: We identified 6,121 eligible hospital encounters, of which 1,354 were complicated by PPH. Of eligible encounters, 3,268 were in the training set, 816 in the validation set, and 2,037 in the test set. In the test set, AUCs for LR, RF, and GBM were .61, .60, and .62, respectively (only modestly predictive).

Conclusions: EHR predictors available at admission cannot reliably predict PPH. Updating model predictions during the hospital encounter may yield better performance.

Sean Meyer and Thomas Klumpner <klumpner@med.umich.edu>

1. Integrative Systems & Design, College of Engineering, University of Michigan 2. Obstetrics & Gynecology, Michigan Medicine 3. Anesthesiology, Michigan Medicine 4. Learning Health Sciences, University of Michigan

Recent technological advancements have enabled detailed investigation of associations between the molecular architecture and tumor heterogeneity, through multi-source integration of radiological imaging and genomic (radiogenomic) data. We integrate and harness radiogenomic data in patients with lower grade gliomas (LGG), a type of brain cancer, in order to develop a regression framework called RADIOHEAD (RADIOgenomic analysis incorporating tumor HEterogeneity in imAging through Densities) to identify radiogenomic associations. Imaging data is represented through voxel intensity probability density functions of tumor sub-regions obtained from multimodal magnetic resonance imaging, and genomic data through molecular signatures in the form of pathway enrichment scores corresponding to their gene expression profiles. We build a Bayesian regression model with the pathway enrichment scores as the response and the probability density functions as the predictors. The hierarchical prior formulation incorporates the grouping structure amongst the predictors induced through the tumor sub-regions. Our analyses reveal several pathways relevant to LGG etiology (such as synaptic transmission, nerve impulse and neurotransmitter pathways), to have significant associations with the corresponding imaging-based predictors. Contact e-mail: shariqm@umich.edu

Shariq Mohammed

University of Michigan, Departments of Biostatistics and Computational Medicine and Bioinformatics

In this work, we introduce a counseling dialogue system that provides real-time assistance to counseling trainees. The system generates sample counselors' reflections -- i.e., responses that reflect back on what the client has said given the dialogue history. We build our model upon the recent generative pretrained transformer architecture and leverage context augmentation techniques inspired by traditional strategies used during counselor training to further enhance its performance. We show that the system incorporating these strategies outperforms the baseline models on the reflection generation task on multiple metrics. To confirm our findings, we present a human evaluation study that shows that the output of the enhanced system obtains higher ratings and is on par with human responses in terms of stylistic and grammatical correctness, as well as context-awareness.

Siqi Shen

University of Michigan. Computer Science

Despite widespread concern for cytokine storms leading to severe morbidity in COVID-19, rapid cytokine assays are not routinely available for monitoring critically ill patients. We report the clinical application of a digital protein microarray platform for rapid multiplex quantification of cytokines from critically ill COVID-19 patients admitted to the intensive care unit (ICU) at the University of Michigan Hospital. The platform comprises two low-cost modules: (i) a semi-automated fluidic dispensing/mixing module that can be operated inside a biosafety cabinet to minimize the exposure of technician to the virus infection and (ii) a 12-12-15 inch compact fluorescence optical scanner for the potential near-bedside readout. The platform enabled daily cytokine analysis in clinical practice with high sensitivity (<0.4pg/mL), inter-assay precision (~10% CV), and rapid operation with 9 min of assay incubation. This test allowed us to perform serial monitoring of two critically ill patients with respiratory failure and to support the immunomodulatory therapy using the selective cytopheretic device (SCD). We also observed clear interleukin-6 (IL-6) elevations after receiving tocilizumab (IL-6 inhibitor) while significant cytokine profile variability exists across all critically ill COVID-19 patients and to discover a weak correlation between IL-6 to clinical biomarkers, such as Ferritin and C-Reactive Protein (CRP). Our data revealed large subject-to-subject variability in a patient’s response to anti-inflammatory treatment for COVID-19, reaffirming the need for a personalized strategy guided by rapid cytokine assays.

Yujing Song <yujing@umich.edu>

Department of Mechanical Engineering, University of Michigan College of Engineering

Preoperative opioid use has been reported to be associated with higher preoperative opioid demand, worse postoperative outcomes, increased postoperative healthcare utilization and expenditures. Understanding the risk of preoperative opioid use helps establish effective pain management for each patient. In the field of machine learning, deep neural network (DNN) has emerged as a powerful means for risk assessment because of its superb prediction power; however, the blackbox algorithms may make the results less interpretable than statistical models. Bridging the gap between the statistical and machine learning fields, we propose a novel Interpretable Neural Network Regression (INNER), which combines the strengths of statistical and DNN models. We use the proposed INNER to conduct individualized risk assessment of preoperative opioid use. Intensive simulations and statistical analysis of 34,186 patients expecting surgery in the Analgesic Outcomes Study (AOS) show that, the proposed INNER not only can accurately predict the preoperative opioid use using preoperative characteristics as DNN, but also can estimate the patient-specific odds of opioid use without pain and the odds ratio of opioid use for one unit increase in the reported overall body pain, leading to more straightforward interpretations on opioid tendency compared to DNN. Our analysis identifies patient characteristics that are strongly associated with the opioid tendency and is largely consistent with the previous findings, evidencing that INNER is a useful tool for individualized risk assessment of preoperative opioid use.

Yuming Sun <yumsun@umich.edu>

Department of Biostatistics, University of Michigan

Branched chain amino acids (BCAAs) supply both carbon and nitrogen in pancreatic cancers, and their increased levels have been associated with increased risk of pancreatic ductal adenocarcinomas (PDACs). It remains unclear, however, how stromal cells regulate BCAA metabolism in PDAC cells and how mutualistic determinants control BCAA metabolism in the tumor milieu. Here we show distinct catabolic, oxidative, and protein turnover fluxes between cancer-associated fibroblasts (CAFs) and cancer cells and a marked branched chain ketoacids (BCKA)-reliance in PDAC cells in stroma-rich tumors. We report that cancer-induced stromal reprogramming fuels this BCKA demand. The TGF-beta/SMAD5 axis directly targets BCAT1 in CAFs and dictates internalization of the extracellular matrix from the tumor microenvironment to supply amino acid precursors for BCKA secretion by CAFs. The in vitro results were corroborated with human patient-derived circulating tumor cells (CTCs) and PDAC tissue slice. Our findings reveal therapeutically actionable targets in pancreatic stromal and cancer cells.

Ziwen Zhu

University of Michigan, Department of Biomedical Engineering