For more details, please also visit MiCHAMP website.
The recent explosion in health data has created unprecedented opportunities for healthcare improvement. However, traditional analytics and existing computational platforms are poorly suited for handling the size and complexity of health data, leading to few real-world examples of ‘big data’ successfully impacting clinical care. One core methodological challenge that currently limits health research is to analyze temporal patterns in longitudinal data for novel discovery and prediction. Although there exists an extraordinary volume of information on patients over time, temporal patterns are frequently overlooked in favor of simplistic, cross-sectional snapshots. This project aims to develop methodologies for understanding longitudinal data, estimating time-varying parameters and predicting patient-specific trajectories. The research team will test their methodologies in the context of two clinical challenges: (1) to improve the accuracy and timeliness of diagnosing acute respiratory distress syndrome onset and (2) to advance abilities to predict progression of chronic hepatitis C virus (HCV) infection.
The research team will establish the Michigan Integrated Center for Health Analytics & Medical Prediction (MiCHAMP), with a group of investigators from five U-M schools and colleges. MiCHAMP will create a vibrant ecosystem that brings together (1) method experts in computer science, engineering, and statistics and (2) health domain experts and clinicians using novel computational platforms built by (3) informatics experts. This tripartite approach improves not only the quality, efficiency, and relevance of multidisciplinary data science in health research, but also its transparency, reproducibility, and dissemination. Through the initial project, the team will gain a deeper understanding of the temporal patterns in complex, real-world patient data through innovative analytic techniques, facilitate earlier diagnosis and treatment in a personalized manner, and build a framework to generalize the methods to other health conditions.
MiCHAMP will build partnership with UM researchers in a Patient Centered Clinical Outcomes Research Institute Clinical Data Research Network, and utilize the rich computing and statistical resources on campus to enable sharing, reusing, and remixing of data and models. MiCHAMP will also incorporate clinical experts and leaders who are well positioned to integrate data science into the day-to-day workflow in the clinics and to spread such practice throughout the U-M community so that new insights will directly impact patient care. The research team’s ultimate goal is not only to develop data science methodologies that directly impact healthcare, but to also revolutionize the way data science is integrated into healthcare.
Brahmajee K. Nallamothu, Principal Investigator, Professor, Department of Internal Medicine
Marcelline Harris, Associate Professor, Department of Systems, Populations and Leadership
Jack Iwashyna, Associate Professor, Department of Internal Medicine
Joan Kellenberg, Research Area Specialist Senior, Department of Internal Medicine
Jeffrey McCullough, Associate Professor, School of Public Health
Kayvan Najarian, Associate Professor, Department of Computational Medicine and Bioinformatics
Hallie Prescott, Assistant Professor, Department of Internal Medicine
Andrew Ryan, Associate Professor, School of Public Health
Kerby Shedden, Professor, Department of Statistics
Karandeep Singh, Clinical Assistant Professor, Department of Learning Health Sciences
Michael Sjoding, Clinical Lecturer, Department of Internal Medicine
Jeremy Sussman, Assistant Professor, Department of Internal Medicine
V.G.Vinod Vydiswaran, Assistant Professor, Department of Learning Health Sciences, and School of Information
Akbar Waljee, Assistant Professor, Department of Internal Medicine
Jenna Wiens, Assistant Professor, Department of Electrical Engineering and Computer Science
Ji Zhu, Professor, Department of Statistics
- MiCHAMP now consists of 69 researchers, including 20 trainees.
- The team is in the planning phase to develop a summer short course aligned with the MIDAS Data Science Certificate Program.
- The team has received R01, K01 and K23 funding support from NIH.
- The team has built a machine learning model that incorporates 1,000 features derived from routinely collected electronic health record (EHR) data, and can predict the onset of Acute Respiratory Distress Syndrome (ARDS) better than the best clinical model currently used. The team is now improving the model by leveraging unlabeled data and semi-supervised learning approaches, as well as incorporating more difficult features in the temporal data.
- The team is developing multi-step forecasting of physiologic waveform data, which could be used to improve early detection of patients with hemodynamic decompensation. The team is investigating novel multi-output deep architectures that explicitly model the joint probability of the signal, which is required for accurate multi-step forecasting (predicting multiple values simultaneously).
- The team has compared longitudinal models and cross-sectional models in their prediction of disease progression among 156,588 veterans with Hepatitis C, and concluded that longitudinal models are superior for this purpose.
- The team has examined data quality and its impact in two ways. It has completed a scoping review of the impact of data quality on the implementation of predictive models. The team is now examining whether current data quality characterizations reflect observed data discrepancies across linked data sources, and identifying new sources of error.
The study’s Aim 1 is to improve accuracy and timeliness of diagnosing Acute Respiratory Distress Syndrome (ARDS) onset. The team is focusing on using data from the first six hours after the patient is admitted to predict ARDS onset. They are examining 395 patient admissions, 868 features (meds, vitals, labs etc.), and 35 cases identified using chart review by three physicians. The preliminary results are promising, with an accuracy rate of 0.82 (Receiver Operating Characteristics). In the next stage, the team will develop semi-supervised approach to leverage large amount of EHR data that has not been through chart review; incorporate time-series data and transform the problem into a survival analysis task; and develop larger data repository for training and temporal distinct cohort for testing.
The study’s Aim 2 is to advance abilities to predict progression of chronic Hepatitis C virus infection. The team is awaiting national data on veterans with Hepatitis C (N=200,000). They are developing methods for model prediction using HALT-C data.
The study team’s effort in education, training and communications has yielded the following results: 1) Three undergraduate students, two graduate students and two postdoctoral researchers are involved in the study. 2) The study’s website was recently launched (http://michamp.med.umich.edu/). 3) A weekly seminar series has been established.