Center for Medical Machine Learning (CM2L)

Clinical care is complex, data-driven and increasingly outstripping the cognitive abilities of clinicians. Machine learning (ML) has the potential to support, enable and improve clinical decision-making to make it faster, accurate, and economical. In particular, ML-enabled realtime and near realtime monitoring, alerting and guiding at the point of clinical care will power the next generation of clinical decision support. The Center for Medical Machine Learning (CM2L) focuses on developing, implementing and evaluating high performance clinical decision support tools that aid in very specific clinical tasks and are powered by ML. CM2L is a core informatics center in the Department of Biomedical Informatics at the University of Pittsburgh and is directed by Shyam Visweswaran, MD, PhD.

Goal

The goal of CM2L is to 1) develop ML tools for unmet clinical needs; 2) demonstrate internal and external validity of tools; 3) ensure that the tools are fair (algorithmic fairness), 4) obtain FDA approval, and 5) monitor for degradation in performance (algorithmic robustness). We are advancing both the science and the engineering of ML-powered clinical decision support to improve health.

  • 1) Develop ML solution. We will identify unmet clinical needs that can be addressed by using ML for pattern recognition or for prediction. The ML solution will generate insightful predictions that are actionable and have the potential to impact patient care. Specifically, we will focus on ML-augmented CDS that is used at the point of care and involves in-person patient interactions. Moreover, we will focus on using ML with a human in the loop rather than on autonomous ML with the human out of the loop.
  • 2) Demonstrate validity. Validation of an ML tool will typically involve multiple iterations. The initial iteration will use retrospective data to evaluate statistical validity, clinical utility and economic utility. Statistical validity addresses the question: does the ML model perform well on metrics of discrimination and calibration? Clinical utility addresses the question: can the ML model improve clinical care and patient outcomes? Economic utility addresses the question: can the ML model produce cost savings and increase efficiency?
  • 3) Ensure fairness. The ML tools should be non-discriminatory (algorithmic fairness) for sensitive attributes such as age, sex, race, and socioeconomic status. Ensuring fairness is vital as ML tools increasingly play an important role in decisions related to health and the potential for harm increases.
  • 4) Obtain FDA approval. The Software as a Medical Device (SaMD) regulatory framework that is evolving at the FDA will enable rapid approval of ML that is designed to aid clinical decision-making. It is critical to obtain FDA certification for real-word deployment of ML tools.
  • 5) Monitor for degradation in performance. ML tools that are clinically deployed need to be evaluated and monitored for degradation in performance (algorithmic robustness) including degradation over time, across different geographical locations, and across populations that differ in disease severity or prevalence of the outcome.

Current projects at the CM2L include:

  • utilizing ML to enable EMR systems to deliver the right data of the right patient at the right time in the intensive care unit (ICU)
  • applying ML to monitor for ischemia in real time during surgical procedures
  • building ML models to predict outcomes in central line-associated bloodstream infections
  • building a Twitter surveillance system that uses ML models
  • predicting new uses and adverse effects of drugs using ML

News

2015 September NLM R01 funded for Development and evaluation of a learning electronic medical record system. Electronic medical records (EMRs) are capturing increasing amounts of patient data that can be leveraged by machine learning methods for computerized clinical decision support. This project focuses on developing a learning EMR system that uses machine learning to provide decision support using the right data, at the right time.

2018 April NCI R01 funded for Leveraging Twitter to Monitor Nicotine and Tobacco-Related Cancer Communication. The goal of this project is to build a Twitter surveillance system that will use machine learning to analyze vaping-related tweets for inferring behaviors and changes in attitudes in response to policy changes.

2020 June UPMC Enterprises and Center for Commercial Applications of Healthcare Data funded Realtime Evaluation for Adverse Events using Intraoperative Neurophysiological Monitoring (READE IONM). The goal of this project is to develop a machine learning based system to detect brain ischemia in realtime from continuous intraoperative neurophysiological monitoring.