I will review our work in extracting clinically relevant characterizations of anatomy and pathology from medical images in two domains. First, joint modeling of image, genetic and clinical data is used to gain insight into the patterns of disease in large heterogeneous clinical populations. Examples include studies of white matter disease in stroke patients from brain MRI, of genetically defined patterns of emphysema in COPD patients as observed in chest CT, and others. The second family of applications aims to provide accurate delineations of pathology and make predictions in medical scans of individual patients. Examples include functional imaging of the placenta and cardiac image analysis for surgical planning.
2016 MIT Digital Health Conference
The Agency for Healthcare Research and Quality was established in 1989 in response to an Institute of Medicine report that pointed out ?escalating healthcare costs, wide variations in medical practice patterns, and evidence that some health services are of little or no value?. More than 25 years later, there has been surprisingly little progress in these three areas. The interest in applying machine learning to clinical practice is increasing yet the practical application of these techniques has been less than desirable. There is a persistent gap between the clinicians required to understand the context of the data and the engineers who are critical to extracting useable information from the increasing amount of healthcare data that is being generated.
For the past decade technologists have been on a mission to disrupt medicine the same way they’ve disrupted practically every other societal system - from the bottom up, by the consumers of healthcare. This would entail replacing much of what doctors do by AI and big data in the cloud and, ultimately, the “Uber-ization of Healthcare”. The result would magically be lower costs and better outcomes.
But my recent work in the digital health space has shown that we technologists can’t approach medicine the same way we’ve approached media and music. Healthcare is a different beast. Anyone who thinks that apps and data alone are going to convince people to change their health-related behaviors - which is the only way to lower costs and improve outcomes at scale - is simply ignoring human nature.
Twine Health was founded on the belief that the real opportunity for technology in healthcare is to strengthen, not weaken, critical human-to-human relationships in the system. We have developed a Collaborative Healthcare IT platform, based on six years of research at the MIT Media Lab, with that principle in mind. We have demonstrated that when patients with chronic conditions like hypertension and diabetes are empowered to take the lead in their health, but with the continuous support and caring of their clinical team, costs drop dramatically and outcomes are greatly improved.
Receptor Tyrosine Kinases (RTKs) are critical for normal human physiology, but can be oncogenic when highly expressed or mutated in a wide array of human cancers. To define the critical components in these networks, we have developed mass spectrometry based methods enabling the absolute quantification of tyrosine phosphorylation sites in RTK signaling networks at high temporal resolution following stimulation with different ligands or inhibitors, in vitro and in vivo. Quantitative phosphorylation data generated in this analysis provides insight into the occupancy of multiple tyrosine phosphorylation sites on the receptor, highlights mechanisms of differential regulation in response to different ligands, and highlights resistance mechanisms to selected inhibitors.
The ability to create increasingly complex genomic data generated directly from patient tumors may impact our understanding of cancer and affect clinical decisions about cancer treatment. As the quantity of genomic data generated from individual cancer patients greatly expands, innovations will be needed to successfully implement large-scale genomics at the point-of-care. These include new ways to 1) interpret large-scale data from individual patients and 2) understand why patients respond (or don't respond) to existing and emerging cancer therapies such as targeted therapies, chemotherapies, and immunotherapies. Dr. Van Allen will explore how the emerging discipline of clinical computational oncology is powering new approaches for the clinical interpretation of large-scale genomic data and how these data are helping physicians understand why certain patients benefit from cancer therapies when others do not. While still in its infancy, this new field of clinical computational oncology may drive the widespread implementation of precision cancer medicine in the years to come.
For a little over a dozen years, our group has been developing, integrating, and testing various bihormonal (insulin and glucagon) bionic pancreas technologies for autonomous regulation of glycemia in people with type 1 diabetes (T1D). The technology has evolved over the years from a crude and clumsy system of interconnected pumps and sensors cobbled together around a laptop computer, to a system that runs on an iPhone, which wireless communicates with two infusion pumps and a sensor, and, finally, to its ultimate embodiment as a dual-chamber infusion pump, a sensor, and mathematical algorithms all housed within a single compact integrated device, which we call the iLet (in homage to the pancreatic islets of Langerhans which contain the alpha and beta cells that secrete glucagon and insulin).
The laptop version of our bionic pancreas was tested first in a diabetic swine model of T1D at Boston University (BU) between 2005 and 2009 and then in inpatient clinical trials with our collaborators at the Massachusetts General Hospital (MGH) between 2008 and 2012 in adults and adolescents with T1D. Between 2013 and 2016 we conducted outpatient clinical trials of the iPhone version of our bionic pancreas together with our clinical collaborators at MGH, Stanford, the University of North Carolina, and the University of Massachusetts. Results of these studies will be presented along with our plans for the final pivotal trials of the iLet and the pathway ahead for regulatory approval.
We present an example of ongoing research in the space of analytics-driven personalized healthcare and showcase an example of a healthcare technology startup spun off of our research endeavor.
The first part of the talk discusses an ongoing research work on personalized diabetes management. Current clinical guidelines for managing type 2 diabetes do not differentiate based on patient-specific factors. We present a data-driven approach for personalized diabetes management that improves health outcomes relative to the standard of care. We modeled outcomes under thirteen pharmacological therapies based on electronic medical records from 1999 to 2014 for 10,806 type 2 diabetes patients from Boston Medical Center. We developed a recommendation algorithm that prescribes a regimen if the expected improvement from switching regimens exceeds a threshold. For patient visits in which the algorithmic recommendation differed from the standard of care, the mean post-treatment glycated hemoglobin (HbA1c) under the algorithm was lower than standard of care by 0.44% +/- 0.03% (p << 001), from 8.37% under the standard of care to 7.93% under our algorithm. A personalized approach to diabetes management yielded substantial improvements in HbA1c outcomes relative to the standard of care. Our prototyped dashboard visualizing the recommendation algorithm can be used by providers to inform diabetes care and improve outcomes.
The second part of the talk presents an overview of MyA Health, a spinoff based on similar research efforts aimed at personalizing health care down to the individual. MyA is powered by a wealth of data sources encompassing historical claims, electronic medical records, wellness and biometric data, wearable device records, and consumer lifestyle data. The backend of MyA is empowered by a high-dimensional analytics engine with: (1) a suite of predictive machine learning algorithms to predict future healthcare costs, disease progression and outcome variability; and (2) robust optimization algorithms to optimize and personalize healthcare decisions that will best mitigate an individual’s financial burden and maximize their healthcare outcomes. To the consumer, MyA is an individual’s healthcare advisor that personalizes decisions ranging from what health plan is best to cover their risk to what drug/treatment is likely to benefit them the most. MyA is unique in that it takes the totality of data sources available to make personalized recommendations, a concept that is made possible given the healthcare data digitization revolution and the increasing adoption of wearable wellness and health monitoring devices.
Cardiovascular disease remains the leading cause of death in the industrialized world. Although research into the etiology and treatment of cardiac disease remains a focus of numerous research groups, the accurate identification of patients who are at risk of adverse events following a heart attack remains a major challenge in clinical cardiology. In this talk I will describe how sophisticated computational biomarkers, which integrate a diverse array of clinical information, can be used to identify patients who are at elevated risk of death after a cardiac event. This work demonstrates that computational biomarkers can provide useful and powerful insights that can help guide clinical decision making.