Download Research Tools
Stuart, a 66-year-old man with diabetes, felt lousy—constantly fatigued, nauseated, and short of breath after just the slightest exertion. His daughter, worried by his increasing frailty, took him to the emergency room at the local hospital. Her concern was amply justified: Stuart was suffering from heart failure. Like 5.1 million other Americans each year who suffer from heart failure, he was admitted to the hospital to treat this serious, often life-threatening condition. The caring medical team stabilized his condition, and Stuart left the hospital after 10 days, glad to be home with words of advice and a few medications. Within a month he was back, once again fatigued, and facing a second episode.
Stuart’s story is far from rare. Hospital readmissions for chronic conditions such as diabetes, chronic obstructive pulmonary disease (COPD), and congestive heart failure (CHF) are both common and very costly. Studies conducted in the United States indicate that nearly 20% of Medicare patients who are hospitalized for chronic conditions are often readmitted within 30 days. Experts at Edifecs indicate that it costs Medicare—and US taxpayers—about $26 billion a year, and often a large majority of these readmissions are actually considered avoidable with accurate prioritization and personalized care protocols. Readmission-related costs have become so onerous that the Affordable Care Act includes financial rewards and penalties to deal with the readmission problem. Hospitals that reduce their readmission rates receive financial incentives; those that do not, lose reimbursement and get penalized.
Holistic tools that can reliably predict heart-failure readmissions—taking into account all aspects of each patient’s condition and risk factors—would significantly help patients and hospitals. The growth in the use of electronic patient records has recently offered the potential for such analysis, but little had been done to harness the collective intelligence contained in hospital patient records augmented with other data sources.
By introducing cloud computing technology and applying some of the latest advances in machine learning techniques, researchers are rapidly changing this situation.
One leading example of this is RaaS (Readmission Score as a Service), a platform that was developed by the University of Washington (UW) Tacoma’s Center for Data Science. RaaS compares a patient’s medical information to a database of heart-failure outcomes, using advanced machine learning techniques to arrive at a risk-of-readmission factor as well as corresponding actionable guidelines for the patient-provider team. Those patients identified with a high risk receive additional treatment: the goal is to reduce their likelihood of readmission and produce overall healthier outcomes across all stages of the patient care continuum.
The hundreds of machine learning models of RaaS are developed by using both the R machine learning language, and Microsoft Azure Machine Learning. This chronic care management predictive platform relies on historical patient data from multiple sources. These sources include anonymized electronic medical records, claims, labs, medications, and psycho-social factors, all labeled with observed outcomes that the machine learning models access and share in sync to provide continuous monitoring for personalized patient alerts.
RaaS is available as an on-premises service as well as via the cloud by using Azure Machine Learning web services and the Azure-based Zementis Adapa scoring engine to make predictions for patients. When deployed using Azure Cloud Services, RaaS performs data preparation at scale.
The UW Center for Data Science team began developing initial models in collaboration with MultiCare Health System in March 2012, using just two on-premises servers. The maintenance, frequent updates, and down times of these on-premises servers posed an ongoing problem, and scalability issues limited the scope of the project by affecting the speed of data exploration and machine learning.
About a year and a half ago, the team applied for and was awarded an Azure for Research grant, taking advantage of the Microsoft Research program that offers training and awards of computing resources to qualified institutions that use the cloud to advance scientific discovery. The award enabled the Center for Data Science team to scale up the project and create a robust prediction engine that generates a readmission risk factor score for patients at every stage of their hospital care: post-admission, pre-discharge, and post-discharge.
The RaaS platform at MultiCare Health enables the care management team to view an electronic dashboard that shows heart-failure patients’ risks of readmission. UW Medicine Cardiology is now collaborating with the Center for Data Science team to study the efficacy of predictive models for augmenting care management guidelines by using machine learning.
—Daron Green, Deputy Managing Director, Microsoft Research—Gregory Wood, MD, UW Medicine Cardiology
In the 1999 American film Bicentennial Man, the late Robin Williams played a robot who strives to achieve the physical, social and legal status of a human being. The character’s growing language capabilities—his capacity to communicate fluently with his human family—proved crucial in his quest. But long before Williams donned his robot suit, people were dreaming about talking with machines naturally, conversing with them as they would with another person.
Earlier this year, Microsoft Korea hosted a roundtable on “Research on Signal Processing and Speech,” describing recent work on human-machine natural language communication. The research, a collaborative effort between Yonsei University and Microsoft Research, was led by Professor Hong-Goo Kang of the Yonsei’s School of Electrical and Electronic Engineering. Its ultimate goal is to make natural conversation between humans and machines possible.
The roundtable highlighted signal processing and speech research using DNNs.(pictured in bottom row: Professor Hong-goo Kang, Yonsei University [left], and Miran Lee, Microsoft Research [right])
The team focused on voice synthesis and text-to-speech (TTS) conversion, two elements crucial in achieving fluent, natural sounding machine speech. The mechanical, depersonalized voice of machines had been a limitation of previous TTS technologies, according to Kang, which is why the team focused on TTS technology based on deep neural networks (DNNs). DNNs attempt to replicate the neural network of the human brain, particularly the way neurons communicate with one another. By so doing, DNN facilitates a sophisticated type of machine learning that researchers call deep learning. Deep learning should allow machines to understand human speech and respond more relevantly and with a more natural sounding voice.
“The copyright of the research result belongs to me, but other IT companies and everyone else can share it,” said Kang. “It’s hard to conduct this kind of long-term project with just the resources in academia. Therefore, we must work with companies, which is why collaboration with Microsoft Research was so meaningful.” Microsoft Research also offered an internship to one of Kang’s students, who subsequently published his research and presented it at an international conference.
This collaboration is indicative of our commitment to create an ecosystem that connects companies and academic institutions, and our ongoing efforts to foster talented young computer-science researchers.
—Miran Lee, Principal Research Program Manager, Microsoft Research
While we know that climate change will likely affect every aspect of the food system—from our ability to grow food, to the reliability of food transportation and food safety, to the dynamics of international trade in agricultural goods—we don’t yet know how to anticipate and mitigate against what may be negative changes. With this in mind, on July 24, 2015, Microsoft, in partnership with the United States Department of Agriculture (USDA), will launch the Innovation Challenge, a contest designed to explore how climate change will impact the United States’ food system with the intent of achieving better food resiliency.
The challenge invites entrants to develop and publish new applications and tools that can analyze multiple sources of information about the nation’s food supply, including key USDA datasets that are now hosted on Microsoft Azure, Microsoft’s cloud-computing platform.
The challenge offers prizes—including a top prize of US$25,000—for applications that make use of the USDA data and provide actionable insights to farmers, agriculture businesses, scientists or consumers. In addition, through the Microsoft Azure for Research program, Microsoft is granting hours of cloud computing time and terabytes of cloud storage to be used to aid university researchers and students who take part in the challenge. With a November 20, 2015, deadline for entries, challenge participants have three months to submit their applications. Winners will be announced in December 2015.
The food resilience theme of the challenge seeks to inspire the creation of tools that help users analyze and explore our food systems. For the first time, key USDA datasets are available in the cloud, where they can be accessed and blended with other data to obtain novel insights or produce new types of end-user applications. Combining USDA data with cloud-computing resources allows even very high fidelity and complex models to be processed in a timely manner and enables results to be delivered to remote users on their laptops, tablets or mobile phones.
The increased prevalence and availability of data from satellite imagery, remote sensors, surveys and economic reports mean that we can analyze, model and predict an extremely diverse set of properties associated with our food production. Applications might combine data from the USDA and other government sources, such as the National Oceanic and Atmospheric Administration (NOAA), the National Aeronautics and Space Administration (NASA) or the United States Geological Survey, and can be targeted at farmers, scientists, food producers, insurance companies or consumers.
Simply put, the intent of the challenge is to stimulate the exploration of the USDA’s data and to encourage new questions to be asked of these data, either in isolation or in combination with other data feeds or tools. We expect that many developers will start from existing data science tools, machine learning algorithms and visualization techniques; whatever the starting point, we are confident that participants will create valuable tools that promote the goal of food resilience.
For more information about the USDA partnership, read the Microsoft on the Issues blog.
—Daron Green, Deputy Managing Director, Microsoft Research