MIT Professor John Guttag Highlights Late Start but Promising Leap for Machine Learning in Healthcare
Today, the growing availability of aggregated datasets and data analytics tools, coupled with federal regulatory mandates for information disclosure, has made machine learning a reality. Machine learning in healthcare holds immense potential to help clinicians, physicians, and researchers uncover patterns within existing datasets, thereby enhancing healthcare efficiency and improving the quality of care. Machine learning is broadly categorized into two types: supervised learning and unsupervised learning, each with distinct applications in the medical field.
John Guttag is a professor at the Massachusetts Institute of Technology and leads the Data-Driven Inference Group within the Computer Science and Artificial Intelligence Laboratory (CSAIL). The group is dedicated to researching the application of advanced computing technologies in the medical field. Current projects include predicting adverse medical events, forecasting patient-specific responses to treatments, developing non-invasive monitoring and diagnostic tools, and advancing telemedicine. VCBeat (WeChat ID: vcbeat) has summarized Professor Guttag’s insights to explore the implications of these two types of machine learning for healthcare institutions, as well as the prerequisites for their implementation.
Guttag believes that although the impact of machine learning has not yet significantly disrupted the industry, its potential is enormous. At its foundational level, machine learning involves uncovering insights from data that are not readily apparent. For example, applying machine learning to patient data from those infected with Zika or other viruses can help identify optimal treatment strategies based on past cases, which can then be applied to future clinical care.
Typically, people use machine learning to build reasoning tools. It helps researchers discover patterns from existing data, enabling them to infer useful information when new data becomes available. Unlike human intuition, machine learning is entirely data-driven.
Let us now examine the importance of supervised learning and unsupervised learning in healthcare, respectively.
In supervised machine learning, the data and certain outcomes associated with the data are known. Taking the Zika virus as an example, if researchers possess information on all patients infected with Zika, they can determine which mothers gave birth to infants with congenital defects. Based on this, researchers can build a model to calculate the probability that a mother infected with Zika will give birth to a baby with congenital defects. Certainly, maternal age is also one of the factors affecting infant health; however, in machine learning models, labels are generated to mark various details about the mother and whether the infant is healthy or not. Therefore, the hallmark of supervised learning is labeling the outcomes of interest.
On the other hand, unsupervised learning operates without any labels. In unsupervised learning, researchers attempt to infer hidden structures from the raw data at hand. For instance, when initially presented with a batch of medical data, one may observe that patients are highly “similar.” Generally, the advantage of unsupervised learning lies in its ability to uncover unexpected insights. Therefore, unsupervised learning is particularly useful when data cannot be labeled for certain reasons.
Machine learning is the fastest-growing technology in computer science today. In recent years, as healthcare institutions have increasingly focused their research on big data analytics, precision medicine, and population health, machine learning, artificial intelligence, and cognitive computing are becoming increasingly valuable.
Although tech giants such as IBM, Google, and Microsoft have been continuously bringing their new technologies to market, it is those that have made significant progress in machine learning thatFinancial Services, Retail...and other industries, and this trend has persisted for approximately a decade. In this regard, the healthcare industry, which has always maintained a wait-and-see attitude toward new technologies, is indeedStarted Too Late。
The healthcare industry faces numerous challenges in adopting new technologies, one of which is the significant time lag between acquiring new technologies and applying them to medical practice. For this reason, Guttag is working to urge major healthcare institutions to more actively integrate machine learning into their current workflows. As he stated, “People should use today’s technology to address today’s needs. Machine learning is a remarkable technology that will undoubtedly bring about substantial changes to the healthcare industry in the coming years.”
Guttag and his students are working closely with Massachusetts General Hospital (MGH) to apply machine learning to clinical workflows in order to reduce hospital-acquired infections. For Guttag, even a small shift in the healthcare system is more impactful than ten purely theoretical papers. Their work at MGH is proceeding systematically, with the expectation of reducing the hospital’s rate of nosocomial infections within one year. Upon success, Guttag hopes to replicate and scale this approach across other healthcare organizations.
Modern healthcare institutions have access to more effective data collection technologies, while federal mandates on data disclosure have compelled hospitals to release previously confidential information, such as infection rates.
Critical MassThis is a prerequisite for putting machine learning into practice. For instance, a small hospital cannot deeply leverage its electronic medical record data. In the past, only a very limited number of hospitals had sufficient data to effectively deploy machine learning. However, the situation has now changed. First, healthcare systems are expanding rapidly, and independent hospitals will soon become obsolete. Alongside the growth of these healthcare systems, aggregated datasets spanning multiple systems are becoming increasingly abundant.
Another prerequisite is sound domain expertise. Machine learning demands a high level of specialized knowledge; unlike more mature technologies that allow for straightforward, user-friendly operation even with only a superficial understanding of the field, it does not. Currently, institutions planning to deploy machine learning must either possess robust in-house expertise or engage professional technical consultants at a cost. Of course, certain proprietary technologies available on the market can be highly beneficial for hospitals implementing machine learning and may also be considered for purchase.
Many companies claim to possess secret weapons in the field of machine learning. IBM’s Watson represents its most innovative breakthrough, with applications already explored in healthcare, finance, and the food service industry. In addition to holding several highly valuable technologies, Google has released a substantial number of tools to the public domain. In the future, machine learning technology is certain to improve continuously and is poised to achieve leapfrog advancements.