Even under the best circumstances, medical diagnoses and treatments are vulnerable to systemic and human errors. A recent technology to address this concern is machine learning, which uses large streams of data to develop complex mathematical models. Researchers wonder about alternative possibilities, including the use of machine learning to address systemic errors and check the reasoning of health care professionals.
Machine learning cannot replace physicians, who are able to use not only knowledge and experience, but also common sense, compassion and emotional intelligence in making decisions. Humans are also better than computers at perceiving small but meaningful nuances in information. So, while computers are no replacement for physicians, machine learning technology may be able to assist physicians.
What Is Machine Learning?
Most artificial intelligence (AI) depends on predefined parameters that operate within set rules, but machine learning actually provides an adaptive form of AI. As more information becomes available, the AI becomes better at analyzing and recognizing patterns. While this process is similar to how humans learn, computers require far more information to produce reliable interpretations, but the subsequent error margin is impressively small. In the context of medical care, machine learning models could process tens of millions of electronic medical records, which represents billions of data points — a feat no human physician could achieve. In this way, machine learning could provide insights a physician, even an experienced one, would not have.
Even the best health care professionals are limited by their experience and technical knowledge. With the exponential rate of scientific discovery, physicians need to work harder than ever to maintain clinical efficacy. At a certain point, it will become impossible for any physician to possess all the relevant clinical knowledge. In this situation, machine learning could become especially helpful. Even today, however, machine learning has the potential to reduce clinical errors by providing a check on individual practitioners who may have missed an important data point.
Machine Learning in Health Care Delivery
Machine learning could be especially helpful for processes that are challenging to most physicians, such as predicting a patient’s prognosis. So many factors are involved in this process, and the research data is lacking for some disease processes. Machine learning could analyze far more data points across time to offer more reliable prognoses. In addition, machine learning could aid in making diagnoses, especially for rare diseases that present with nonspecific symptoms. A quick, accurate diagnosis in this case could potentially lead to better clinical outcomes.
Machine learning may also play a role in connecting patients to the best treatments. Most modern health record systems already provide warnings when physicians try to prescribe suboptimal or contraindicated drugs, but machine learning could take this assistance further. With machine learning, a wider range of variables can be accounted for when determining the optimal treatment strategy.
In addition, machine learning could streamline clinical workflows. For example, machine learning could be used to simplify electronic records for a simpler presentation to clinicians. With a computer facilitating the process of reading and updating medical records, physicians and nurses could have more time with their patients, instead of typing information into a database. Furthermore, machine learning could help patients, especially those living in remote areas, by enabling the development of technologies that warn patients when their symptoms demand urgent attention and when they merely require a non-emergency appointment.
Limitations of Healthcare-Based Machine Learning
Of course, machine learning is not perfect. It’s important to keep expectations in check when considering how this technology could be implemented in the clinical setting. The technology will not solve all issues in healthcare delivery, nor prevent all mistakes in care delivery. Importantly, machine learning is only as good as the data it is given — a fact that underscores the need for high quality health care data. This limitation becomes especially pronounced when it comes to ethnically and racially diverse populations, because high quality medical data that includes these populations is particularly lacking. However, as more researchers become interested in population health outcomes, hopefully more data will be collected.
Other limitations of machine learning are directly related to its technical aspects. One issue is the fact that information between institutions, and even within them, is not easily exchanged because the data often exists in many different formats in different record systems. This creates a technical challenge in collecting data and building compatible systems for information exchange. In addition, privacy and security are major concerns. Health data is heavily protected by law, so any machine learning system would have to keep the information secure while processing it. Moving forward, it is important to recognize these limitations and address them in a way to instill confidence among patients and health care providers.