Paper #1: Embedding electronic health records onto a knowledge network recognizes prodromal features of multiple sclerosis
and predicts diagnosis
C. A. Nelson, R. Bove, A. J. Butte and S. E. Baranzini
Journal of the American Medical Informatics Association, 00(0), 2021, 1–11 https://doi.org/10.1093/jamia/ocab270
Bottom Line:
Two major goals of modern medicine are to prevent disease and to treat individuals in a personalized fashion, a fashion that acknowledges an individual’s unique constellation of genes, habits, and environment. While there are ways of reducing the risks of developing MS, such as maintaining high normal levels of vitamin D early in life and avoiding obesity, no two individuals with the illness are identical, either in clinical course or in response to disease-modifying therapies. As a result, providing personalized care to persons with MS is a challenge. Several physician-scientist groups published paper suggesting that there is a “prodromal” phase to MS, that is a phase in which individuals who will eventually be diagnosed with MS present with symptoms and medical difficulties long before there is clear evidence of neurologic dysfunction. Determining what these early or “prodromal” features are may allow early intervention, such as advising changes in behaviors and possibly even initiation of treatment. Determining prodromal symptoms may also provide clues as to what triggers MS, an illness, I believe, may have different causes.
Identifying prodromal MS may be possible utilizing techniques of the burgeoning field of “systems medicine.” I’ll explain this in greater detail in the following section. Using the powerful techniques of systems medicine the authors of the featured paper were able to identify prodromal signs and symptoms in over 5,000 persons with definite MS. This was done by searching electronic health records (EHR)and other biologic data of persons with MS before and after their diagnosis of MS. Using statistical and mechanical learning tools the authors evaluated clinical symptoms, medications, genes, and other biologic characteristics to look for correlations or associations between these variables. They compared results with a control group of over 2 million individuals who did not have MS. Going back as far as 7 years prior to the diagnosis of MS, strong and distinct correlations were found between the above variables that allowed the identification of persons who went on to develop MS year later. At three years prior to MS diagnosis the accuracy of prediction improved to more than 80%.
The techniques used in this study are not yet widely available and continue to evolve into even more powerful tools. Once the methods and techniques of systems medicine become more widely implemented, they will have the potential to revolutionize the care of persons with MS as well as those with other illnesses.
Some Background Information:
What is systems medicine? Definitions vary, but essentially it involves utilizing techniques from the fields of engineering, physics, genetics, statistics, and mathematics to determine significant correlations and interactions between a particular disease and a wide spectrum of clinical and biologic variables. Variables can include genetic makeup, patterns of immune response, metabolism pathways, patterns of protein synthesis, environmental factors, drug utilization, and clinical symptoms. With the aid of artificial intelligence and mechanical learning techniques networks are generated showing the degrees of interactions between variables. Some interactions are very infrequent. Some are very robust. The stronger the levels of interaction between variables, the greater the probability that such interactions are important in the disease process. Since the search for interactions or relationships is not based on any theory or hypothesis, unexpected and surprising interactions can and have been discovered. Several have led to new understandings of disease processes and the development of new treatments. There are several good introductions and reviews of systems medicine.
Key Points Related to the Featured Paper:
1. The researchers of the featured paper used a systems medicine approach to identify characteristics of persons with MS that occurred prior to their diagnosis. Their goal was to see if there were “prodromal” features of the disease that would allow them to predict the development of MS.
2. They examined electronic health records (EHR) of 5752 persons with definite MS diagnosed and treated at the University of California in San Francisco (UCSF). They focused on individuals who had received medical care in the years prior to their diagnosis of MS, care that was provided by either or both primary care physicians and/or specialty physicians. When possible, the researchers went back as far as 7 years prior to diagnosis, reviewing the reasons for the visits, their outcomes and treatments. Records from more than 2 million non-MS individuals were used as controls.
3. They then summarized the clinical features of the two patient populations as described in their electronic health records and embedded them into a huge network of basic science information called SPOKE. SPOKE contains massive amounts of data on genes, immunologic functions, multiple metabolic functions and pathways, and drugs and drug interactions.
4. The researchers then used a technique commonly used in engineering, called mechanical learning to randomly look for interactions and correlations between an individual’s clinical features and the vast amount of basic science data present in SPOKE.
5. The resulting networks of interactions showed, as expected, strong correlations and interactions between the clinical features of persons subsequently diagnosed with MS and features associated with changes in myelin, in brain and muscle functions, and with particular patterns of immune response. Unexpectedly, the networks also showed correlations with genes related to other central nervous system diseases. Non-MS individuals did not show these correlations.
6. Using these outcomes, the researchers were able to predict the development of MS with up to greater than 80% probability within three years of diagnosis. The predictive abilities increased as information was obtained closer to the time of diagnosis and with information obtained from health records of neurologists compared to health records prepared by primary care physicians.
7. The authors note that this was a retrospective or backwards-looking study with great variability in the data obtainable from the electronic health records. They propose that if, in the future, such records incorporate more specific and definable patient data, the accuracy of their predictive algorithms will be even greater, further increasing the likelihood of providing truly personalized, preventive health care, not just in persons with MS, but to persons with a wide variety of other illnesses.
8. The use of systems medicine approaches to study diseases, disease processes and treatment outcomes has the potential to revolutionize medical care in the next several decades.
Discussion:
Providing individualized, that is personalized, preventive health care has now become an achievable, though yet not an attained, goal of modern medicine. With the advent of digitized electronic health records (EHR) and the availability of huge, digitized libraries of biologic data, the use of mathematical, statistical, and computerized techniques from the fields of engineering and physics has permitted an entirely new approach to the analyses of disease. Whole genome sequences are available for hundreds of thousands of individuals, both normal populations and populations with a multitude of different illness. Digital libraries of medications, their side effects, and modes of action, are available as are large libraries of metabolic pathways, proteins, protein interactions, and detailed analyses of immune functions. With the help of mechanical learning techniques and artificial intelligence algorithms researchers are now able to search these databases in an unbiased fashion, looking for interactions and associations not previously detectable. Such approaches already resulted in the discovery of unexpected pathways in persons with allergies and in the testing of new medications.
Using the above noted methods researchers at UCSF were able to discover relationships between the clinical features of persons with MS as described in their EHR and thousands of genetic, metabolic, chemical, and immunologic factors. Analyzing the frequency and strength of the interactions with mechanical learning algorithms the researchers were able to identify patterns of interactions that were different in persons with MS compared to controls. These analyses of interactions allowed the identification of persons with MS up to 7 years prior to their actual MS-related presentation.
The researchers acknowledge that a shortcoming of their study was its retrospective nature and the great variability in the quality and content of the EHR. However, with the insights learned from this study, specific data can now be entered prospectively into an individual’s EHR that should allow greater accuracy in defining studied variables. This could lead to greater insights into unrecognized disease processes of MS with the potential to discover new disease pathways and new treatments.