Skip Navigation
request an appointment my chart notification lp musc-logo-white-01 facebook twitter youtube blog find a provider circle arrow
MUSC mobile menu


An MUSC blog
Keyword: machine learning

MUSC Health Rheumatologist Dr. Jim Oates


Summary: Medical University of South Carolina (MUSC) investigators report preclinical research showing that prognostic models for lupus nephritis that include novel biomarkers have significantly improved predictive power over models using only traditional markers, in the August 2016 issue of Arthritis & Rheumatology. Data reveal that chemokines, cytokines, and markers of cellular damage were most predictive of patients' therapeutic response. This is a critical first step to developing clinically meaningful, decision-support tools in lupus nephritis.



Caption: MUSC Health rheumatologist Jim C. Oates, M.D.

Results of preclinical studies by investigators at the Medical University of South Carolina (MUSC) reported in the August 2016 issue of Arthritis & Rheumatology demonstrate for the first time that including novel biomarkers in lupus nephritis (LN) prognostic models significantly increases their power to predict therapeutic efficacy. Identifying biomarker models with sufficient predictive power is a critical step toward developing clinical decision-making tools that can rapidly identify patients who require a change in therapy and potentially reduce onset of renal fibrosis during induction therapy.

Approximately half of all patients with systemic lupus erythematosus (SLE) develop LN, an immune complex-mediated glomerulonephritis. Lupus nephritis, in turn, leads to renal failure in up to 50% of patients within five years. American College of Rheumatology guidelines recommend changing LN treatment after six months of induction therapy if response to therapy is not achieved. However, 'response to therapy' is not clearly defined and renal damage can occur during the six-month induction period.

Currently, clinicians monitor response to treatment via blood pressure measurements, serum complement levels, anti-double-stranded DNA (anti-dsDNA) antibody levels, urinary sediment, urinary protein-to-creatinine ratios, and surrogates of renal function. Unfortunately, predicting disease progression is difficult using these traditional biomarkers due to their low sensitivity and high LN heterogeneity at presentation. Even when machine learning models are employed, traditional biomarkers are only 69% accurate in predicting a LN diagnosis among SLE patients. There is a need for individualized, decision-support tools that can better define 'therapeutic response' at the start of therapy and allow clinicians to tailor induction therapy to disease severity to prevent renal damage and unnecessary drug toxicity.

"We saw our colleagues' frustration in trying to come up with predictive models,” said Jim C. Oates, M.D., Associate Director of the MUSC Clinical and Translational Research Center, Associate Professor of Rheumatology, and senior author on the article. “The traditional markers we use in clinic today have quite limited predictive capacity. All lupus patients have varying degrees of kidney damage and levels of involvement of the different kidney structures. So, we wanted to account for this heterogeneity and the stages of disease progression. We wanted to include markers for pathways of inflammation as well as for damage."

The research team hypothesized that a targeted panel of urinary biomarkers reflecting initial resident and inflammatory cell activation (cytokines), signals for homing to the kidney (chemokines), activation of inflammatory cells (growth factors), and damage to resident cells, combined with artificial intelligence/machine learning modeling, might provide an early LN decision-support tool that could predict outcomes better than standard biomarkers alone. The team also chose to assess urine biomarkers rather than serum/plasma markers to increase the tool's sensitivity and specificity to signals of renal (rather than systemic) processes.

Urine samples from 140 patients with biopsy-proven LN who had not yet started induction therapy were analyzed for a panel of novel biomarkers using pre-mixed, commercially available kits. Univariate, receiver operating characteristic (ROC) curves were generated for each biomarker and compared to ROC area under the curve (AUC) values from machine learning models developed using random forest algorithms. Outcome models using novel biomarkers plus traditional clinical markers demonstrated greater AUC and significance compared to models developed with traditional markers alone ([AUC 0.79; P<0.001] vs. [AUC 0.61; P=0.05], respectively). The combined models also demonstrated greater power to correctly predict LN therapy outcomes (responder versus non-responder) than models using only traditional markers (76% vs. 27%, respectively [p<0.002]).

The team identified chemokines, cytokines, and markers of cellular damage as most predictive of LN therapy response. Race, anti-double-stranded DNA antibodies, and induction medication did not significantly contribute to the model.

"We were somewhat surprised by some of the analytes that were important in the model,” said Oates. "One traditional marker, protein-to-creatinine ratio, was the third most important, and a standard kidney function measure was the ninth. I was also surprised to see interluekin-8 so high. This is in keeping with recent publications highlighting the importance of neutrophils in the pathogenesis of lupus, however."

Including multiple mechanisms of disease pathogenesis and cellular damage likely provides a more effective diagnostic approach by better reflecting the multi-stage, heterogeneous nature of LN. This is the first study to combine a broad biomarker panel with machine learning techniques to optimize disease outcome models. "This could apply to any model where there is kidney inflammation leading to damage,” said Oates. "It's proof of concept for other kidney diseases that you can take a discovery model and incorporate machine learning to develop and validate predictive models."

The team is now testing other biomarkers and applying the model in a larger patient population to ensure external validity and improve power. They are also exploring other inputs.

"Our next approach is to harness existing data in the medical record to enhance predictions,” said Oates. “This is much more immediately translatable in the clinic than getting through a long FDA validation process and the industry pipeline. Using medical record data is cheaper, and there are patient and system factors in the medical record that you can't measure with an assay, such as economic and societal disparities, which affect outcomes. This approach could also be used to enhance biomarker predictive models”

Predicting speech fluency after stroke. Brain images showign features of damage to grey-matter and white-matter regions of brain, reflecting their importance in predicting speech fluency.

Image Caption: Predicting speech fluency after stroke. These are features of gray-matter cortical regions (left) and white-matter tracts (right), reflecting their importance in predicting speech fluency scores. Regions/connections are marked in red when they strongly influence speech fluency, in blue when their influence is moderate, and are left uncolored when the influence is weak or non-existent. Image used courtesy of Dr. Leonardo Bonilha and Dr. Grigori Yourganov of the Medical University of South Carolina, who own the copyright for the image. Published in the June 22 issue of the Journal of Neuroscience (doi:10.1523/JNEUROSCI.4396-15.2016).

Loss or impairment of the ability to speak is one of the most feared complications of stroke—one faced by about 20% of stroke patients. Language, as one of the most complex functions of the brain, is not seated in a single brain region but involves connections between many regions.

In an article published in the June 22, 2016 issue of the Journal of Neuroscience (doi:10.1523/JNEUROSCI.4396-15.2016), investigators at the Medical University of South Carolina (MUSC) and the University of South Carolina (USC) report that mapping all of the brain’s white matter connections after stroke, in addition to imaging the areas of cortical tissue damage, could better predict which patients will have language deficits and how severe those deficits will be. The totality of the brain’s connections is referred to as the connectome.

“Imaging the connectome of patients after stroke enables the identification of individual signatures of brain organization that can be used to predict the nature and severity of language deficits and one day could be used to guide therapy,” said MUSC Health neurologist Leonardo Bonilha M.D., Ph.D., senior author on the Journal of Neuroscience article, whose laboratory focuses on connectome imaging, particularly as it relates to language loss after stroke. Grigori Yourganov, Ph.D., is the first author on the article. Julius Fridriksson, Ph.D., Chris Rorden, Ph.D., and  Ezequiel Gleichgerrcht, Ph.D, aphasia researchers at USC who recently received NIH funding to establish a Center for the Study of Aphasia Recovery and who are long-time collaborators of the Bonilha laboratory, are also authors on the article.

This study is the one of the first to use whole-brain connectome imaging to examine how disruptions to white matter connectivity after stroke affect language abilities. White matter fiber tracts are the insulated wires that connect one area of the brain to others. White matter is named for the myelin sheaths (insulation) that cover the many axons (wires) that make up the fiber tracts.

“If you have two brain areas and both of them have to work together in order to carry out a function and the stroke lesion takes out axons that connect those brain areas—the two areas are intact but the communication between them is disrupted and so there is dysfunction,” said Yourganov.

Currently, structural magnetic resonance imaging (MRI) is used after stroke to assess lesions in the cortical tissue—the brain’s grey matter. However, the extent of cortical damage often does not correlate with the severity of language deficits.

“Stroke patients sometimes have significant impairments beyond the amount of cortical damage,” said Bonilha. “It is also hard to predict how well a patient will recover based on the cortical lesion alone.”

Could connectome-based imaging be a useful complement for assessing damage to the brain’s connections after stroke and for guiding rehabilitative therapy?

The study led by Bonilha took an important first step toward answering these questions. The study, which enrolled 90 patients at MUSC and USC with aphasia due to a single stroke occurring no less than six months prior, assessed four areas related to speech/language using the Western Aphasia Battery—speech fluency, auditory comprehension, speech repetition, and oral naming—as well as a summary score of overall aphasia. Within two days of behavior assessment, each of the patients underwent imaging studies—both T1- and T2- weighted MRI, typically used after stroke to map cortical damage, and diffusion imaging, used for connectome mapping. The team then used a type of machine learning algorithm—support vector regression (SVR)—to analyze the imaging results and make predictions about each patient’s language deficits.  In essence, an algorithm was created that could derive the WAB score from either a feature relevant to imaging of the grey matter damage by structural MRI or a feature relevant to connectome imaging of the brain’s white matter fiber tracts. The team used 89 of the 90 patients as training sets for SVR and then used the algorithm to predict language defect/preservation in the 90th patient. This was done for each of the 90 patients and, in each patient, for both features identified via structural MRI and connectome imaging.

The accuracy of the algorithm’s prediction of WAB score for each patient was then assessed by comparing it to the WAB score determined via behavioral testing. Connectome-based analysis was as accurate as cortical lesion mapping for predicting WAB scores. In fact, it was better at predicting auditory comprehension scores than was lesion-based imaging using structural MRI and only slightly less accurate at predicting speech fluency, speech repetition, and naming scores.

The study demonstrates that damage to the white matter fiber tracts that connect the brain’s regions plays a role beyond cortical damage in language impairment after stroke. Furthermore, this study also discloses that connections in the brain’s parietal region are particularly important for language function, especially fluency. This region is less likely to sustain damage after stroke, even in patients who experience aphasia, suggesting that damage or preservation of the brain’s connections in this region could play a key role in determining who will experience aphasia and who will have the best chances for recovery. The integrity of these connections could not be mapped with conventional structural MRI but can now be assessed through connectome-based analysis.

The study findings also suggest that connectome-based analysis could be used to inform a more individualized approach to stroke care. Because the algorithms developed using these study patients as the training set are generalizable to a broader stroke population, connectome-based analysis could one day be used to identify the distinctive features of each patient’s stroke—which connections have been lost and which preserved—and then the algorithm could be used to predict the type and severity of language impairment and the potential for recovery. This information could then be used to direct rehabilitative therapy to improve outcomes. 

“By mapping much more accurately the individual pattern of brain structural connectivity in a stroke survivor, we can determine the integrity of neuronal networks and better understand what was lesioned and how that relates to language abilities that are lost,” said Bonilha. “This is, broadly stated, a measure of post-stroke brain health. It is the individual signature pattern that could also be used to inform about the personalized potential for recovery with therapy and guide treatments to focus on the deficient components of the network.”

Subscribe to Progressnotes

Submit a Story Idea

Current Issue of Progressnotes

Digital EditionPDF | Home

Progressnotes - Spring 2017 cover thumb

Back Issues of Progressnotes

past issues of progressnotes