Methods to Identify Patient Clusters and Build Precision Analytics for Diagnosis, Prognosis, and Treatment

Nicole Meister1, Hannah Cowley1, Corban Rivera1, Karla M Gray-Roncal1, Kathryn Fitzgerald2, Claudia Allshouse3, Anna Duerr3, Aalok Shah3, Paul Nagy3, Peter A Calabresi2, Antony
Rosen3, Ellen M Mowry3 and William R Gray-Roncal1
1Johns Hopkins University Applied Physics Laboratory, Laurel, MD; 2Johns Hopkins University School of Medicine, Baltimore, MD; 3Johns Hopkins University, Baltimore, MD

BACKGROUND
• Precision medicine promises great advances in the treatment of MS
through leveraging data science techniques for rapid diagnosis and
targeted treatment
• Disparate datasets lead to challenges in creating large datasets (Fig.
1) and a lack of a standardized data science framework makes it
challenging to extend or explore various approaches
MS PATHs
OBJECTIVE
RESULTS
• We provide an initial toolbox to deploy data science methods across various patient data,
providing access to standardized models, novel features, and visualization (Fig. 3)
• Clustering based on clinical expertise and machine learning
methods helps researchers find sub-cohorts that may be
used in prediction and treatment (Fig. 4)
• Our packages allow for quickly switching between research
questions and models; we demonstrate these tools to predict
25-foot walk time scores and Patient-Determined Disease
Steps (PDDS) (Fig. 5)
A. Histogram of PDDS score absolute errors (n=1653)

• Develop a proof-of-concept toolkit for data

EHR OCT

Cohort percentage
fusion, quality assessment, sub-cohort Data Results
identification, and predictive analytics
Figure 3: Pairwise relationships between selected variables illustrate the
• Lower the barriers to adoption and problem complexity and opportunity when exploring features and patient
facilitate data science MRI Results information at scale
A. Clinician-selected clusters B. Data-driven clusters in PCA space
Figure 1: Creating enriched datasets PDDS error
allows for more accurate prediction of B.
METHODOLOGY MS outcomes Histogram of 25-ft walk time absolute errors (n=1653)

• Our tools are organized into easy-to-use python packages to be

Cohort percentage
deployed in a Jupyter notebook environment
• We support both data-driven exploration and clinician-guided
discovery to enable flexible and iterative experimentation (Fig. 2)

Figure 4. K-Means clustering to discover sub-cohorts using clinician-defined

Time (seconds)
and data-driven features. A) Visualization of 3 sub-cohorts identified using the
clinician-selected features of cognitive function, years of education, and fatigue Figure 5. Using our tools and the JHU
scores. B) Visualization of 5 sub-cohorts discovered using data-driven clustering. MS PATHS cohort, researchers can
model and visualize their results for two
DISCUSSION complementary research questions. (A)
Logistic regression modeling of PDDS
• We present an initial suite of tools to support cohort score (absolute error) (B) Linear
discovery and predictive analytics for precision medicine regression model predicting 25-ft walk time
• Using Generalized linear models provides accurate (absolute error)

predictions of 25-ft walk time and PDDS score as an engineering example

• Future work will package and extend these tools to a larger clinical research community to
Figure 2: Two synergistic approaches capture and extend clinical insights through hypothesis and facilitate extensible and reproducible analysis
data-driven approaches
Acknowledgments: This work was supported by internal research, Biogen, JH Precision Medicine, MS Center of Excellence and NIH R01NS082347.