1. Exemplars in Computational Pathology
U. Balis – PI 2016
Exemplars in Computational Pathology
Ulysses G. J. Balis, M.D.
Professor of Pathology &
Director, Division of Pathology Informatics
Director, Computational Pathology Lab Section
University of Michigan Health System
Pathology Informatics 2016 Annual Meeting May 25, 2016

2. Notice of Faculty Disclosure
In accordance with ACCME guidelines, any individual in a position to influence and/or control the content of this ASCP CME activity has disclosed all relevant financial relationships within the past 12 months with commercial interests that provide products and/or services related to the content of this CME activity.
The individual below has disclosed the following financial relationship(s) with commercial interest(s):
Ulysses G. J. Balis, MD
Inspirata, Inc.
Scientific Advisory Board
Stock
U. Balis – PI 2016

3. 1. An approach to diagnosis that incorporates multiple sources of raw data (eg, clinical electronic medical records, laboratory data including ‘‘-omics,’’ and imaging [both radiology and pathology imaging]); extracts biologically and clinically relevant information from these data; uses mathematic models at the molecular, individual, and population levels to generate diagnostic inferences and predictions; and presents this clinically actionable knowledge to customers through dynamic and integrated reports and interfaces, enabling physicians, patients, laboratory personnel, and other health care system stakeholders to make the best possible medical decisions
2. More generally, using computation for the interpretation of multiparameter data to improve health care.

4. Emergence of the Computational Pathology Lab Section
A multi-disciplinary team consisting of pathologists, data scientists, pathology informaticians, bioinformaticians, statisticians, application programmers and user experience (UX) engineers working together the develop and then deploy clinically actionable tests, based upon modeled, inferenced, extrapolated, interpolated and imputed transformations of primary clinical results from existing laboratory data.
Provides value-added, clinically actionable information to clinicians, often via reflexively ordered and resulted computational assays.
Offered tests are subject to the same level of requisite validation and quality engineering as conventional bench-based assays.

5. Emergence of the Formally Trained Computational Pathologist / Laboratorian
With the emergence of increasingly computational and multi-dimensional bioinformatics data that can be constitutive to the rendering of diagnoses, a competent and established cohort of Computational Pathologists will become essential.
Although molecular pathology trained individuals would appear to be adequate for this role, there are additional skill sets in information theory and data science that are needed

6. Imputation
The process of replacing missing data with substituted values. When substituting for a data point, it is known as "unit imputation"; when substituting for a component of a data point, it is known as "item imputation".
U. Balis – PI 2016

7. Population(s)
In the setting of the larger context of personalized and precision diagnostics, the data adjacency problem becomes daunting to detection by human cognition alone, if encoded data is indeed present.
From: Athey and Omenn, 2010

8. Assessing presence of absence of disease states distributed between a multitude of deviated analytes is not trivial in the setting of the data being encoded among many discrete data elements.

9. Supercomputing Threshold
Population-Level NGS Time-Series Data
Supercomputing Threshold
NGS Time-Series Data
Single NGS Data Set
Library of Whole Slide Images
1020 data elements
1015 data elements
Single Whole Slide Image Data
1012 data elements
250 Data Elements: encoded data including prognostic data likely present
105 data elements
1011 data elements
28 Data Elements: encoded data present beyond human cognitive limit
Increasing Complexity
Expression Array Data
108 data
elements
data elements
Tissue Microarray Study Expression Data
7 data elements: limit of experiential threshold of encoded data extraction
Time-Series Routine Lab Studies
Comprehensive Chemistry + CBC
1 data element: Simple linear inference model
Chem 7
Threshold for complete cognitive data extraction
Single Analyte

10. The Operational Challenge of Making Encoded Data Clinically Actionable
It is becoming clear that our existing routine laboratory results may contain many levels of “encoded” data, with this information being: real, distinct, diagnostic, prognostic and theranostic.
At present, there is no systematic approach to:
Orchestrate the automated capture of such data in the LIS itself
Implement validated computational solutions and rule sets
Easily reproduce such capture models across disparate LIS architectures
Obtain constitutive input data electronically from requesting laboratories
Reflexively forward integrative results to downstream repositories (e.g. data consumers) in a manner that preserves semantic layers of information

11. Typical Ordering Paradigm
Pre-analytical
Initial Primary Tests Ordered
Specimen Forwarded to Lab
Specimen Accessioned and Set up for Analysis
Analytical
Specimen Analyzed
Post-Analytical
Primary Results Reviewed and Verified
Primary Results Released to Clinical Systems
U. Balis – PI 2016

12. Extended Computational Pathology Ordering Paradigm
Computational Server
LIS
Pre-analytical
Reflexive Ordering of Assay
Initial Primary Tests Ordered
Primary Results Extracted to
Computational Pipeline
Specimen Forwarded to Lab
Specimen Accessioned and Set up for Analysis
Analytical
Computational Analysis
Specimen Analyzed
Post-Analytical
Computational Results
Reviewed and Verified
Primary Results Reviewed and Verified
Primary Results Released to Clinical Systems
Computational Results
Released to Clinical Systems
U. Balis – PI 2016

13. Exemplar 1 (a Multi-Analyte Assay with Algorithmic Analysis): Thiopurine Metabolite Imputation
Inflammatory Bowel Disease (IBD) affects at least 1.6 million individuals in the US
Current first-line therapy includes use of biological agents (immunotherapy)
>$30,000 per year in therapeutic agents
>$5B expense for pharmacotherapy
Second-line therapy (thiopurine analogs) is far less expensive but difficult to manage clinically
Current bench-based molecular metabolite testing available but results are expensive and questionable
6-MT therapy cost as low as $500 per year (60x less expensive)
Consequently, there is strong motivation to develop an improved test for clinical management of the IBD patient

15. An ideal candidate for cloud-based LIS computing
An Exemplar of Encoded Data Use in Diagnostics: Computationally-Based 6-MP Testing
Entirely computational, based on previously ordered tests (CBC & Comprehensive chemistry panels)
No actual bench testing
Better predictive power than the molecular bench-based test
Heavy computational burden for initial validation and continued performance certification
Adds ability to predict
patient compliance
detection of metabolic shunting (toxicity)
An ideal candidate for cloud-based LIS computing

17. A rules-based clinical algorithm: the ideal situation for automation in clinical reporting

18. Variable Importance of Each Analyte for an Overall Clinical Response Algorithm

19. Comparing Machine Learning Algorithms to 6-TGN
Notices that the green lines shows us an ROC of 0.59
And the Orange line an area under the curve of 0.85
This is highly significant with 0.001

20. How ThioMon Works Integration with LIS orders and primary results
CBC and Comprehensive Chemistry results constantly monitored for new paired entries
Paired Results reflexively order a ThioMon Computational Assay
The Primary results are looped to an LIS-attached R server
11,500 optimized / validated random forest rules on the R Server generate a Computational Report
The Computational report is returned to both the LIS and to a separate Web Server / Web Services Layer
The clinician is then able to review the result

21. The ThioMon Test in Actual Clinical Use
Random Forest
Algorithm on R Server
Clinical Order for
CBC and
Comprehensive Chemistry
(written or electronic)
Bench Testing
(<3 hours)
2.1 seconds
Decision Support within minutes

22. Present State of Reference Lab Thiomon Testing
Clinical Order and
Transport of Specimen
Reference Clinical Laboratory
Pathology Informatics
Data Center
With the current service model, external patient blood specimens must be physically transported to the U-M clinical Lab
This is inefficient
A better model would be to simply transport the CBC and Chemistry electronic results data (concept of eTubes)
This required a split study to demonstrate equivalence
Final Delivery of
Computed Result

23. A Split Study to Demonstrate Utility of ThioMon Testing via the eTubes Concept
Clinical Order
Remote Pathology
Clinical Laboratory
Initial Lab Results from External site electronically transferred to the central eTubes server
UMHS Pathology
Clinical Laboratory
Initial Lab Results from local lab electronically transferred to the central eTubes server
Final Comparison of
Computed Results Between
Remote and local Locations
2.1 seconds

24. ThioMon Testing: Realized Features and Benefits
More predictive than the current gold standard test (Prometheus Labs)
Immediate resulting capability from routine ordered CBC and comprehensive chemistry panels as input data (no additional bench testing needed)
Web Infrastructure combined with the eTubes concept makes this technology available on a global basis, without transport of specimens
Operational resources required are very modest and all open-source
R server (virtual machine)
Web Portal (virtual machine)
0.1 FTE for website and data pipeline maintenance

25. Current Use In use for past 3.5 years
Implemented in the SCC-Soft Lab Information System
Fully interfaced to the local EMR (Epic MiChart)
Date/Time
SHUNT
CLINRESP
NONCOMP
THIO INTRP
10/05/2010
1.4
80.6
0.4
GOOD: see text
THIO INTRP
Interpretation: Assuming this patient has been on the same dose of thiopurine medication for at least 4 weeks and is at steady state: This patient has had a good hematologic and chemistry response to thiopurines, has a low probability of shunting, and a low probability of noncompliance. If this patient has not obtained a good clinical response, they may have a non-inflammatory cause of symptoms, or if inflamed, may need a different form of therapy, as dose increases are unlikely to produce large therapeutic gains.

26. Thiomon results, now rendered as a webpage:

27. Initial Findings With 3.5 Years of Use:
Performance of ThioMon greatly exceeds the bench-based molecular test, in terms of predicting biologic response
Adoption of use has increased by 100-fold since the third publication
Now possible to deploy the test as an orderable item by other institutions, as a web-based, eTubes test
Widespread conversion to 6-MT based therapy could save as much as $4.5B annually, in the US alone
Computational assay a fraction of the cost of the bench-based assay.

28. Exemplar 2 (a Multi-Analyte Assay with Algorithmic Analysis): C
Exemplar 2 (a Multi-Analyte Assay with Algorithmic Analysis): C. Difficile exacerbation prediction for patients admitted from the Emergency Department
Hypothesis:
Identify patients who are most likely to be chronic carriers of C. difficile who will exacerbate with infectious colitis, upon admission from the ED
Input variables:
Age
Gender
CBC
Comprehensive Chemistry
Pilot study findings
96% sensitivity and 81% specificity (AUC 0.78)
Larger cohort now under investigation

29. Exemplar 3 (a Multi-Analyte Assay with Algorithmic Analysis): Prediction of Progression to Hepatocellular Carcinoma for patients with Chronic Hepatitis C
Observation
With the availability of treatment for Hepatitis C, there is a need for stratification of patients selected for initial therapy, owing to significant expense (>$80k for a course of therapy)
Hypothesis:
Identify patients who are most likely to progress to harboring malignancy, thus identifying the cohort in most urgent need of immediate treatment (ethical principle of beneficence)
Test required only:
Age
Gender
CBC
Comprehensive Chemistry
AFP
Pilot study findings
94% sensitivity and 98% specificity (AUC 0.94)
Larger cohort (VA hospitals) now under investigation

30. Closing Thoughts
Computational pathology is no longer a conceptual approach
Many diagnostic targets remain unexplored
Tools are now sufficiently mature to address the significant computational and operational requirements

31. U. Balis – PI 2016