1. A Data Revolution is Coming - Revitalizing our AP data through Natural Language Processing and Federated Data Sharing
Rebecca Crowley Jacobson, MD, MS
Departments of Biomedical Informatics and Pathology
University of Pittsburgh
Michael Feldman, MD, PhD
Department of Pathology and Laboratory Medicine
University of Pennsylvania

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):
Rebecca Jacobson, MD, MS
Nexi, Inc
Shares
Shareholder/Consultant
Michael Feldman, MD, PhD
Inspirata - consultant, SAB
Philips - consultant
XIFIN - medical advisory board
Perkin Elmer - consultant
Virbio - advisory board

3. Q1. How are you currently accessing pathology data, images and biospecimens (FFPE, FF) for research and QI across your organization? Q2. How do you share and work with collaborators within and across institutions?

4. What is TIES? http://ties.pitt.edu
An NLP pipeline for de-identifying, annotating and storing millions of clinical documents
A system for indexing research resources (FFPE, FF, images) with document annotations
A system for querying large repository of annotated clinical documents and obtaining resources locally, using an honest broker model
A platform to support data and tissue sharing among networks of cancer centers and other institutions
Open source and freely available to not-for-profits

5. Radiology reports with DCIS in impression followed within 3 months by pathology reports with DCIS in final results

6. Pathology Report on patient meeting these criteria with NLP annotations

7. Response time over three retrievals#
Complexity
Query
Response time over three retrievals
Performance metrics
Number Reports Retrieved
Mean time to first results (sec) SD
Mean time to all results (sec)
Number of Reports or Report Sets (complex) Classified
Agreement TP FP
Precision 1
Low
Men, with prostatic adenocarcinoma on prostatectomy
1792
1.08
0.62
4.63
1.92 50
0.98 49 2
Women, with atypical endometrial hyperplasia
792
0.70
0.19 33
1.00 3
Patients, with phaeochromocytoma 54
0.95
0.31
0.96 4
Patients with hemangiosarcoma of scalp 17
0.49
0.13 5
Patients 10-30, with cystosarcoma phylloides 18
0.59
0.07
0.94 16
0.89 6
Patients with superficial spreading melanoma, metastatic
0.46
0.08 7
Patients with medullary carcinoma in thyroid gland 27
0.26
0.60 26 8
Patients with adenocarcinoma in brain
156
0.65
0.33
0.44 9
Men with invasive ductal carcinoma of breast 29
0.53
0.15 10
Patients, >60 with Hodgkins disease
549
0.64
0.17
0.84
0.22 34
0.68
All Low Complexity Queries
3439
0.67
0.20
1.07
1.26
329
308 21

9. TIES Functionality
NLP; Concept annotation with NCIM; ontology indexing with NCIT using Lucene
Infrastructure to code and recode; parallelize coders
De-identification, encryption, separation of PHI, auditing, X.509, quarantining
Honest Broker model built in to software. HBs see identifiers when working with investigator
Workflow to request FFPE, Frozen Tissue, Radiology Images
Virtual Slides
Support other datatypes (e.g. Cancer Registry data)

10. Technology Overview Three Tier client-server architecture
User Interface: Java Swing based UI client deployed using Java Webstart
Middle Tier: encrypted web services, processing pipelines
RDBMS: Supports MySQL or Oracle as the backend RDBMS data store
Primarily written in Java
GATE – General Architecture for Text Engineering based NLP pipeline
OGSA-DAI – Open Grid Services Architecture- Database Access and Integration – based web services
Apache Lucene index for search

11. Jacobson et al, Cancer Research 2015

12. Current TIES NLP Pipeline CHUNKER; SENTENCE SPLITTER
NOBLE Coder optimized terminology format
Terminology
Tokenizes words, punctuation, numbers and spaces
Marks synoptic text and eliminates from further processing
Partitions section text into sentences and phrases using ConTEXT stopwords
Cleans document, deletes existing annotations
NER
SYNOPTIC DETECTOR
CHUNKER; SENTENCE SPLITTER
RESETTER
TOKENIZER
NOBLECODER
ANNOTATION
TRANSFORMER
ConTEXT
Detects Negation, Temporality, Degree of Certainty
Organizes output annotations

14. Structured Data Support
Allows you to import data from data sources, e.g. Cancer Registry, Tissue Bank etc.
Combine search criteria across all data sources and the TIES concept coded reports.
Easy to setup and use. Ability to import from MS Excel/CSV/BSV or other delimited file formats.
You can also search data sets from other institutions.

15. Demonstration of the TIES system using 10K TCGA publically available cases (NLP processed path, CR, WSI) Get a demo account at

16. University of Pennsylvania Roswell Park Cancer Institute
U Pittsburgh
University of Pennsylvania
Roswell Park Cancer Institute
Georgia Regents/Augusta
Thomas Jefferson University (new)
Stony Brook University (soon)
……And others are preparing to join

17. Network Trust Agreements
Instrument of Adherence
IRBs agree that use of data for investigators is NHSR, no additional IRB protocol even for record level de-id data
Establishes governing body
Policies and Processes
QA and validation
User authorization
Auditing
Incident Reporting
Joining of new members
Governance

18. Jacobson et al, Cancer Research 2015

19. Use in Tissue Bank Honest Broker functionality is the key
Order biospecimens and images from within TIES, or export manifest for another system
Tags and Structured Data can be used to import information from LIMS, enabling search from within TIES
Whole Slide Images

20. Querying across TCRN

21. Being a node on the network
Regulatory foundation
What you need IT
Getting out the word
How will this help you and end users
Prep to research
Finding rare diseases
Temporal queries for cohort identification
Clinical service lines
Diagnostic rates across providers

22. Regulatory Foundation
All site IRB protocols approved
4+ million records
All site Network Agreements executed
Query across sites
Common Authorized User Agreement
Standard process to management users
Agreement to use NIH UBMTA
Getting material moved between sites couple weeks
Policy and process across sites
How we manage membership
95% of regulatory paperwork and effort covered
Prep to research no IRB
PI’s only need to get:
Join network as a research
Submit projects to access data across network
Approval of individual project to see deidentified reports
UBMTA to move tissue between sites

23. QA De-ID Review of QA results identified
Automated search identified 300 records out of 900K records
Manual scrub of records
Missed name (Full name, Lname, Fname) most common
Reports were manually scrubbed and returned into data set

24. Team to stand this up as a new member
Hardware – running on small VM slice and small amount of storage – 5-8K
Part of a system admin to manage and maintain TIES node, TIES coder and HL7 feed – 10K
Setting up the system –
HL7 feed from LIS – vary by site
Setting up database, coders, application layer
De-ID vs open source scrubber
Trainer
Web based – best on young folks
Hands on one on one – best for faculty

25. Prep for research Sarcoma Group
Osteosarcoma 83
Chondrosarcoma
141
Liposarcoma
318
Angiosarcoma
101
Leiomyosarcoma
467
Myxofibrosarcoma 89
DFSP 98
Fibrosarcoma
152
Ewings/PNET 37
MPNST 64
Sarcoma as search concept

26. “Breast Papilloma study”
Find all breast needle cores with diagnosis of a papilloma or papillomatosis
but nothing worse (atypia, DCIS, IDC, LCIS, ILC) at the time of the core biopsy who then went on to a subsequent resection
In the resection after the papilloma core biopsy, what is the frequency of finding either in situ or invasive carcinoma
Compare the carcinoma rate to carcinoma in a random core biopsy population with BIRADS4

27. Query
Breast
Core
Papilloma
Carcinoma
Atypia
Atypical hyperplasia

28. Result view

29. Research study Rosai Dorfman Disease
Rosai Dorfman disease is an idiopathic reactive condition characterized by exuberant macrophage reaction in lymph nodes or soft tissue
Etiology is unknown but some studies have implicated a virus, Herpes virus 6 in some cases
Pathochip is a microarray technology
All known pathogenic virus and bacteria and fungi arrayed
Allows FFPE to be probed for infectious signature in lesional tissue compared to normal controls
Metagenomic assay for identification of microbial pathogens in tumor tissues. mBio Sep-Oct; 5(5): e PMCID: PMC

30. Pathochip A. PathoChip a microarray-based
Probe sets for parallel DNA and RNA detection of viruses, bacteria, fungi, parasites and other human pathogenic microorganisms.
B. The current version of the PathoChip has 60,000 probes per array, representing all known viruses, 250 helminths, 130 protozoa, 360 fungi and 320 bacteria.
C. The array contains 2 types of probes: - Unique probes for each specific virus and microorganism, - Conserved probes which target genomic regions that are conserved between members of a family of viruses, thereby providing a means for detection of previously uncharacterized members of the family

31. Rosai Dorfman Query Penn – 10 cases identified
Tissue blocks have been cut and ready for extraction
Pitt – 40 cases identified
Review of cases showed 8 suitable for use
Case are on the way
Georgia Regents University
3 Case tissue sent
Control tissue will be reactive lymph nodes

32. Clinical Pathways Cancer Center
Cancer center needed quick way to find patients
All patients with urinary bladder resection (partial or whole cystectomy) with urothelial carcinoma
All prostate cancer patient with Gleason 6 on biopsy with no Gleason pattern 7 or more to see how many patients were offered active surveillance

33. TIES and the TIES Cancer Research Network
TIES Team
Girish Chavan
Eugene Tseytlin
Kevin Mitchell
Julia Corrigan
Liz Legowski
Adi Nemlekar
Yining Zhao
Vanessa Benkovich
Liron Pantanowitz
Rajiv Dhir
Penn
Michael Feldman Nate DiGiorgio
Tara McSherry
Joellen Weaver
GRU
Roni Bollag
Samir Khleif
Jennifer Carrick
Nita Maihle
And more…..
Roswell Park
Carmelo Gaudioso
Monica Murphy
Mayurapriyan Sakthivel
Amanda Rundell
Funding
NCI U24 CA Enhanced Development of TIES

34. Extra Slides

35. Data Processing
Report text is processed in four stages; each stage is handled by a separate data processing service. Each report has a specific status field in the database that indicates the stage of processing.
IMPORT
Read HL7 or delimited files into TIES schema
Support for multiple report types
Handle preliminary reports and addendums accordingly
Correctly assign reports to patients using a multi-attribute patient matching algorithm
DE-IDENTIFY
De-identify report and patient metadata and transfer to de-identified database
De-identify report text using third party de-identifier and additional de-identification scripts
Support for multiple de-identifiers including, DeID and Harvard Scrubber
CODE
Detect report sections using pre-configured section headers
Detect concepts in text using the NobleCoder algorithm using NCIM terminology
Detect negation in text using ConText algorithm
Store annotated document in GATE xml format
INDEX
Index concepts and text using Apache Lucene
Use NCIT ontology hierarchy to enhance index with ancestry information

36. High throughput coding using Java Messaging Service(JMS)
Each TIES coding service can be configured to run multiple processes internally to utilize multi-core CPUs effectively
Additionally, TIES can use Java Messaging to utilize multiple servers for coding a large dataset. This reduces the load on the database server by using a JMS provider like ActiveMQ to act as intermediary
CODING SERVER
CODING SERVER
PRODUCER
Apache ActiveMQ
TIES DATABASE
CONSUMER
CODING SERVER
CODING SERVER

37. Multi-layered approach to data security
TIES separates the PHI and de-identified data into separate databases that can be hosted on different servers for additional protection
OGSA-DAI grid services encrypt all communication between the client and servers using RSA-1024 encryption
Role based access control allows for data access granularity at three different levels
Users can quarantine any reports containing PHI, which immediately hides that report from all users until an QA admin reviews it
All queries and document views are logged by user and study. Auditing view lets you easily retrieve past activity for auditing purposes

38. Authentication and Authorization
Authentication happens at user’s institution
Authorization happens at Hub server for the network
After successful authentication, X.509 proxy certificates with a 12 hour validity are generated and used to communicate with any nodes in the network
Services are further secured using gridmaps that only allow specific individuals to access them

39. Structured Data Support
Patient
Pathology Report
Patient Or Report
Cancer Registry
Tissue Bank
Dataset
ER Status
PR Status
Materials Available
IPOX Stains
No. of lymph nodes
Text Attribute
Numeric Attribute
Category Attribute
Boolean Attribute
Disease free survival
Recurrence
Materials Available
IPOX Stains …