Connecting data for future insights · • Data governance and access model that balances privacy with data accessibility • Engaged and ‘data literate’ hospital community The

Connecting data for future insights

DATADEX 2019 © - Private and Confidential

Background

Francis Jeanson, PhD CEO & Scientific Director

Head of Datadex, a data augmentation company

Former Manager at the Ontario Brain Institute

AI Software Engineering

PhD, MSc, BSc in Cognition, Adaptive Systems, AI

Member of the GA4GH

Startup and Academic Advisor in Data Science and AI

DATADEX 2019 © - Private and Confidential Connecting data for future insights

DATADEX 2019 © - Private and Confidentia l Connecting data for future insights

Diagnostic errors occur in ~10% of case s 18 Million

PRIMARY CARE ERRORS PER YEAR IN THE US


Clinical trial success is diminishing… From 18% in the 90’s to 9% in the 2000’s

~11% increase per year…

Trials costs are increasing


Cost of healthcare is increasing faster than GDP GDP: 2% per year

Healthcare : 3% per year

Solut ions


Patient centered preventat ive care DATA: re al-tim e , we arab le s, te le m e d ic ine … AI assisted d iagnosis/ p rognosis DATA: b iom arke rs, e nvironm e nt, e thical, … Short cycle t ranslat ional re se arch DATA: sharing , s im ulation, se curity, p rivacy

Digit izat ion of healthcare


We are all eager to harness the power of AI of innovators are

using AI to identify opportunities in data.

61%


We are capturing more data than ever before but insights re m ain e lusive

of data are leveraged by organizations on average

5% LESS THAN


Your data Lake

No way to explore sensitive data before granting acces s

No governance over s ens itive data means increased risks

No ability to link cus tomer data without violating privacy

No way to grant access to jus t the data the analys t needs


Our solut ion


Our solut ion

Automated Data Indexing


Our solut ion


Feature exploration Data matching


Our solut ion



Big data joins & unions Private data linking


Our solut ion




Access management Fine grained sharing


Your insights




Access management Fine grained sharing


Healthcare insights with Datadex

“What is the likelihood of my patient being readmitted?”

EHR patient medical history database

LIMS blood data

Other clinic HL7 files

Wearable devices and mobile apps

Pharmacy medication database


Link Data Privately


Name Age Ph#

Susan Tao 23 +1 416 234-6343

Ethan Bim 45 234.543.2431

Pete Low 19 +1 647 994-5237

Name Outcome Score

Adas Ides re-admitted -3.4

Susan Tao Cured 4.1

Leor Pas a re-admitted -2.8

Name Drug Dosage Height

Adas Ides Lozepram 150 mg 6.3

Hani Ito As pirin 0.5 g 5.11

Susan Tao Ibuprof. 250 ml 5.6

Age Height Drug Dos age Outcome Score

23 5.6 Ibuprof. 250 ml Cured 4.1

Name

76WX-RT23

Name

Susan Tao

Hash Pre d ictors Outcom e s

AI

Fine Control led Sharing


ID Name Income Score

423 Adas Ides 89K 3.4

556 Suan Tao 143K 4.1

768 Leor Pas a 56K 2.8

As s ign digita l acces s policies for governance compliance

Auditable acces s agreements

Fine control over fields for sharing

Secure cloud or direct data exchange

Tangible Benefits


Standardized data

Increased sample sizes

Increased entity features

Controlled data access

Digital access policies

Improved statistics

Greater significance

Reveal new correlations

Stronger security

Future proof governance

ROI for Users


Over 10 times reduction in time to data access

Creat ing New Opportunitie s

Over 3 times increase of data value

Make your data FAIR ● Findable ● Acces s ible ● Interoperable ● Reusable

Build your innovations from within and dis s eminate them

Accelerate Innovation

Our Partners & Network


datadex.net [email protected] +1 647 217 2232 Toronto, Canada

DATADEX 2019 © - Private and Confidential

The Era of Artificial Intelligence in Healthcare

Muhammad Mamdani, PharmD, MA, MPH Vice President – Data Science and Advanced Analytics, Unity Health Toronto

Scientist – Li Ka Shing Knowledge Institute Faculty Affiliate – Vector Institute

Adjunct Senior Scientist – Institute for Clinical Evaluative Sciences Professor – University of Toronto

October 2019

The AI ‘Hype’ in Healthcare

What is Artificial Intelligence??

• Google dictionary

• the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages

How do computers ‘learn’?? Data… lots and lots of data

Unity Health Toronto: St. Michael’s Hospital

• St. Michael’s Hospital • Tertiary care teaching hospital in downtown Toronto • Established in 1892 with the founding goal of taking care of the

sick and the poor of Toronto’s inner city • 1 of two adult trauma centres in the GTA • 463 beds and numerous outpatient clinics

• > 6,000 staff • > 900 physicians; > 1,600 nurses

• Approximate annual patient volumes • > 75,000 ED visits • > 500,000 ambulatory visits • > 25,000 inpatient visits

• Research: Li Ka Shing Knowledge Institute & Keenan Research Centre

• > 200 investigators; > 800 staff

• Fully affiliated with the University of Toronto • Part of the Toronto Academic Health Sciences Network (TAHSN)

‘Advancing’ Advanced Analytics in the Hospital

Vision Infrastructure

People & Culture Process Lessons

The Vision

• Vision – What Do We Want to See? • Make data and analytics an integral part of clinical and management decision-

making to drive better patient outcomes and improve hospital efficiency

• Components of the Vision

• Readily available, ‘up to date’, high quality data • Highly skilled data scientists • Software and hardware to enable advanced analytics • Data governance and access model that balances privacy with data accessibility • Engaged and ‘data literate’ hospital community

The Advanced Analytics Platform at the Network:

Our Journey

The Data Infrastructure

The People

The Analytical Tools

Siloed Datasets that are not ‘analytics ready’ ‘Analytics ready’ consolidated data warehouse

Establishing a Data Science Unit LKS-CHART Data Scientists Computer Science: Machine Learning Engineering: Simulation modeling and optimization Statistics: Traditional biostatistics

Analytical Tools Designed for Advanced Analytics / Artificial Intelligence

Phase 1 of data warehouse completion:

March 2018

DSS

People and Culture: The LKS-CHART Team and Community

Presenter

Presentation Notes

We’re a bunch of data geeks who really like math. But we’re more than that. The reason why our math works so well is that we’re a community… the questions we work on don’t come from us – they come from our nurses, doctors, pharmacists, physiotherapists, hospital administrators, and patients. These very people who care for our patients and ARE our patients not only ask the questions, but they lead the projects with our support, so we as a community collectively own it. Why does our model work? We all work together because we have a common cause: life.

The Process • Project ideas / questions come from end-users (management and clinical) NOT from the LKS-CHART

• Projects the LKS-CHART will work on (from highest to lowest priority):

– Initiatives that have promise to improve patient outcomes AND reduce costs – Initiatives that have promise to improve patient outcomes at no additional cost – Initiatives that have promise to reduce cost but not adversely affect patient outcomes – All other initiatives that are prioritized by management and/or clinical leadership

• The majority of projects must have SMH leadership (management and/or clinical leadership) support

to enact CHANGE – Where relevant, projects will have an evaluation component to assess impact and ROI

• Target project duration: 3-6 months

• Current Project Portfolio: 25+ projects at any time

AI at St. Michael’s Hospital: Examples and Lessons

Predict 24H in Advance: ICU Transfer or Death

> 98% accuracy ↑ Patient Outcomes

Predicting When Patients Do Poorly

Emergency Department Wait Times

Forecasting patient volumes 3 days to 3 months in advance

94-96% accuracy

Optimizing Nurse Staffing

Optimize Nurse Resource Team Staffing at St. Michael’s and St. Joseph’s

$1 million cost reduction annually

↓ Wait Times ↑ Quality of Care

↓ Cost ↑ Quality of Care

Key Lessons: Human Factors, Visualization, and Workflow

Key Lessons: End-User Needs and Customization

Key Lessons: User Acceptance, Legal/Ethical, Implementation

Drug-Related Projects • IV to Oral Antibiotic Conversion - implemented

• Algorithm using structured data and natural language processing to asses clinical stability and absorption status of patients on IV antibiotics

• Automated daily list of patients eligible for conversion to oral antibiotics to pharmacists • > 95% accuracy

• Oral Anticoagulation (OAC) - implemented

• Algorithm using structured data and natural language processing to identify patients eligible for OAC but not receiving it

• Automated daily list of patients eligible for OAC to cardiology team • > 90% accuracy

• Warfarin Dosing Algorithm – under development • Reinforcement learning algorithm that examines individual patient characteristics, lab values

(including INR values), medications (including warfarin dose), orders (including diet orders), consults, vitals, and text notes every 6 hours to guide pharmacists on warfarin dosing

Discussion

Website: www.chartdatascience.ca Twitter: @Chart_DataSci

https://www.google.ca/url?sa=i&rct=j&q=&esrc=s&source=images&cd=&cad=rja&uact=8&ved=0ahUKEwjPsIyMrZjQAhWBcCYKHRKoAWYQjRwIBw&url=http://tedtalks.soup.io/tag/Design?since%3D324081956&psig=AFQjCNE78oT5DDnz_v3r6acQIKApemvG-g&ust=1478666742094780

http://www.chartdatascience.ca/

IBM Watson / © 2018 IBM Corporation | IASO-1747 Rev 1.0

The Era of Artificial Intelligence in Healthcare — Medical Imaging Marwan Sati Development Executive, Cognitive Clinical Advisors IBM Watson Health Imaging

Presenter

Presentation Notes

IBM's statements regarding its plans, directions and intent are subject to change or withdrawal without notice at IBM's sole discretion. Information regarding potential future products is intended to outline our general product direction and it should not be relied on in making a purchasing decision. The information mentioned regarding potential future products is not a commitment, promise, or legal obligation to deliver any material, code or functionality. Information about potential future products may not be incorporated into any contract. The development, release, and timing of any future features or functionality described for our products remains at our sole discretion.


Data explosion and industry challenges: Creating opportunities for AI More physicians are

experiencing burnout

51% of physicians experienced at least one symptom of burnout in 2016, a 25% increase in the last four years

Misdiagnoses entail huge costs for organizations

$4 billion is spent on false-positive mammograms in the US each year

Administrative tasks take up significant time

64% of radiologists’ time spent on non-interpretive tasks

Imaging is generating a huge volume of data

60 billion medical images were generated in 2015 across the U.S.


Sources: 1. http://www.cnbc.com/2015/04/06/breast-cancer-misdiagnoses-cost-4-billion-study.html; http://content.healthaffairs.org/content/34/4/576.abstract

2. http://www.jacr.org/article/S1546-1440(15)00196-9/fulltext 3.http://www.medscape.com/features/slideshow/lifestyle/2017/overview

4-5. http://www.ibmbigdatahub.com/video/ibm-big-data-minute-transforming-unstructured-data-better-healthcare-outcomes

Patient data is often unstructured

80% of patient data in organizations is unstructured, often lacking relevant context

Presenter

Presentation Notes

The explosion of data in our industry is creating challenges that are opportunities for AI. Misdiagnosis entails huge costs for organization (for example $4B spent on false-positive mammograms.) Administrative tasks are taking up significant time. (For example Radiologists spend 64% of their time not actually interpreting images.) More Physicians are overloaded and experiencing burnout at alarming rates (with 50% of radiologists experiencing symptoms of burnout that is up by 25% over the last 4 years.) This is due in part because imaging is generating a huge volume of data (60 billion images in the US alone in 2015). And patient data is often unstructured, about 80% is unstructured data. This data an administrative overload is forcing Radiologists to pick and choose where to focus their attention. Transition: so how can AI help?

The power of Watson


Human + machine = greater than the sum of its parts

Humans excel at:

AI systems excel at:

Common Sense Dilemmas Morals Compassion Imagination Dreaming Abstraction Generalization

Natural Language

Pattern Identification

Locating Knowledge

Machine Learning

Eliminating Bias

Endless Capacity

Presenter

Presentation Notes

We strongly believe that people working with cognitive computing technology can outperform people alone or computers alone. The trick is to identify what we humans do best and what computers do best. We humans excel at and enjoy a number of activities such as common sense, solving dilemmas, showing compassion, imaging dreaming. Whearas computers on the other hand are happy to fulfill a number of mundane tasks and won’t complain unless, of course, we program them to do so. For example NLP, looking for patterns, locating and consolidating patient information etc. Transition: So how to we fit AI’s value into everyday Radiology workflow?

AI in radiology workflow


Provide clinical context to radiologists by surfacing relevant patient data

Inform decisions by providing clinical insights* • Optimize workflow • Flag key findings • Automate tedious

tasks

Highlight gaps in care by empowering retrospective review • Problem list gaps • Missed findings*

Analysis of structured and unstructured data

*This technology is in the research and development phase and has not been evaluated by any regulatory agencies (such as USFDA) for safety or efficacy. It is not available for any commercial or non-commercial use. Information about R&D stage technology is shared only for purposes of feedback.

Presenter

Presentation Notes

As mentioned in Radiology there is an overload of diverse data from many sources that goes into clinical decision making. AI can analyze both structured and unstructured data to provide value at many levels for example: Before a radiologist receives a study AI can provide better clinical context to radiologists for example by surfacing key relevant patient data such as patient history, relevant priors, genomics consolidated problem lists etc.; During a read AI can Help inform decisions by providing clinical insights such as: flagging key findings, automating tedious tasks AI can also help retrospectively for example by finding gaps in problem lists and missed incidental findings. Transition: So how do you design AI solutions that fit into clinical workflow?

AI algorithms that work in clinical conditions Designed for:

• Ease of use

• Efficient workflow

• Varied clinical histories

• Range of ages, genders, ethnicities

• Variety of PACS systems

• Diverse imaging protocols

• Multiple device manufacturers


Presenter

Presentation Notes

You have certainly notice a large number of AI research papers and startups. AI is the latest engineering hammer and it seems like it is being applied to every possible problem. There is, however, a huge difference between a research paper and a product that can work in the sometimes wild clinical ecosystem. For AI algorithms to work in clinical conditions they need to be designed to: Be easy to use, not interfere with workflow rather help create efficient workflow While working across a variety of clinical histories Over a range of ages, genders, ethnicities Working with a variety of PACS and other healthcare IT systems Supporting diverse imaging protocols and multiple device manufacturers. Transition: So how do we go about developing systems that meet all these constraints?

Deep learning on medical images


Traditional AI Deep Learning

Flat hierarchy Deep hierarchy with layers of abstraction

Manually defined features Automatically learned features

More detailed local annotations Less detailed global annotations

Sequential computation Multiple decisions simultaneously

Deep learning is a branch of machine learning that makes use of multiple processing layers and hierarchical representations to drive the learning process.

AlexNet trained on ImageNet Data

Prob=0.8 for class fibrosis

Input (fibrosis or normal)

Presenter

Presentation Notes

Transition: Now onto the Deep learning algorithms on medical images. You may have heard the term Convolutional Neural Networks or CNN (that is not the news channel) but is a class of deep learning most often used in visual imagery. Deep learning is a branch of machine learning that uses multiple hierakikal processing layers to drive the learning process. Whereas Traditional AI uses a flat hierarchy, manually defined features (ex lung nodules), more detailed local annotations and sequential computation, Deep learning has deep layers that encode learned features. It requires less detailed global annotations for training and can perform multiple decisions simultaneously. Here is one example of a type of CNN called AlexNet (designed by Alex Krizhevsky cited by 30K publications) that outputs a probability score for a certain disease type on an image. Probability scores are a natural output of these Deep Learning algorithms. Transition: Now a Radiologist would not just look at the image so how do we program Deep learning algorithms with multiple inputs similar to how we analyze a patient case.

IBM Medical Imaging AI Development - Works in Progress -

AI Development Activity Examples

• Diagnostic Discrepancy Detection

• Liver Cancer Detection

• Breast Cancer Screening

• Prostate Cancer Screening

• Etc.

IBM Watson / © 2018 IBM Corporation

Automated analysis of chest x-rays * (*Works in Progress)


• Detection of 14 common diseases/conditions

• Achieved near state-of-the-art performance

Cardiomegaly (enlarged heart)

Pleural effusion (fluid at base of lungs)

Presenter

Presentation Notes

Another application for incidental findings we are working on if for chest x-ray screening. We used Deep Learning algorithms to identify the pobabilty of a number of diseases in chest x-ray. Here we can see heat maps indicating areas of suspicion for a specific disease. Action Get write-up from Yiting.

AI Application Example - Unreported pneumothorax

72B8D5B4BDBC8274A3D5C11BB02DD57C_R1

Zoomed ROI

Watson Health © IBM Corporation 2019 | IBM Confidential, INTERNAL USE ONLY

AI Imaging Example - Unreported rib fracture

ce93ae3525a34e3cb8d2b87acfcd4b9c_57573313

Zoomed ROI


Automated assessment of aortic aneurysms in CT scans* (*Works in Progress)


Detection of aorta centerline in 3D

Extraction of 2D aorta cross-sections

Aorta diameter measurement

Distance from aortic root (cm)

Aorta Diameter

(cm)

Aortic aneurysm

4 cm diameter limit

Aneurysm detection

Presenter

Presentation Notes

We also have developments underway for incidental findings. In this case we apply deep learning algorithms for automatic detection of aortic aneurysms in CT scans. This type of tool is value for helping with peer review and eventually providing a second set of eyes for incidental findings. Get other details from Ben.

AI Example - Unreported aortic dilatation

Axial slice Sagittal slice

Unreported ascending thoracic aortic aneurysm measuring at least 5.6cm


Liver Cancer Detection * (*Works in Progress)


Presenter

Presentation Notes

Here is an example of some of our preliminary automated liver and lesion detection algorithms from Deep learning algorithms.

IBM Watson Imaging Care Advisor for breast * (work in progress)

IBM Watson / © 2018 IBM Corporation | IASO-1747 Rev 1.0 This technology is in the research and development phase and has not been evaluated by any regulatory agencies (such as USFDA) for safety or efficacy. It is not available for any commercial or non-commercial use. Information about R&D stage technology is shared only for purposes of feedback.

Presenter

Presentation Notes

Care Advisor for Breast is a Deep Learning application for Mammography screening. A combination of various Deep Learning algorithms and image-based filters have been applied to breast screening studies. The objective is to help radiologists save time by triaging screening exams to allow radiogist to focus more time on suspicious cases. Confirm that we can say the following: our ultimate goal is to perform secondary and eventually primary reads of normal mammograms. Preliminary work was presented in a poster on educational session Monday at RSNA by my colleague David Richmond. We also have similar developments underway in breast tomography screening

Detection and clustering of signature dynamic contrast patterns* (*Works in Progress)


Presenter

Presentation Notes

And we are also working on incorporating pharmaco-kinetic models including detection and clustering of dynamic contrast patterns.

Deploy AI algorithms at scale


Presenter

Presentation Notes

To scale AI algorithms we leverage IBM’s large international cloud infrastructure and core API’s on which we build our AI applications. The medical imaging archive is our Merge VNA that is highly scalable and onto which we can connect our cloud viewers, worklists and gateways to input data for processing. Although we will have some on premise solutions our ultimate goal is a cloud-based solution for maximum scalability and cost-effectiveness. Our AI applications are designed to plug into this highly scalable cloud infrastructure for worldwide deployment. Transition: So let’s now look at how we integrate out offerings into clinical workflow.


Legal Disclaimer © IBM Corporation 2018. All Rights Reserved. The information contained in this publication is provided for informational purposes only. While efforts were made to verify the completeness and accuracy of the information contained in this publication, it is provided AS IS without warranty of any kind, express or implied. In addition, this information is based on IBM’s current product plans and strategy, which are subject to change by IBM without notice. IBM shall not be responsible for any damages arising out of the use of, or otherwise related to, this publication or any other materials. Nothing contained in this publication is intended to, nor shall have the effect of, creating any warranties or representations from IBM or its suppliers or licensors, or altering the terms and conditions of the applicable license agreement governing the use of IBM software. References in this presentation to IBM products, programs, or services do not imply that they will be available in all countries in which IBM operates. Product release dates and/or capabilities referenced in this presentation may change at any time at IBM’s sole discretion based on market opportunities or other factors, and are not intended to be a commitment to future product or feature availability in any way. Nothing contained in these materials is intended to, nor shall have the effect of, stating or implying that any activities undertaken by you will result in any specific sales, revenue growth or other results. Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance that any user will experience will vary depending upon many factors, including considerations such as the amount of multiprogramming in the user's job stream, the I/O configuration, the storage configuration, and the workload processed. Therefore, no assurance can be given that an individual user will achieve results similar to those stated here. All customer examples described are presented as illustrations of how those customers have used IBM products and the results they may have achieved. Actual environmental costs and performance characteristics may vary by customer. IBM, the IBM logo, ibm.com, and Watson Health are trademarks of International Business Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the Web at “Copyright and trademark information” at ibm.com/legal/copytrade.


Documents

Connecting data for future insights · • Data governance and access model that balances privacy with data accessibility • Engaged and ‘data literate’ hospital community The