putting it all together: unlocking the potential of ......de-identification per mobility. de-id...

Post on 07-Jul-2020

2 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

1

Putting it all together: unlocking the potential of tomorrow with ML

and the CloudSession INT6, February 11, 2019

Aashima Gupta, Director, Global Healthcare Solutions, Google Cloud

2

Aashima Gupta

Has no real or apparent conflicts of interest to report.

Conflict of Interest

3

• Describe the realities of evolving technology changing the interoperability landscape.

• Assess the role of use cases in driving policies and technologies to support value-based care.

• Evaluate the anticipated benefits and challenges of TEFCA implementation across stakeholder groups

• Analyze the remaining gaps as data exchange is expanded to broader stakeholder groups in support of innovation.

• Describe the value of payer and provider data exchange within the healthcare ecosystem.

Learning Objectives

1 in 20 searches are health related

An example: mental health

An example: public health

Make health information universally useful and accessible

Organize the world’s healthcare and life sciences data, make it accessible , secure and useful .

Organization

Make it easy to ingest, normalize, and join data

Accessible

Support open standards, APIs, interoperability, and discovery through search

Useful

Insights to action through analytics and machine learning

Secure

Lead the industry in security, privacy and compliance from endpoint to data center

What if

Query for all females, age 45 -60 with a specific biomarker, with history of breast cancer that haven’t been screened in the last 24 months

01

Three pillars to democratize innovation for healthcare

01Democratizing Technology Foundation

Google Cloud Infrastructure, scale security and compliance

0 2Democratized Machine Learning

Cloud AutoML, TensorFlow

0 3Democratizing Patient Health Records

Interoperability and Standardization

Democratizing Infrastructure

Eight cloud products with over one billion users each,all powered by the c loud.

Google is a Data Company

uploads per minute

500hrsusers

1B+search index

100PB+query response time

0.25s

Compliance for Healthcare

Global Scale 18 regions by end of 2018

Google Compute Engine guarantees 99.95% uptime

Health DataInteroperability

Fostering data interoperability and nurturing FHIR Community

“The HL7 FHIR Foundation is thrilled by this generous contribution of Google Cloud Platform services to support the ongoing activities of the FHIR community to help advance our goal of global health data interoperability,” said Grahame Grieve, HL7 FHIR Foundation Board Member and HL7 FHIR Product Director. “The future of health computing is clearly in the cloud, and FHIR will serve to accelerate this transition.”

Grahame Grieve , FHIR Principal

Healthcare Enterprise Landscape

Cloud and AI Enabled Digital Health Platform

Payer/Health Plans/Pharma

EMR, Legacy Systems

Genomics Data , Imaging PACS

On-prem Data warehouse

Data/Analytics Engine

Remote Monitoring Devices

Wellness and Fitness Devices

Patient Mobile and Tablet Apps

External Data

OLD will not get you the NEW

ChallengeBridging Chile’s healthcare

Chile’s 1,400 connected health facilities and 1,000 remote medical facilities lacked connectivity – while many of its healthcare systems couldn’t easily, or securely, interoperate.

Previous attempts to resolve failed due to the overwhelming costs and complexity of the solution.

New ApproachDigital Architecture

Nationwide API-based architecture

● Data and applications are easily, yet securely, available

● Enables Public Alerts, Population Health Management Programs, etc

● Connects HC Centers – large and small

● Reinforces and leverages digitizing all clinical and administrative processes

Google Cloud Healthcare API

EHR ImagingHealthcare data

Genomics

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare datasetCloud Healthcare API

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Per Mobility

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Per Mobility

More modalities in the future More modalities...IoT

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Support standard protocols GeonomicsReads

VariantsPipelines

DICOMwebWADO-RSSTOW-RSQIDO-RS

FHIRCreate UpdateRead

HL7v2MLLP

Msg storeAPI

Per Mobility

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Transformations/conversions

Per Mobility

Convert to FHIR

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

De- identification

Per Mobility

De- identified datasetDe- id

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Data joining and export to analytics engine exploration, processing and machine learning

De- identified datasetDe- id

Join, Export

CloudDataproc

CloudDatalab

Cloud MLBigQuery

Per Mobility

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Model development and testing De- identified datasetDe- id

Join, Export

Per Mobility

ML ModelsCloudDataproc

CloudDatalab

Cloud MLBigQuery

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Model inference workflow integration

De- identified datasetDe- id

Join, Export

Per Mobility

ML ModelsCloudDataproc

CloudDatalab

Cloud MLBigQuery

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Advanced operations and transformations

Per Mobility

Search

NLP

MPI

Google Cloud Healthcare API

EHR ImagingHealthcare data Genomics

Healthcare dataset

HL7v2 FHIR GA4GH APIsDICOM

Cloud Healthcare API

Serving to applications

Per Mobility

Digital data exchange with Cloud Healthcare APIs

Medical RecordsEMRGenomicsMedicationsImagesEnvironmentSocial determinants

Clo

ud H

ealth

API

sH

L7, F

HIR

, DIC

OM

, Gen

omic

s

Data warehouse, analytics and

machine learning

Research and clinical processing

and workflows

Health APIx: Developer portal

Machine Learning for All

Machine Learning with Google Cloud

orUse our models

● Leverage Google’s domain expertise

● No tools or expertise required

Train your own models

● Build on your own specialized domain expertise

● Use Google tools for building and training models

Machine Learning with Google Cloud

orUse our models

● Leverage Google’s domain expertise

● No tools or expertise required

Train your own models

● Build on your own specialized domain expertise

● Use Google tools for building and training models

Diabetic retinopathyDiabetic retinopathy:Fastest growing cause of blindness

415M people with diabetes

North America and Caribbean

2015 44.3 million2040 60.5 million

Europe2015 59.8 million2040 71.1 million

Middle East and North Africa

2015 35.4 million2040 72.1 million

Western Pacific2015 153.2 million2040 214.8 millionSouth east asia

2015 78.3 million2040 140.2 million

Africa2015 14.2 million2040 34.2 million

South and Central America2015 29.6 million2040 48.8 million

INDIAShortage of 127,000 eye doctors45% of patients suffer vision loss before diagnosis

Diabetic retinopathy

Diabetic retinopathy

How DR is diagnosed:Retinal fundus Images

No DR

Mild DR

Moderate DR

Severe DR

Proliferative DRHealthy

Hemorrhages

Diseased

Diabetic retinopathy

Adapt a Google deep neural network to read fundus images

130k images

Deep convolutionalneural network

54 ophthalmologists

880k diagnoses

No DR

Mild DR

Moderate DR

Severe DR

Proliferative DR

Diabetic retinopathy

Adapt deep neural network to read fundus images

No DR

Mild DR

Moderate DR

Severe DR

Proliferative DRConv Network - 26 layers

Image Quality

L/R eye

Fie ld of View

130k images

Cloud AutoML Vision

ML starts with getting hold of your data

CaptureCloud Health APIs

Pub/Sub

Storage Transfer Service

Data Transfer Appliance

ProcessCloud Health APIs

Dataprep

Dataflow

Dataproc

StoreCloud Health APIs

Cloud Storage

BigQuery

Cloud SQL

Datastore

BigTable

AnalyzeBigQuery

Dataflow

Datalab

Insight

Cloud ML Engine

Cloud AutoML

Step 1: Training images from a Dermatology Clinic.

IMPORT: User uploads CSV file with image names and labels into the app

Most popular image file formats are supported (jpg, gif, png, etc.)

LABEL: Images are tagged with “labels” (“Nevus”, “Lentigo NOS”, “Healthy”, etc.)

Step 2: Automatically train ML model

TRAIN: user initiates training of the model

Training takes between 10 min to 24 hours (weeks/months of DIY project)

EVALUATE: Results of the training are shown (Precision, Recall, etc.)

Step 3: Time for predictions

PREDICT: model assigns class probabilities for new images

Prediction only takes seconds

Access: model via UI, command line, or API and integration in Clinical Workflows

Skin disease classification using AutoML

Input:Images with associated “classes”

Model Evaluation Result:Confusion matrix

Prediction:Image classification

Machine learning: from big data to meaningful insights

Raw Health Data

Clinical Impact

data mining

machine learningvisualization

analysis

Three pillars to democratize innovation for healthcare

01Democratizing Technology Foundation

Google Cloud Infrastructure, scale security and compliance

0 2Democratized Machine Learning

Cloud AutoML, TensorFlow

0 3Democratizing Patient Health Records

Interoperability and Standardization

We’re not just thinking about now. We’re thinking about, what is next.And how we can build a healthier future...

Thank you

top related