Privacy In The Era Of Personalised Medicine Dr Kelvin Ross, Dr Brent Richards, Dr Zoran Milosevic, Dr Mark Pedersen Dr Kelvin Ross Director, IntelliHQ kelvin@intelliHQ.com.au
Machine Intelligence Progress
intelli Innovation Ou Innovation Out • Idea Development • Venture Establishment • Team Development • Funding Advisory • Commercialisation Pathways Innovation In Innovation In • • AI Innovation Centre for Attract Innovators • Readiness Assessment • Customer Validation & Trials Healthcare • Benefit Evaluation • NFP partnership between GCUH, research and industry
Delivery Model Industry Partners • Platform: AWS • Devices: GE, Philips. … • EMR: IMDSoft, … • SMEs: TechConnect, KJR, Advisory Board Veriluma Investment Fund Investment Fund • Other: Verily, CIMIT Investment Fund Selection Board Government • Qld Govt, Advance Queensland Commercialisation Pathway • Federal Government GCUH Mentoring / Advisory Customer Universities Innovation • Local: Griffith, Bond Ventures Partner Ventures • Regional: QUT, UQ, SCU Ventures Ventures • National: Curtin, … Ventures • International: NUS, MIT, Tapei, … Precision Medicine Data Platform Professional Associations • College of Intensivists, Heath Technology RACS, … Assessment Facility Health Economics Global Innovation Hubs Advisory • Startup Health, TMCx, CIMIT, NIA, … Other Research & Education • Data 61, AEHRC • Health & Knowledge Precinct
PMDP a patient physiology data platform supporting AI-based decision systems intrinsic to emerging Health Technologies. 1. Improve Health Outcomes 2. Create local opportunities for skills, jobs and venture development 3. Attract Entrepreneurs & Investment 4. Achieve Global Recognition for Research, Innovation & Commercialisation
Machine Learning INTERVENTION ANALYSIS PREDICT OUTCOME Demographics Pre-Admission Data Interventions Treatment Time-Series Outcomes Physiologic Data Lab Results Orders AI and machine learning techniques provide prescriptive/prescriptive analytics and decision support to optimise health outcomes. Comprehensive predictive and prescriptive models of disease and intervention can be realised from deep data sets. Interventions Prediction algorithms can pre-empt onset of sepsis in advance of organ failure, allowing for early intervention. Reinforcement learning can be used to understand optimal policy Outcomes for interventions for individualised patient conditions and diagnoses. This can augment clinical decision processes to advise on optimal intervention, such as drug dosage.
Precision Medicine Precision Medicine Interventions DAT DATA Journey Journey Actions Actions Bio-signals Data Data Outcomes Go Goals als
Public Data Sets https://www.kaggle.com/tags/healthcare Other resources: ● https://physionet.org ● http://www.ukbiobank.ac.uk ● https://healthcare.ai ● https://www.datasciencecentral.com/profil es/blogs/10-great-healthcare-data-sets ● https://cloud.google.com/blog/big- data/2017/05/new-healthcare-and- population-datasets-now-available-in- google-bigquery
Source: https://ai.googleblog.com/2018/05/deep-learning-for-electronic-health.html
Big Data Growth ICU Rooms • 25 Beds, 250m rows Inpatient Patches • 250 patients IOT devices will grow Additional demand for Units, health data Outpatient Hospitals, analytics Health Regions Patches • 2,500 patients Community Care • 12,500 patients
Platform Architecture Data Sources Data Science Platform Prediction Models Electronic Medical Records (EMR) Heart Rate Analysis Toolset: ETL Analytics Mortality Analysis Visualisation Big Data Compute Bio-Signal Monitors Data Modelling (map/reduce) Store Tools Septic Shock ETL Systems Integration Tools
Deeper ICU Data Sets # patients High Level Demographics and ANZICS Vitals Summary ~ 1,400k (A+NZ) First 24 hr chart data MIMIC III eICU-CRD ~ 60k (US) ~ 200k (US) Orders, Labs, etc (timestamp) 200+ research publications Low frequency chart data GCICU (15 min) ~ 20k (GC) Med frequency PMDP chart data (1 min) Other hospitals and HHSs High frequency chart data (240 Hz)
Deeper Data First 24 hrs HR_HI = 118 HR_LO = 44
Information Loss Hourly Minutely
Deep Medical Datasets Clinicians and data scientists need access to deep de-identified data. Deep data creates new tiers of predictive power, and enables new front end applications. Data sources from many hospitals need resources that will scale.
Need Data Scale 2k 250k
Data Is The New Oil 00101100101001010111001010010110010100101100101100101100101001011001010010110010100101100 10110100111010010101100101000010110010100101100110100101100101001011001010010110010100101 10010100101101110010100101100101001011001010010110010100101100111010010110010100101100101 10010101110010100101100101010101100101001011001010010110010100101100101001011001010010101
Flatiron Flatiron Health makes software to improve the workflow in cancer clinics and aggregates anonymized data from that software to share with pharmaceutical companies and research institutes. "relentless curation of data--an army of "data janitors" transforming EHR data into analyzable, actionable information." Reduced Control Groups via Real World Evidence http://cancerres.aacrjournals.org/content/78/4_Supplement/P3-17-03 https://www.forbes.com/sites/davidshaywitz/2018/02/18/the-deeply-human-core-of-roches-2-1b-tech-acquisition-and-why-they-did-it/
Data Is The New Oil Injest Refine Models Products Consolidating, cleaning and de- Machine learning models Data is ingested from identifying data is the first step. are incorporated within hospitals. Hospital Machine learning models products to deliver value, contribute to a data are constructed from the Value adding to the data sets, such translating decisions into consortia, joining their data data sets. Rich data sets as labelling, augmentation, and actions, capturing more with many others to provide provide the underlying quality assurance enhances data data into the machine a richer data set. basis for powerful models. sets for optimised model creation. learning feedback loop.
UK House Of Lords AI in the UK: ready, willing and able? • Government (Public) data including health data is of considerable value. Data is the infrastructure for artificial intelligence - providing a fundamental framework for AI systems to function • Could provide considerable economic benefit to the community • Concerned that data could lack equitable access and be monopolized by large technology companies • Commercial use of public health data raises community concerns • Community expects societal benefit from data sharing • But data on its own is raw, siloed and of limited value. Significant work is required to extract it, merge and add value. • Public sector lack guidance on how to negotiate agreements https://publications.parliament.uk/pa/ld201719/ldselect/ldai/100/100.pdf
Stages of Data Readiness Adapted from: https://arxiv.org/abs/1705.02245 Source: Ready…set…AI — preparing NHS medical imaging data for the future, Dr Hugh Harvey, https://towardsdatascience.com/ready-set-ai-preparing-nhs-medical-imaging-data-for-the-future-8e85ed5a2824
Democratisation Of Data Public Private How to Leverage Private Investment? How to grant data access equity? What is considered acceptable use?
‘Identifiability spectrum’ Credit: Understanding Patient Data
De-Identification Process Hosp A PRIVATE OPEN LIMITED Identifiable patients HIPAA Safe Harbor: Limited Data Sets HIPPA Safe Harbor: incl timestamp de-identification ● PHI under tight access ● PKI used to provide confidential constraints mapping between patient/provider IDs, ● Data linkage between and de-identified IDs ● 18 safe harbor rules data sets less timestamp de-identification HIPAA “Safe Harbor” De -identification Rules 1. Names 6. Fax numbers 14.Health plan beneficiary numbers 2. Geographic subdivisions smaller than a state 7. Device identifiers and serial numbers 15.Full-face photographs and any comparable 3. All elements of dates (except year) for dates 8. Email addresses images that are directly related to an individual, and 9. Web Universal Resource Locators (URLs) 16.Account numbers 10.Social security numbers 17.Any other unique identifying number, all ages over 89 and all elements of dates (including year) indicative of such age 11.Internet Protocol (IP) addresses characteristic, or code, except as permitted by 4. Telephone numbers 12.Medical record numbers paragraph (c) of this section [Paragraph (c) is presented below in the section “Re - 5. Vehicle identifiers and serial numbers, including 13.Biometric identifiers, including finger and identification”]; and license plate numbers voice prints 18.Certificate/license numbers
Sepsis Survivability Improvement Over 10 Years Scrambling admission date loses predictive power Source: Kaukonen et al, Mortality Related to Severe Sepsis and Septic Shock Among Critically Ill Patients in Australia and New Zealand, 2000-2012, JAMA, 2 April 2014
Recommend
More recommend
