Kraaij infrastructures for secure data analytics def brussel 2017
2 likes328 views
The document discusses the challenges and advancements in secure data analytics within healthcare, highlighting the need for individual control over data access while complying with GDPR regulations. It emphasizes the importance of anonymization and the risks of re-identification when combining personal data. The Holland Health Data Cooperative is presented as an innovative model that empowers citizens to manage their health data and utilize it for research without compromising privacy.
Kraaij infrastructures for secure data analytics def brussel 2017
- 1. Infrastructures for secure data analytics Wessel Kraaij TNO & Leiden University
- 2. EU: Digital technology will redefine health and care • Care: transition to Value Based Health Care • PROMs • Aggregated cost of care paths • Patient: self management of health, patient science • N=1 , personalized health • Find comparable health trajectories • Combining the right data may lead to new insights • BUT Data storage is fragmented • BUT The GDPR limits the combination of data sets • Challenges: • Provide a trusted environment, individual control on data access and sharing • Supporting secure and legal data analytics for combined datasets
- 3. Some barriers for statistical analysis and ML • Data is horizontally partitioned • Distributed learning • Personal Health Train (J van Soest) • Data is vertically partitioned • Existing practice: Trusted 3rd party (TTP) • Health Data Cooperative (Midata) • Prana Data (example of secure multiparty computation) ID age income sex 1 55 70000 M 2 45 60000 F ID Age sex 1 55 M 2 45 F 3 20 F 4 22 M ID age income sex 3 20 25000 F 4 22 20000 M ID income 1 70000 2 60000 3 25000 4 20000
- 4. Traditional solution – trusting a third party • For research: • create anonymized/ pseudonymized datasets, possibly using a trusted third party • Anonymization: remove identifiers • Export anonymized dataset for research (academic/ commercial)
- 5. Assessment • The more personal data is combined, the easier it is to re- identify a profile and the more difficult the anonymization process is • E.g. personal well-being data (Fitbit, smartphone apps) are quite personal and could lead to re-identification using external data. • Personal data may seem innocent, but can lead to valuable insights • Strict regulation on data processing and storage • Data leaks can lead to substantial fines • Professionals are reluctant to use big data approaches
- 6. Towards a citizen driven healthcare economy: > Citizen’s together form a Cooperative and a Community > The cooperative delivers the platform and governance structure > Enables individuals to collect their data (medical and lifestyle) > Provides services for members and delivers services to customers > Data is controlled by citizen and patients themselves > Rely on their cooperative for support CONTROL SELF- MANAGEMENT HOLLAND HEALTH DATA COOPERATIVE: THE GAMECHANGER TRUST REWARD VIRTUAL SAFE
- 7. MIDATA Server(s) Data Services and Access Management 3rd Party Data Sources Anony- mised export Lifestyle Health Others Research environment Smartphone 3rd Party Apps MIDATA Portal Quan- tified self App Coach- ing App Treat- ment App Follow- up App HHDC HHDC architecture
- 8. PRANA DATA (WWW.PRANADATA.NL) Privacy preserving analyse of sensitive data Results: Proof of principle: secure ‘babies like mine’ , homomorphic encryption, for predictive mean matching (characteristics of babies with similar growth curve) User study White paper Collaboration with Personal health train, PEP Follow up in H2020 BigMedilytics (ErasmusMC, Achmea, TNO and VWDATA
- 9. PREDICT EVOLVEMENT USING PATIENT MATCHER
- 10. PREDICT EVOLVEMENT USING PATIENT MATCHER
- 11. PREDICT EVOLVEMENT USING PATIENTS MATCHER
- 13. GOAL: ANSWERING CHILD DEVELOPMENT QUESTIONS WITHOUT SHARING SENSITIVE DATA
- 14. SECURE PATIENTS LIKE ME: 1/3 DATA INGEST
- 15. SECURE PATIENTS LIKE ME: 2/3 ANALYSIS Apply ‘predictive mean matching’ in the encrypted domain
- 16. SECURE PATIENTS LIKE ME: 3/3 DECRYPTION
- 17. QUESTIONS ANSWERED? PRIVACY PRESERVED? Alice’s parents/doctor can: Plot expected growth curve Spot growth issues that might occur Obtain benchmarks about parent’s age, doctor visits, obesity, etc. Privacy preserved: Sensitive records nor the aggregations are learned by the system Analysis server learns distribution of comparison metric The distribution of child lengths is not sensitive