Improving health care outcomes with responsible data science
1 like472 views
The document discusses improving healthcare outcomes through responsible data science, focusing on value-based approaches, secure federated analysis, and patient empowerment. It highlights the importance of personalized health data, the need for collaboration in overcoming fragmentation, and examples of using patient-generated data to inform research and care decisions. Key conclusions emphasize the necessity for ethical data practices, transparency, and the role of both patients and data science in enhancing health outcomes.
Improving health care outcomes with responsible data science
- 1. Improving health care outcomes with responsible data science Wessel Kraaij Leiden University & TNO w.kraaij@liacs.leidenuniv.nl NVPHBV spring meeting 2018
- 2. Outline 1. Overview of value based approaches to health care and data science 2. Secure federated analysis of health data 3. Patient Empowerment: Patient Forum Miner
- 3. Digital Health • Improving Health care outcomes using data science (and AI) • Includes E-health, M-health etc • What do we want? simultaneously: 1. Better outcomes 2. Better access to healthcare (best healthcare for all) 3. Reducing cost
- 4. Typical machine learning challenges in health / life sciences: diagnosis, progression, prognosis
- 5. $$
- 6. More attention for value and principles
- 7. Precision medicine, Personalized health, vitality Prediction, prognosis based on personal profiles Reasoning/decision making with incomplete information Needs lots of data, data science methods, trusted infrastructure
- 8. Early diagnosis and prevention Integration of multiple weak biomarkers Anomaly detection Self measurement Avoid costly acute care ..BUT, this hurts current business cases
- 9. Increasingly stronger role for citizens and patients Participatory research Shared decision making Patient centered PROM’s Late effects
- 10. Tailoring in information access, communication, coaching Segmentation Avoid one size fits all Sociodemographic differences E.g. chronic diseases, elderly, low literacy
- 11. Value Based Health Care 𝑉𝑎𝑙𝑢𝑒 = 𝑃𝑎𝑡𝑖𝑒𝑛𝑡 𝑟𝑒𝑝𝑜𝑟𝑡𝑒𝑑 𝑜𝑢𝑡𝑐𝑜𝑚𝑒𝑠 𝐶𝑜𝑠𝑡 + 𝐵𝑢𝑟𝑑𝑒𝑛 Objective function Michael Porter
- 12. Machine readable, reusable data : FAIR • Findable • Accessible (under well defined conditions) • Interoperable: Machine readable • Reusable • Secundary goal: combine datasets for analysis • FAIR does not imply open!! Michel Dumontier, Maastricht University
- 13. GO FAIR • Internet of FAIR data and services • FAIR Metrics • Implementation Networks (IN) • Example IN: Personal Health Train
- 14. Responsible Data Science • Fairness: avoiding prejudice due to e.g. biased training data • Accuracy: quality of information, prediction , how certain? • Confidentiality: GDPR, also commercially confidential • Transparency: Can we explain conclusions drawn from big data? (opening the Black Box)
- 15. FAIR data (reusability) Fairness (avoiding bias) Accurate (incl AI) Confidential Transparent (incl AI) HCI (incl AI) Communicating with (a diverse group of) patients Improving outcomes of health care system (care, cure, prevention) using different types of data Making choices based on normalized added health value per cost unit Relating data science challenges to health care trends
- 16. 1. Some conclusions • Combining / comparing personal data across individuals is needed for personalized, preventive health • Digital Health challenges pose quite relevant use cases for Data Science, AI and HCI • Existing costing/business models may actually be a barrier for better patient value
- 18. Quantified self 19 bron: MIT A movement of citizens and ‘makers’ that aim to explore the possibilities of self-tracking. Gary Wolf (Wired): “Almost everything we do generates data”. bron: RescueTime WHAT COULD WE LEARN IF we could measure and record activities, social context, environmental context, physiological parameters, food intake, sleep across our entire lifetime? (sometimes called Exposome). Now we have sensors for most of these.. WE CAN LEARN: Personalized health advice based on systems view and population data Integrating evidence based medicine and data driven predictions
- 19. The mutual dependence of personal and population health data 20Bron: cbw.ge en wikimedia.org Van ‘big’ naar ik
- 20. Towards personalized e-advice Model Personal health profile (longitudinal data) Population health profiles Health professional: duty of confidentiality For e-services: How can we avoid that competitors learn the model? How can we avoid that a health app copies personal data?
- 21. Health Data Challenges • Data is organized around treatments • Silos • Heterogeneous data • Patient generated data not linked • Need patient/citizen centered data • Undo fragmentation • Link person generated data • Deal with uncertainty and imperfections • Quantify accuracy • Missing data
- 22. Some barriers for statistical analysis and ML • Data is horizontally partitioned • Distributed learning • Personal Health Train • Data is vertically partitioned • Existing practice: Trusted 3rd party (TTP) • Health Data Cooperative (e.g. Midata) • Secure multiparty computation ID age GP visits gender 1 55 70 M 2 45 60 F ID Age gender 1 55 M 2 45 F 3 20 F 4 22 M ID age GP visits gender 3 20 25 F 4 22 10 M ID GP visits 1 70 2 60 3 25 4 10
- 23. Sign the manifesto at: Personal health train • Distributed data • FAIR data stations • Bring the algorithm to the data • Approximate global analysis • FAIR data stations include • Clinical repositories • Personal lockers (PGO) • Umbrella stations (e.g. HDC) • Trains implement secure workflow for researchers and patients/citizens https://www.dtls.nl/phtmanifesto/
- 24. NWA startimpuls VWDATA • Resonsible Data Science • FACT, FAIR • Learning from distributed data • Develop secure algorithms for learning from a ‘join’ of databases with confidential/personal data
- 25. Build prognostic model using secure regression H2020 BigMediLytics • Developing “secure regression” using secure multiparty computation methods • Aim: improve KPI with 20% in large scale trials • Erasmus MC Heart Failure pilot • KPI’s • Increase medication compliance • Decrease number of hospital re- admissions Ergo Achmea claims EMC clinical data Develop and evaluate intervention informed by the joint longitudinal dataset (based on risk and cost analysis)
- 26. Proposed solution: Multi party computation • Several data owners intend to perform to ‘learn’ from the combination of the datasets • However, they cannot release their data • MPC enables to overcome this barrier: • learning without disclosing • Extensive communication protocol between partners • Proven non disclosure • TNO/CWI/Philips are developing and evaluating MPC methods • homomorphic encryption • garbled circuits • secret sharing
- 27. Simple Example: Secure summation using MPC F(n1)=R+n1 F(n5)-R F(n2)=F(n1)+n2 F(n3)=F(n2)+n3F(n4)=F(n3)+n4 F(n5)=F(n4)+n5 Σ =
- 28. Assessment • Cryptographic MPC protocols have been around for decades • Applicability was hampered by inefficiency of several protocols • Recent advances in protocols and computing power, increased attention for privacy have resumed interest in development of MPC • MPC has the potential to become a crucial enabler for improving health and health care by learning from data
- 29. 2. Conclusions Challenge is to fight fragmentation (need collaboration, usual model is competition) It is crucial to restore the balance in data governance in order to avoid that platform companies increasingly control crucial aspects of our civil society NL well positioned to take the lead in building a public infrastructure for health data analysis.
- 30. 3. Patient Forum Miner An example of collective “empowered” participation: Patient Platform Sarcomas
- 31. Sarcoma: a rare cancer
- 32. Rare cancer challenge • Low incidence: small patient cohorts for research • Limited data in national registration • Need for international cooperation • Limited budgets for research • In most hospitals not enough experience to treat rare cancers: need for concentration
- 34. Patient generated tissue bank
- 35. Open Facebook group GIST Support International
- 36. Discussion forums: ● Information on: ○ Development of disease, treatment, side-effects ○ Strategies to cope with side-effects ○ Adherence to therapy ○ Co-morbidity ○ Quality of life ● Also: lots of emotional discussions, support ● Use data mining, NLP, machine learning to extract valuable information caregiver for mother dx with GIST 12/31/12 31x19x9 cm primary tumor removed 12/12/12 400 mg gleevec 1/15/13 to 12/6/13 Recurrence and removal of 3 masses from the omentum (3.5 to 4.4 cm) on 12/6/13 Sutent 25 mg. since 1/15/14 37.5 mg. since 8/14 (failed Sutent 6/15) Sorefinib (naxavar) 7/3/15 more tumors removed 7/13/15 Votrient 9/4/15 to 12/30/15 50 tumors removed 1/16/16’ Wildtype Gist NO Kit or PDFRA mutation, SDH intact
- 37. Analysis method Public Facebook Group 4 year 40,000 posts UMLS tagger Word2Vec Categories Database UMLS: thesaurus with medical terms Word2Vec: semantic relationships Summarisation: machine learning Elastic search User interface: relevant posts Automatic summarisation
- 38. User interface for textmining of forums
- 39. Examples of findings Normal dose of imatinib for GIST patients is 400 mg per day. In case of progression: increased to 800 mg per day. In one hospital, policy was to advise patients to take full 800 mg at dinner. Question: would it be better to split the dose? Based on forum discussion analysis: yes, most patients experience less side effects when splitting the dose. Based on result, policy of hospital was changed. Something else was found: patients reported that taking the medicine with dark chocolate greatly diminishes nausea.
- 40. Example: co-morbidity ○ Searching for patients with PDGFRa mutation we found two patients who mentioned that they also had thyroid cancer. ○ Searching for this specific combination we found around 20 cases ○ coincidence or is there a relationship? platelet-derived growth factor receptor α (PDGFRA) mutation, occurs in 10% of GIST cases
- 41. Example: research agenda ● Patients participate in agenda setting for research in university hospital ● Longlist of topics was composed on basis of analysis of (inter)national discussion fora: What is important for patients? ● Longlist was used to consult patients through web survey: What are your priorities and why? ● 75 patients responded ● 3 topics were selected and are now included in a PhD project ● 1 topic could be addressed on basis of existing knowledge published in magazine of patient organisation
- 42. OPTION THEME TOTAL # PER THEME 1A Surgery of metastases 31 1B RFA 11 1C Embolisation 9 2A Cramps and nausea 8 2B Skin problems 5 2C Fatigue 22 2D Long term side effects of imatinib 30 3A Best way to take imatinib 21 3B Interaction with food 28 4A Last phase of disease 17 4B Micro-gists 8 4C Hereditary aspects 15 4D Combination with other cancers 6 TOTAL 211
- 43. Further development ● Building experience: project with 4 other cancer patient organisations in the Netherlands, combine with surveys, build experience and improve algorithms ● Strengthening scientific basis: detecting patterns, quantification, statistical analysis, validation, filtering by linking to existing knowledge, clinical testing of hypotheses (PhD project)
- 44. 3. CONCLUSIONS ● Patient experience discussion forums can be very valuable, in particular in case of rare diseases; worldwide community ● Patient forums can reveal unexpected patterns and can provide information that would otherwise never reach doctors ○ Care must be taken to avoid experimenter bias ● This project demonstrates the power of partnership between patients and medical professionals
- 45. Finally,.. • Improving Health Care affects us all: • Data, algorithms can support patients and care professionals. • Patient value should be more important than shareholder value • Health data is sensitive, people should be in control regarding access. • GDPR stimulates innovation in data science • Explainability • Protection of personal data
- 46. References • P4 • Hood, Leroy, and Stephen H. Friend. 2011. “Predictive, Personalized, Preventive, Participatory (P4) Cancer Medicine.” Nature Reviews Clinical Oncology 8 (3): 184–87. doi:10.1038/nrclinonc.2010.227. • VBHC • Porter, Michael E. 2010. “What Is Value in Health Care?” New England Journal of Medicine 363 (26): 2477–81. doi:10.1056/NEJMp1011024. • FAIR • GO-FAIR: https://www.go-fair.org/ • Wilkinson, Mark D., Michel Dumontier, Ijsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas Blomberg, et al. 2016. “The FAIR Guiding Principles for Scientific Data Management and Stewardship.” Scientific Data 3 (March): 160018. doi:10.1038/sdata.2016.18. • FACT • http://www.responsibledatascience.org/ • NWA VWDATA • https://wetenschapsagenda.nl/start-vwdata-onderzoeksprogramma/ • Personal Health Train • Damiani, Andrea, Mauro Vallati, Roberto Gatta, Nicola Dinapoli, Arthur Jochems, Timo Deist, Johan van Soest, Andre Dekker, and Vincenzo Valentini. 2015. “Distributed Learning to Protect Privacy in Multi-Centric Clinical Studies.” In The 15th Conference on Artificial Intelligence in Medicine, edited by J. H. Holmes, R. Bellazzi, L. Sacchi, and N. Peek, 65–75. Pavia, Italy: Springer. http://eprints.hud.ac.uk/23905/. • Secure multiparty computation • Meilof VEENINGEN, Supriyo CHATTERJEA, Anna Zsófia HORVÁTH, Gerald SPINDLER, Eric BOERSMA, Peter van der SPEK, Onno van der GALIËN, Job GUTTELING, Wessel KRAAIJ, and Thijs VEUGEN. Enabling Analytics on Sensitive Medical Data with Secure Multi-Party Computation. In Proceedings of Medical Informatics Europe 2018, Gothenburg, 2018. • Patient forum miner • Oortmerssen, Gerard van, Stephan Raaijmakers, Maya Sappelli, Erik Boertjes, Suzan Verberne, Nicole Walasek, and Wessel Kraaij. 2017. “Analyzing Cancer Forum Discussions with Text Mining.” In Proceedings of Second International Workshop on Extraction and Processing of Rich Semantics from Medical Texts. Vienna.