Decision support based on machine learning for effective and efficient care

Transcription

1 Decision support based on machine learning for effective and efficient care September 21th 2017

2 Health care is currently depending on locality and not data driven Vision: now Care provider Academia and research Hospital Care Patient perspective Location Provider Patients are being treated depending on the local situation in the local institution The more complicated the patients case, the more specialised the care provider that delivers care is Knowledge Data Primary care (GP) Patient and informal care Less specialised Patient Integration Decisions are not based on patients personal situation and informal care is small part of health care Very little integration and communication through multiple lines of care Knowledge Very little exchange of data. Knowledge transfer is slow and studies and protocols do not cover the variety and complexity of daily practice

3 Digitisation of health care allows for specialised and personalised care outside of the clinics Vision: future Care provider Data sources Patient perspective Academia - decision making Research data Proteomics Location Health care is digitising and moving from clinics to home Knowledge Data Hospital Care - executing Primary care (GP) - guiding Integrated and data driven care path per disease Patient and informal care - monitoring Genomics Imaging Sensor data Medical files Wearables +apps Lifestyle data Open data Analysed with Machine Learning Provider Patient Integration Knowledge All decisions are made based on large scale data and world leading expertise, other care providers execute and guide Health care becomes personalised: patient is involved in decision making and receives tailor made care Integrated care paths per disease that are optimised for the specific case Through technology and machine learning doctors learn from every patient continually and directly

4 Knowledge is based on average results in a small population of irrepresentative patients and is slow and inaccessible Problem 17 years Biased research on a fraction of patients Applied on a great variety of patients

5 Pacmed builds tools that provide doctors with a tailor-made evidence-base for individual patients Solution Direct Present relevant outcomes and advice of guideline for patient during consultation Learn from every patient using machine learning which patients are similar

6 We started our proof of concept by supporting Dutch GPs with the treatment of urinary tract infections Primary Care Proof of Concept General Practitioner Generalist Little scientific research Strategically easier to start a company Treatment of Urinary Tract Infections Common (1,2 million/y) and innocent Suitable for Machine Learning Little scientific research Netherlands Strong and digitised primary care Innovative Strong network

7 Observational data can add enormously to the evidence base in the consultation room Patient population in development of Decision Support System (NIVEL) Patient population in studies on which NHG protocol is based Healthy females: Pregnant females: Patients with Diabetes: Patients with abnorm. in urin. tract: Patients with signs of tissue invasion: Males: Healthy females: Pregnant females: Patients with Diabetes: Pat. with abnorm. in urin. tract: Pat. with signs of tissue invasion: Males: year

8 We finished software for UTI that will be implemented with 100 GPs paid by CZ, Menzis, ZK Primary Care Proof of Concept Gebaseerd op 2100 mannelijke patiënten tussen 72 en 81 waarvoor ook geldt: Geen tekenen van weefselinvasie

9 We will investigate the use and effect of our software in a controlled study Primary research questions Secondary research questions 1 How many times is the software used? For which patients is the software used more often? Which doctors used the software more often? 2 What is the change in prescription behaviour while using the software? Do doctors treat in line with Pacmed algorithms more often? For which patients and treatments do doctors change behaviour more often? Which doctors change behaviour more often? 3 What is the change in effectivity of treatment of urinary tract infections while using the software? Did doctors treat patients more successfully (when they changed behaviour)? What are the costs saved? For which treatment were outcomes improved most/least? For which patients and doctors were outcomes improved most/least? 4 How do doctors experience the use of the software? Would they keep working with the software? How could the software be improved? For which diseases would they be most willing to use such software? Academic partners:

10 In hospital care we started at the ICU by supporting doctors with discharging patients Hospital Care Intensive Care Intensive Care Enormous amounts of high quality data High impact on patient s lives High costs Discharging patients Decision that is made daily for all patients Suitable for Machine Learning Life and death decision with clear cost savings VUmc Experienced researchers with Big Data projects High quality and complete data of pts Cooperations with other IC units

11 We supported Meetbaar Beter in improving their predictive models and are investigating new opportunities Hospital Care Cardiology Cardiology A lot of relevant data High impact on many patient s lives High costs Which cardiovascular intervention Very common decision Suitable for Machine Learning clear cost savings Meetbaar Beter Experienced researchers that know data well High quality outcomes Cooperations with many hospitals

12 We aim to make Psychiatry more outcome driven Hospital Care Psychiatry Psychiatry Very ill understood with current methods High impact on patient s lives High costs Which antipsychotic/antidepressant (in clinic)? Decision that affects many patients Suitable for Machine Learning (similar to other work done) Clear cost savings for patients in clinics UMCU Experienced researchers with Big Data projects (Relatively) high quality and complete data Cooperations with Primary Mental Health Care

13 We aim to combine the best of the technical and medical world Team Core team Wouter Kroese Co-founder Medicine Logic Nationale DenkTank Willem Herter Co-founder Physics Nationale DenkTank Hidde Hovenkamp Partner Econometrics Financial Mathematics Nationale DenkTank Daan de Bruin Senior Data Scientist Econometrics MIcompany Hiring Data Scientist(s) Annanina Koster Graduate Intern Econometrics Research team Prof. Mattijs Numans GP and head Family Medicine LUMC Dr. Egge van der Poel Clinical Data Scientist Erasmus MC Prof. Niels Chavannes GP en e-health researcher LUMC Dr. Mark Hoogendoorn Researcher Artificial Intelligence VU Dr. Bart Knottnerus GP and researcher AMC Jochem Cornelis Software Developer Dr. Robert Verheij Head NIVEL primary care registrations Dr. Elbers Intensive Care doctor and researcher VUmc Dr. Tobias Bonten GP en e-health researcher LUMC Patrick Thoral Intensive Care doctor and researcher VUmc Hine van Os MD/PhD LUMC Berend Beumer Student in Medicine and Econometrics

14 The opportunity for Pacmed to improve a specific decision depends on both value and feasibility Value Feasibility Criteria for cases Scale Availability of data Bigger scale means more data, more often used software and more impact Essential to develop decision support software Impact on patient Relevant features present in data More societal value, more urgency with doctor and patient, more PR Features that most likely influence outcomes need to be available in data Clear and multiple decision options for doctors Outcome measure can be deduced from data Sign that there is room for improvement and opportunity for support Preferably clear, acute outcome that measures quality of life of patient Variation in provided care and it s success Possibility of integration in IT system Sign that there is room for improvement and a natural experiment to learn from Essential for use of software and to use patient s data as input for algorithms High (medication) costs Clear business case and available budget More value to gather, medication costs are easiest to save directly Clear and structural financial benefit for relevant stakeholder Poor (scientific) evidence-base More likely that software will give added value and will be used Recent changes in policy and protocols Could create urgency or incentive to use the software