Empowering Patients through Machine Learning-Driven Health Interventions

Mareena Mallory from Memotext talks about how to empower patients using tools built with Python.

Transcript

1. #ACTIONYOURHEALTHDATA memotext.com mtxt.io Mareena Mallory, Data Engineer MEMOTEXT Empowering Patients through Machine Learning-Driven Health Interventions
2. Who am I? Bachelor of Computer Science Specialization in Biomedical Computing Master of Health Informatics
3. What do I do?
4. BACKGROUND
5. OUR MISSION Our mission: Make health data useful We empower patients to reach their healthcare goals and improve the bottom line for healthcare stakeholders.
6. A4i A platform for supporting the schizophrenia and psychosis recovery process Our Mission: to extend care beyond the walls of the current mental health continuum, extend support, reduce isolation, and provide insights on the risk of relapse
7. Complex Behavioral Health ADHD ALZ Dementia Maternal Health • Canada’s first prenatal adaptive SMS education program • Commercialized and endorsed by the Society of Obstetrics of Canada • Predicting readmissions, ↑ support, reducing isolation for Schizophrenia populations • Feasibility: retention 94% and significant improvements in psychiatric symptoms and personal recovery • Simple Online Family Intervention for ADHD • Evidence-based adaptive adherence for pharma manufacturers and physician groups • Feasibility shows 50% improvement in intervention group • Awarded grant funding by the Centre of Aging and Brain Health Innovation • Pilot a solution with a large home care provider in Canada, utilizing the Amazon Echo to communicate with seniors with dementia and to support caregivers MEMOTEXT PARTNERSHIP PROGRAMS All technology is built on MEMOTEXT Sentinel™ Core Engine
9. Social Activation Living with Schizophrenia Stress and Anxiety Motivation and Cognition 15% 20% 20% 45% Intake assessment • 7 questions targeting the 4 content domains • User responses feed into segmentation model • Model produces custom distribution of content in real time for each person. • Updated Content Domains Custom Content Distribution Content Segmentation Algorithm
10. Use Cases + Results Data Integration Customer Medication Adherence ↑31.4% Claims Ambient Phone Data Patient Self-Report Treatment Discontinuation ↓50% statin/hypertension Wearables Pharma/Pharmacy Payors PBM Providers MEMOTEXT SYSTEMS: SELECT USE CASE RESULTS Refill Persistence ↑37.3% HEDIS/Star Ratings Disease Literacy +85% Care Coordination ↓1.5 hr/week case manager time Patient Retention 91% Predictive Surveillance
11. WHY AI/ML?
12. • A necessity to better navigate and manage chronic illness • Lack of timely targeted communications • Need for engagement to meet health goals • An inability to predict patient outcomes, behaviors and costs • Disconnect between appropriate interventions and risk identification • Gaps in care coordination Patient Stakeholder Why AI/ML?
13. Where we are Where we need to be Descriptive & Retrospective Predictive & Prescriptive Why AI/ML?
14. Why AI/ML? Machine Learning Disease Trajectory Mapping Connected Health Data Personalized Support • Claims • Risk Assessment Data • Demographics • Wearables • Ambient Phone/iOT • Feature Engineering • Adherence Clustering • Predicting Treatment Complexity / Therapy • Targeted Adherence Interventions • Care Navigator Bots • Multi-Channel Disease Interventions + Risk Identification
15. IMPLEMENTATION
16. Implementation Business Problem Definition KPI Definition, Validation Data Collection / Cleaning, Feature Engineering Data Exploration & ML Modelling Identify Intervention Nodes Identify Target Population(s)
17. Step 1: Identify the Business Problem ✓ Frame your business problem • Describe the business opportunity, threat or issue at hand Example: ‘Hospitalization is a poor health outcome and very costly for schizophrenia/psychosis patient population. We want to know who is at risk.’
18. MEMOTEXT Examples: Q: Can we identify those at risk of switching off of monotherapy? Q: Can we figure out who is likely to develop X in the next year? 2 years? Q: Can we determine who will likely stop taking their medication? Q: Can we identify who will become a high cost patient?
19. Step 2: Define the KPIs • Consider your stakeholders and their business goals • If we are interested in high cost patients, how are we defining high-cost? Starting on expensive therapy? Top 5% of high cost users in a given year? OR
20. Step 3: Collect, clean, build features! • Extract the data that is relevant in solving the business problem • Clean it • Removing irrelevant data • Ensure validity, remove corrupt/ inaccurate records • Structural errors • Outliers • Account for missing data • Proper data types • Data linking
21. ✓ Size ✓ Location ✓ Bedrooms Step 3: Collect, clean, build features! Feature = An ATTRIBUTE or CHARACTERISTIC of something
22. Provider Complexity Raw Data Patient Pharmacy Sarah A Alex B Alex D Alex C Sarah A Alex B Sarah A Feature Engineering Engineered Feature Patient Number of Unique Pharmacies Visited Sarah 1 Alex 3 WHAT IS FEATURE ENGINEERING? Step 3: Collect, clean, build features!
23. Unsupervised Learning We try to uncover patterns within the data with no labels provided. Diabetes Supervised Learning We train the models on labelled data (known outcome) to predict on unseen data. No Diabetes Step 4: Data Exploration and ML Modelling
24. Step 4: Data Exploration and ML Modelling Historical Data Training Model Prediction Learning from past cases Is this patient likely to discontinue treatment?
25. Step 4: Data Exploration and ML Modelling Historical Data Training Model Prediction Learning from past cases Is this patient likely to discontinue treatment? Actual Outcome Data Surveillance Patient stopped claiming for medication X.
26. Step 5: Identify Intervention Nodes • Can the problem be solved? Where in the patient journey will we intervene to make a difference? What are the levers? Identify intervention node
27. Step 6: Identify Target Population • Who are the right patients to receive the intervention? • What types of patients will have a different experience within the intervention? Entire patient population Those who stand to benefit! High risk group No Risk High Risk of Tx. Discontinuation
28. Which individuals will be affected by this tool? At which point in the journey will this be implemented? Into whose workflow will this fit into? What’s the end goal and how will this tool help? Overview: Putting All of the Pieces Together
29. Which individuals will be affected by this tool? At which point in the journey will this be implemented? Into whose workflow will this fit into? What’s the end goal and how will this tool help? Overview: Putting All of the Pieces Together Patients with schizophrenia/ psychosis. Reduce readmissions. By identifying those at risk to intervene early. Outside of the clinical setting based on interaction with app. Patient’s physician. Example: schizophrenia/psychosis readmission.
30. Challenges - Stakeholder Alignment
31. Empowering patients by providing them with the tools and resources they need to take control of their health
32. Personalizing and tailoring programs with machine learning- driven insights to optimize intervention uptake, adherence, and success
33. Delivering the right content to the right person at the right time
34. Evidence Based • Personalized • Validated • Integrated • Secure Thank you! GET IN TOUCH Mareena Mallory, Data Engineer mareena@memotext.com www.memotext.com 877.MEMO.TXT @memotext