A Strategic Partnership Proposal presented to Rogers CommunicationsSeptember 2017Dr. Heather RossTed Rogers and Family Chair in Heart FunctionSite Lead, Ted Rogers Centre for Heart Research

A Strategic Partnership Proposal presented to Rogers CommunicationsSeptember 2017Dr. Heather RossTed Rogers and Family Chair in Heart FunctionSite Lead, Ted Rogers Centre for Heart Research Preventing Heart Failure Hospitalizations with Artificial IntelligenceDate: April 17th, 2020Dr. Heather Ross MD, MHSc(Bioethics), FRCPC, FACCHead, Division of CardiologyPfizer Research Chair in CardiologyLorretta A. Rogers Chair in Heart FunctionSite Lead, Ted Rogers Centre for Heart ResearchProfessor of MedicinePeter Munk Cardiac Centre Disclosure§I have no relevant disclosures Why do we want to predict risk of hospitalization…….to intervene sooner…….§Mortality§Quality of Life §Cost §Days lost from work/lifeRisk Prediction ScoresMonitoring tools MAGGICN= 39,372 patientsHFpEF, HFrEFLarge international database from 30 cohort studiesFU 2.5 years13 clinical variablesPurpose: predict mortalityMultivariable piecewise Poisson regression methods with stepwise variable selection Meta-Analysis Global Group in Chronic (MAGGIC) Heart Failure Risk Score Pocock et al, EHJ 2013 34, 1404–1413Rich et al, J Am Heart Assoc. 2018;7:e009594. Systematic Review of Predictive models Circles are mean C-statistic values for each modelRED = mortalityPURPLE = mortality/readmGREEN = readmissionC-statistics for models predicting:Mortality = 0.71Mortality or HF hospitalization = 0.63HF hospitalization = 0.68 Ouwerkerket al, JACC HF 2014;2:429–36 Why do predictive models fail at such a high rate?§Storage of data in medical silos prevents deployment and creation of algorithms§Algorithms are NOT sufficiently transparent/explainable§Algorithms are there but not used….§Lack of trustof clinicians in predictions that are generated by algorithmsBut most importantly….§Predictive models are insufficiently predictive © www.therecylcer.com 1824-1907 1909-2005 Data Amount of Data Time Available info Important info Adapted from Johnson, K.W. et al. J Am Coll Cardiol. 2018;71(23):2668–79; Obermeyer and Lee, 2017; 377:1209. You are here TMI!!!!The complexity of medicine exceeds the capacity of the human mind AI, ML, DL Learning Types§Supervised§Algorithms use a dataset labeled to predict the desired and known outcome§Great for classification and regression problems§Time consuming§Requires labelling §Unsupervised§Seeks to identify novel disease mechanisms, genotypes, or phenoytpesfrom hidden patterns present in the data…§Find the hidden pattern without feedback from humans§Reinforcement §A hybrid of supervised/unsupervised learning §Aim is to maximize the accuracy of algorithms using trial and error MHI2002-2019 | Emergent Topics in Health Informatics Uses of AI: assisted, augmented, automatic AIAssisted AIConsidered a weak form of AI and it is mainly used to automate simple tasks © Reuters Augmentation AISupport human decisions, rather than simulate independent intelligence Automatic AIThe final and most feared state of artificial intelligence; autonomousintelligence that can make decisions without human intervention https://www.tgdaily.com/technology/assisted-augmented-and-autonomous-the-3-flavours-of-ai-decisions Clinical Application Dr. J. CafazzoExecutive Director M. YeungManager Dr. E. SetoAssistant Professor Mala DoraiProduct Manager Mary O’SullivanClinical RN Lead Patrick WarePhD student Requirementsto close the Circle of Home Management of Heart FailureMany connections are required to allow for incorporation of physiological information obtained from patients at home to trigger interventions and potentially improve outcomes by means of heart-failure disease management. Adapted from Desai and Warner Stevenson, NEJM 2010;363;24 Rule Base Expert SystemAlerts…. Recommendations Alerts…. Medly’sbrain is a decision treevDesigned to mimic clinicians decision making processvMathematical rules based algorithm can only handle a limited number of factors for decisionvDecision process tends to be conservativevcan generate many false positives Setoet al, JMIR 2012 Jan-Feb; 14(1): e31Setoet al, IntJ Med Inform 2012:81:556-65Setoet al, JMIR 2012;10:14(1):e25. doi:10.2196/jmir.1923Setoet al, J Cardiovasc. Nurs2011;26:377-85.Setoet al, JMIR 2010;12(4):e55. doi: 10.2196/jmir.1627. Clinical OutcomesLimitations –pre and post analysis Number of Hospitalizations 0,750,54 00,20,40,60,81 6 months before enrollm ent6 months after enrollmentSignificant finding Length of Stay (Days) 8,426,3 0246810 6 months before enrollm ent6 months after enrollmentSignificant finding Count Days Ware et al, J Med Internet Res 2020;22(2):e16538) doi: 10.2196/16538 P<.001* * ** MHI2002-2019 | Emergent Topics in Health Informatics Uses of AI: assisted, augmented, automatic AIAssistant AIConsidered a weak form of AI and it is mainly used to automate simple tasks © Reuters Augmentation AISupport human decisions, rather than simulate independent intelligence Automatic AIThe final and most feared state of artificial intelligence; autonomousintelligence that can make decisions without human intervention https://www.tgdaily.com/technology/assisted-augmented-and-autonomous-the-3-flavours-of-ai-decisions ML Prediction of Mortality and Hospitalization in HFpEF Angraalet al, J Am Coll CardiolHF 2020;8:12–21 CAVEATSClinical trial not representative of broader race, ageUsed prediction and not time to event analysisData was limited to that collected as part of the clinical trial N=1767 pts subset–Canada, US, Brazil, Argentina carefully curated data Prediction of 30-Day All-Cause Readmissions in Patients Hospitalized for HF –ML vs. LR Frizzell, et al, JAMA Cardiol. 2017;2(2):204-209 Registry-based study linking patients from the Get With the Guidelines Heart Failure registry with Medicare dataPrimary outcome -readmission within 30 days following discharge for index HF hospitalizationN=56,477 patientsAge > 65N-250 variablesThe study sample was randomly divided into training(70% of sample) and validation(30% of sample) cohorts Use of ML algorithms did not lead to improved prediction of 30-day HF readmissions compared with traditional statistical models. The LINK-HF Multicenter Study N=100 Stehlik et al, Circ Heart Fail. 2020;13:e006513 An ML analytics algorithm using continuous remote monitoring data from a wearable sensor will predict HF rehospitalization with ≥70% sensitivity at a specificity level of 85%. 87 completed 30 days74 completed 90 daysClinical alerts preceded hospitalization by a median time between 6.5 and 8.5 days AI challenges§HYPE??? AI winter§Disconnects between reality and expectation§Biased data for AI model development§What is our goal? Do we real expect CI = 1???§Applying AI outside of populations represented in the training and validation sets§Disregarding the ‘law of unintended consequences’§Impact on care or ptclinical relationship§Limited data on ACTUAL effects on ptoutcomes and cost of careMatheny et al, JAMA 2020;323:509 A Strategic Partnership Proposal presented to Rogers CommunicationsSeptember 2017Dr. Heather RossTed Rogers and Family Chair in Heart FunctionSite Lead, Ted Rogers Centre for Heart Research Preventing Heart Failure Hospitalizations with Artificial IntelligenceDate: April 17th, 2020Dr. Heather Ross MD, MHSc(Bioethics), FRCPC, FACCHead, Division of CardiologyPfizer Research Chair in CardiologyLorretta A. Rogers Chair in Heart FunctionSite Lead, Ted Rogers Centre for Heart ResearchProfessor of MedicinePeter Munk Cardiac Centre