1、Using Artificial Intelligence to prevent healthcare errors from occurringSECOND GLOBAL MINISTERIAL SUMMIT ON PATIENT SAFETYCONFIDENTIAL AND PROPRIETARYAny use of this material without specific permission of McKinsey&Company is strictly prohibitedPresentation|29thMarch 20172McKinsey&CompanyWhy is Art
2、ificial Intelligence/Machine Learning different,and why now?1Where is the opportunity in patient safety/patient care?2How can we enable change?3Agenda for today3McKinsey&CompanyWhy is Artificial Intelligence/Machine Learning different,and why now?1Where is the opportunity in patient safety/patient c
3、are?2How can we enable change?3Agenda for today4McKinsey&CompanyWhy is machine learning different?How Traditional stats sees itTraditional stats will fit a predetermined“shape”into the phenomenon(e.g.linear,quadratic,logarithmic models)the square peg into the round hole!The actual phenomenon(real hi
4、storical data)Real life phenomenon come in“all shapes and flavors”showing patterns that are usually complex,non-linear and apparently disorganizedHow Machine Learning sees itWhile ML algorithms are adapting themselves by spotting&recording patterns without clinging to any predetermined corsetv1v25Mc
5、Kinsey&CompanyImproving injury prediction in premiership footballSOURCE:QuantumBlack90Improvement in accuracy of forecasting non-impact injuries:Forecast 170 of 184 non-impact muscle injuries across four squads and two years%All content Copyright 2017QuantumBlack Visual Analytics Ltd.ImpactApproachF
6、ull data capture of all network activities,customer and geolocation data to predict faultsSituationLeading European telecoms and broadband provider needing to improve faultsFaults predicted75%Predicted faults prevented90%60%Inbound service calls reductionIndustry leading customer satisfaction scoreT
7、elecomsImproving fault rateAll content Copyright 2017QuantumBlack Visual Analytics Ltd.ImpactApproachSituationStation was suffering from low availability,unplanned maintenance was 3x the global averageThree components found to be key drivers of failureGoal to reduce unplanned losses due to mill fail
8、uresCollected data 7 different data bases and logsIdentified key failures and validated failure eventsDeveloped user interface to plan maintenance on time and implemented to shop floorUsed machine learning to define and predict failuresPredictive maintenance at a coal fired power station helped to r
9、educe mill downtime by 50%Model predicts failure 3 months in advance with 75%accuracyDown time reduced by 50%42-50%BeforeAfter8McKinsey&CompanyWhy is Artificial Intelligence/Machine Learning different,and why now?1Where is the opportunity in patient safety/patient care?2How can we enable change?3Age
10、nda for today9McKinsey&CompanyUS healthcare has seen increased adoption but only captured 10%of 2011 value estimate9547164108333SOURCE:Expert interviews;industry surveys on HIT adoption;McKinsey Global Institute analysis521621236Major changes in industry incentives brought about by the ACA and rapid
11、 uptake of EMRsfrom 30%in 2011 to 90+%today has contributed to general opening of healthcare dataSlow regulatory processes and monopolistic and segregated industry structure has resulted in only 10%progress toward realizing BDAA value from 2011CategoriesValue potential realized,$B2011 Value potentia
12、l,$BAverage value realized by adopters%2016 Adoption rate,%10%value capturedClinical Ops TotalLevers25-7010-2030-5090-10060-8010-3020-3035-5040-10060-80Public HealthPublic health surveillanceNew business modelsAggregating and synthesizing patient clinical records and claims datasetsOnline platforms
13、and communitiesR&DPredictive modelingStatistical tools and algorithms to improve clinical trial designAnalyzing clinical trials dataPersonalized medicineAnalyzing disease patternsAccoun-ting/PricingAutomated systemsHEOR and performance-based pricing plansComparative effectiveness researchClinical de
14、cision support toolsTransparency about medical dataRemote patient monitoringAdv.analytics applied to patient profiles10McKinsey&Company83.684.581.4ClinicalcodesCombined codes and measurementsClinical measure-mentsImpactProof of concept established for machine learning based predictive modelling to d
15、etect adverse drug events using EHR dataSummaryMachine learning(random forest model)can be used to accurately predict ADEs(adverse drug events)using data from EHRs(electronic health records)Clinical coding data(used to record diagnoses and prescribed drugs)has higher predictive performance than clin
16、ical measurements though both used in combination have higher predictive performance for certain ADEsFeature selection to reduce dimensionality and sparsity further improves predictive performanceADEs are responsible for 5%of hospital admissions internationallySystems based on voluntary spontaneous reporting(pharma-covigilance)fail to capture 94%of ADEsExtracting data for machine learning from EHRs0.750.760.66Combined codes and measurementsClinical measure-mentsClinicalcodesp 0.007p 0.0001SOURCE
免费区 