AI: innovation in healthcare Dr Ali Connell Senior Research - PowerPoint PPT Presentation

AI: innovation in healthcare Dr Ali Connell Senior Research Scientist, Google Health UK

Health systems aim to meet The Quadruple Aim.. . 1 Betuer clinical outcomes Don Berwick Ex-Head of Medicare and Medicaid 2 Reduced costs 3 Improved patient experience 4 Improved experience of care provision Private and Confidential

...but face signifjcant challenges >50% of healthcare not Stafg burnout rates on the rise Care continues to be episodic Intractable increases in evidence based vs integrated healthcare costs Failure to deliver shared Unwarranted variation exists > 10% of patients experience Focus on illness at the expense decision making for patients across healthcare delivery harm in hospitals of prevention Private and Confidential

Can digital technologies and AI help solve some of these challenges?

AI is already working in tools used by billions of people Photos Gmail Translate

Groups across Google/Alphabet have been working in healthcare Google Health Cloud Google Assistant Brain Google Fit Streams Google Search Apigee Private and Confidential

Dr David Feinberg Bringing our health efgorus across Google into one health team

Together we are applying AI to a range of clinical challenges OPHTHALMOLOGY PATHOLOGY MEDICAL RECORDS RADIOLOGY

We are commitued to the highest standards of peer-reviewed academic research

Medical imaging

Training a model to read fundus photographs 130k examples No DR Mild DR Labeling tool Moderate DR 54 ophthalmologists 880k diagnoses Severe DR Proliferative DR Image Quality L/R eye Inception model Inception Field of View

The model now pergorms on par with retina specialists 100 Specialists Feb 2018 90 80 Sensitivity, % 70 Generalists 60 Dec 2016 50 40 0 2 4 6 8 10 1-Specificity, % J. Krause, et al. Ophthalmology , 2018 V. Gulshan, et al. JAMA , 2016

ARTICLE Age Prediction of Self reporued sex cardiovascular risk Cardiovascular risk factors from retinal fundus photographs via deep learning Predicting non-ocular health

Used in DR screening today 14

Electronic Health Records

A collaboration with Veterans Health Afgairs Test orders and results Procedures Microbiology Demographics and Unique patients 521,000,000 labs 181,000,000 instances 465,000 reporus health factors ~700k 227,000,000 instances Medications Discrete vitals Clinic visits, note titles Comorbidities Admissions, discharges and 293,000,000 prescriptions 356,000,000 records 123,000,000 instances 166,000,000 ICD-9 codes transfers 1,500,000 admissions 825,000 transfers

Continuous prediction of patient deterioration from EHR data Hospital admission 48 hours Billing code Sepsis Model AKI Moruality Hypoglycemia

Enabling anticipatory care in AKI Kidney function (GFR) 0 Time Model warns clinicians 48 hours Current AKI fmags

Enabling anticipatory care in AKI ROC AUC 92.1% Network predicted 90.2% of all AKI that needed dialysis Predicting 7 lab tests: substantial increases correctly predicted in 88.5% of instances

The challenges...

Bias and representative datasets Data security and privacy Dataset shifu Generalisation to new populations Evidence of safety & efgectiveness

“Technology projects fail because we don’t take suffjcient account of complexity. ” Prof Trish Greenhalgh

Understanding complexity Greenhalgh T. et al. Beyond Adoption: A New Framework for Theorizing and Evaluating Nonadoption, Abandonment, and Challenges to the Scale-Up, Spread, and Sustainability of Health and Care Technologies. J Med Internet Res. 2017 Nov 1;19(11):e367.

Real world complexities

Model metrics != clinical applicability

Model metrics != clinical applicability It is vital to know how the model, user and socio-technical environment interact

Understanding complexity: design and deployment Collaborative Design Process Human Computer Interaction Ethnographic Studies Dr. Paisan Ruamviboonsuk and the DR nurse team Pathumthani, Thailand Multidisciplinary perspectives

Understanding complexity: impacts Processes of care Patient outcomes Hospital metrics Qualitative review Economic analysis

For the fjrst time there’s enough intelligence, compute, data and structural supporu to make real impact

Clinician-centred solutions

Patients at the hearu of everything we do

Thank you @draconnell