AI for Health: A Partnership between Clinician, Patient & Data

Transcription

1 AI for Health: A Partnership between Clinician, Patient & Data Morris Wellman Assistant Professor of CSE University of Michigan, Ann Arbor

2 Machine Learning for Healthcare diagnoses medications vitals procedures locations lab results admissions genetics wearables imaging notes

3 Dartmouth Hitchcock Medical Center, 1955

4 Boston Children s Hospital, Today

5 The Problem: Physician shortage Physician burnout Medical errors Aging baby boomers and the millions of newly insured people are expected to tax the medical profession in coming years Source: Association of American Medical Colleges; American Academy of Family Physicians

6 The Problem: Physician shortage Physician burnout Medical errors

7 The Problem: Physician shortage Physician burnout Medical errors Causes of death, US, 2013 Medical Errors 251k

8 Time-Series Analyses Transfer Learning Representation Learning 10.0 EYE Contour Plot 7.5 θ DnK X Q θ K Regularization Techniques Z Y Causal Inference An Opportunity for AI

9 Healthcare-associated Infections Graft Versus Host Disease Cystic Fibrosis Alzheimer s Disease Diabetes An Opportunity for AI

10 CLOSTRIDIUM DIFFICILE Bacteria that takes over the gut We lack an effective clinical tool for measuring patient risk

11 Our Data-Centric Approach Leverage the contents of the electronic health records (EHR) to learn accurate models for predicting C. diff infections (CDI) Increasing Risk Applied to UM study population: n=191,014 features: d=4,836 AUROC=0.82 (95% ) predicted CDI 5 days in advance [JW, JG, EH; JMLR 2016] [JO,MM,, JW; ICHE 2018]

12 Where do we go from here? Collect more data -using ambient sensing technologies Higher Accuracy Greater Utility [Yeung et al., NEJM 2018]

13 Where do we go from here? Key characteristics for safe and meaningful adoption [JW, ES; CID 2017] Accurate à the model gets it right Credible (trustworthy) à agree, in part, with what is known Robust à adapt to changes over time & pop. Actionable à how to reduce a patient s risk

14 We need domain expertise Accurate à the model gets it right Credible (trustworthy) à agree, in part, with what is known Robust à adapt to changes over time & pop. Actionable à how to reduce a patient s risk domain expertise [JW, JO, HW, JW; KDD 2018] [PS,, JW; MLHC 2018] [MM, JG, JW; AAAI 2018]

15 Incorporating domain expertise about known risk factors [JW, JO, HW, JW; KDD 2018] J(, r) =k(1 r) k 1 + q k(1 r) k kr k2 2 expert knowledge r 2 {0, 1} d favors a solution that is sparse in set of unknown & dense in known increased credibility and increased robustness unknown known

16 diet activity locations wearables environment medication behavior interactions history diagnoses medications vitals procedures locations lab results admissions genetics wearables imaging notes

17 DIABETES 30.3 MILLION diabetics 50% $245 BILLION increased risk of death total medical costs and lost work wages What can we do? prevent or delay (type 2) manage (type 1 and 2) TYPE 1 TYPE 2 Body doesn t make enough insulin Body can t use insulin properly

18 Continuous Glucose Monitor: Feb- Mar 12am 12pm Avg. glucose 125 mg/dl Std. dev. 31 mg/dl Continuous Glucose Monitor: Mar Apr Avg. glucose 144 mg/dl 12am Lauren, PhD Student, Type I 12pm Std. dev. 42 mg/dl

19

20 AI for Automated Blood Glucose Control Ultimate goal - artificial pancreas Requirement generate good forecasts [IF, RP, LA, JW; KDD 2018] increases in blood glucose Closed loop system predict and deliver appropriate amount of insulin patient specific predictions decreases in blood glucose

21 AI augmenting current clinical practice Critical need for AI for health: Reduce costs and medical errors clinical decision support Increase global access mobile applications and telehealth Domains where ML algorithms excel Domains where humans excel [AE, et al., Nature 2017] To improve health we need both