The Good News. More storage capacity allows information to be saved Economic and social forces creating more aggregation of data

Transcription

1 The Good News Capacity to gather medically significant data growing quickly Better instrumentation (e.g., MRI machines, ambulatory monitors, cameras) generates more information/patient More storage capacity allows information to be saved Economic and social forces creating more aggregation of data

2 The Bad News Clinicians spending ever more time: Studying data about their patients And still ignoring most of the data Tracking onslaught of new medical data And not keeping up One of many contributors to waste in healthcare system

3 Waste In the U.S. Healthcare System Inst. of Medicine

4 Predictive Models Can Help Input: Information about an individual, environment, treatment Output: Probability of an event (good or bad) occurring Information about why that is the probability Uses Improved outcomes Earlier detection Matching treatments to patients Optimization of processes Resource allocation Clinical trials Hypotheses about causality

5 Building Predictive Models Patient data Clinical predictive models

6 What Is Machine Learning? Modern statistics meets optimization

7 How Are Things Learned? Memorization Accumulation of individual facts Limited by Time to observe facts Memory to store facts Generalization Deduce new facts from old facts Limited by accuracy of deduction process Essentially a predictive activity Assumes that the past predicts the future

8 Machine Learning Methods Many different ones, all try to learn a model that is a generalization of examples Supervised: given a set of feature/label pairs, find a rule that predicts the label associated with a previously unseen input Unsupervised: given a set of feature vectors (without labels) group them into natural clusters All have four components Training data and evaluation method Representation of the features Objective function and constraints Optimization method for learning the model

9 Some Considerations in Machine Learning Choosing training data and evaluation method, e.g., Holdout set or cross validation Representation of the features, e.g., How to represent an ECG signal Objective function and constraints, e.g., Maximize AUROC Optimization method for learning the model, e.g., E.g., logistic regression. SVM, deep learning

10 Machine Learning: the Big Data Challenge Too Fast The data arrives in volume each day Fields, distributions, etc. constantly changing Models need to adaptive Too Hard Multiple modalities Always incomplete, often incorrect, outcomes ambiguous Ground truth often hard to come by Site-to-site variation Too Big: images, videos, signals 10 s of millions of patients, billions of bits/patient Learning needs to parallelized, use sub-linear algorithms,

11 The Big Data Challenge Too Fast The data arrives in volume each day Fields, distributions, etc. constantly changing Too Hard Multiple modalities Always incomplete, often incorrect, outcomes ambiguous Ground truth often hard to come by Site-to-site variation Too Small Never enough obviously relevant data

12 An Example: Predicting Chronic Kidney Disease Total patients : 3.5 million patients 1 CKD patients : 44,000 patients w/ hypertension and/or diabetes : 10,000 patients Other Inclusion criteria: 1. Early stage CKD, 2. At least 2 years of medical history Final population size : 6400 patients

13 To Make it Worse: Missing and Misleading Data Uneven sampling Inaccurate coding Varying levels of abstraction (beware of ICD-10) Underlying condition: Acute Systolic Heart Failure Recorded information: Heart Failure No ground truth for labels Diagnoses can be wrong or missed Most tests not ordered 428 Heart failure Congestive heart failure Left heart failure Systolic heart failure."."."" Unspecified Acute Systolic heart failure Chronic Systolic heart failure.".".""

14 Dealing with Not Enough Data Novel approaches to feature engineering Reducing dimensionality Finding the right levels of abstraction Computational biomarkers E.g., to convert 100 s samples/sec. to useful data Making productive use of data from other sources Different features Different outcomes Different conditional distributions Novel machine learning techniques

15 Computationally-Generated Biomarkers Biomarker (broadly defined) A characteristic that can be objectively measured and provides an indicator of normal or pathological processes or expected responses to a therapeutic intervention Computationally-Generated Biomarker A biomarker generated by applying computation to medical data

16 Cardiovascular Risk Stratification Acute coronary syndrome (ACS) common: ~1.25M/year in U.S. 15% - 20% of these people will suffer cardiac-related death within 4 years Have lots of good treatments Eat better Get an ICD Choosing who should get which (or any) is hard Fine-grained risk stratification the key

17 Approaches to Identifying High Risk Cases Clinical characteristics E.g., gender or high blood pressure Traditional biomarkers E.g., cholesterol levels Echocardiography E.g., LVEF Electrocardiography (ECG) We combine all of these Innovation on use of ECG Jason Grow/The Human Face of Big Data

18 Morphological Variability Detect signs of small and transient electrical instability in myocardium by detecting minor differences in shape of normal appearing heart beats Invisible to human eye Patient (died)! (MV > 2 x high risk threshold)! Patient 1593 (survived)! (MV < 1/2 x high risk threshold)!

19 Data from MERLIN TIMI-36 dataset (~1B beats) About 4,500 ACS patients 1 year follow-up for cardiovascular death (193 events)

20 MV and Other Risk Variables Multivariate Analysis of Patients with LVEF > 40%* Parameter Adjusted Hazard Ratio (*) 95% Confidence Interval P Value MV BNP BMI CrCl * Also adjusted for age, hypertension, diabetes, hypercholesterolemia, prior MI, prior angina, ST changes, cardiac markers

21 Using Advanced ML to Tame HAIs Medicare Shift Fails to Cut Hospital Infections -- Oct. 10 th 2012 Hospitals Fail to Take Simple Measures to Thwart Deadly Infections, Survey Says -- April 8 th, 2013 More Aggressive Action Urged to Curb Hospital Infections -- May 29 th, 2013

22 Health Care Associated Infections

23 Example: Predicting Clostridium difficile Bacteria that takes over the gut Transmitted through the mouth Causes severe diarrhea, intestinal diseases Treatment: metronidazole and oral vancomycin 20% of cases relapse within 60-days 178,000/year in U.S. Approx. same as invasive breast cancer

24 The Data Hospital A: ~180 beds and ~10,000 admissions per year Hospital B: ~250 beds and 15,000 admissions per year Hospital C: >900 beds and >40,000 admissions per year Hospitals have different patient mixes, related to incidence of C. diff.

25 Risk Factors Time Invariant Collected at the time of admission e.g., admission complaint, previous admissions, home meds Time Varying Changes during the hospitalization e.g., current meds, current procedures, current location, hospital conditions PATIENT RISK

26 Feature Spaces e.g., assigned location: 2NE e.g., lives in city X e.g., history of vancomycin Hospital A Hospital B Hospital C 1. Specific to A : Specific to B : Specific to C : Common to A,B & C : Common to A & B(only) : Common to A & C (only) : Common B & C (only) : 11

27 Incorporating Time Typical Approaches in Clinical Literature Calculate risk using only a snapshot of patient s state a. At time of admission [Tanner et al., 2009] b. n days before index event [Dubberke et al., 2011] Our Approach Calculate risk using the entire evolving risk profile 1a Estimated Risk 2 Time (days) 1b Index Event

28 Evaluating the Model Actual Class Applied to a held-out validation set of 25,000 admissions to a single hospital, from a single year, the model achieved an AUC=0.81 (95%CI ) Positive Negative Predicted Class High Risk Low Risk ,151 23,006 Fraction of Identified Cases An ~6X improvement! Predicted at least X days in advance

29 Technology Transition IRB for clinical study at MGH approved this week!

30 Taking My MIT Hat Off

31 Material omitted

32 Wrapping Up Exciting times at the intersection of Medicine and data analytics Data analysis has transformed many fields Biology from molecules to genetics to systems Sports Moneyball looks old fashioned Finance it takes a computer to lose $10,000,000/minute Medicine is just starting We have the opportunity to truly change the practice of medicine and the health care industry