Biomedical Computing�6.872/HST.950 (Informatics)

Peter Szolovits, PhD (psz@mit.edu)Course Secretary: Fern Keniston (fernd@mit.edu)

On-line Web Site:https://people.csail.mit.edu/psz/6.872Monday, Wednesday 9:30-11:00am, 36-155

Who Am I?

• Background is in Artificial Intelligence (AI)• Medicine as an application of AI, in 1974• Knowledge-based systems for diagnosis, therapy selection & management• Early emphases

• Reasoning based on causal, pathophysiological models• Time, space and probability

• 1990’s revolution brought tons of data; e.g., MIMIC contains terabytes of ICU data• Shift from knowledge-based to empirical models• Data mining and machine learning: logistic regression, naive Bayes, Bayes

networks, neural nets, support vector machines, conditional random fields, deep neural nets, etc.

• Challenge: combining general knowledge with experience• Natural Language Processing to enable use of clinical data locked up in narratives• Shift from institutional to personal systems: Guardian Angel (http://ga.org)• NRC studies on patient privacy, health care challenges to IT, CMS IT revitalization

Outline

• Changing nature of health care• More based on biology, scientific theory, engineering analysis• Computerized

• What is addressed?• Diagnosis, Therapy, Prognosis• Public Health• Mortality, morbidity, cost

• The Care Cycle• Genomics and Precision Medicine• What is Health?

Why do People Care?

• “The Learning Health Care System” — Institute of Medicine• Growing availability of clinical data on past patient care

• E.g., Partners has a Research Patient Data Registry (RPDR) containing records of ~4M patients

• E.g., Kaiser Permanente has ~32 petabytes of data (much of it imaging)• 2009 Stimulus Bill subsidizes hospitals, clinics, doctors’ offices to install

electronic health records to meet “meaningful use” criteria• Improving machine learning methods to exploit such data

• “The $1000 Genome” — spread of genomic medicine• Phenotype = f(Genotype, Environment); what is f?• Theories of genetic causation: (a) Mendelian, (b) Common Variants, (c) Rare

Variants, (d) neo-Mendelian, (e) ???• Cost

• E.g., a Chevy contains more in employee health care costs than steel!

Inexorable GrowthHealthcare Spending as % of GDP

0

4.5

9

13.5

18

198219841986198819901992199419961998200020022004200620082010201220142016

USA USA (OECD) France UK UK (OECD)Taiwan

Estimates Vary…

6

https://en.wikipedia.org/wiki/List_of_countries_by_total_health_expenditure_per_capita

The Challenge of Productivity

• Productivity = goods & services produced / resources consumed• Staying in place is not good enough

0%

10%

20%

30%

40%

50%

60%

70%

80%

1840 1850 1860 1870 1880 1890 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

farm labor (% of population)

0

10

20

30

40

50

60

70

80

1840 1850 1860 1870 1880 1890 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

Farm productivity

What is the Productivity of Health Care?

• Hard to know what outputs to measure• Life expectancy• Childhood mortality• Quality of life

• Difficult to estimate resources consumed

Comparison of 1901 and 2003 survival curves for US white women.From http://www.state.gov/g/oes/rls/or/81537.htm

0"10000"20000"30000"40000"50000"60000"70000"80000"90000"100000"

0" 5" 10"

15"

20"

25"

30"

35"

40"

45"

50"

55"

60"

65"

70"

75"

80"

85"

90"

95"

100"

105"

110"

115"

MODEL based on real 2004 mortality data

UK

Standing Still is Not Good Enough

Chris Dede, Harvard Ed

http://theincidentaleconomist.com/wordpress/the-health-care-productivity-problem/

Less Efficient Sectors Come to Dominate

• Hypothetical sector growth over 30 years, assuming constant demand for each sector

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Mining

Construction

Other services

Health, education

Arts

Prof. services

Finance

Transport

Nondurable

Utilities

Retail trade

Wholesale trade

Argiculture

Information

Durable goods

Hypothetical sector growth over 30 years, assuming constant demand for each sector. Productivity rates from OECD 2009.

“The Learning Health Care System”

• “one in which progress in science, informatics, and care culture align to generate new knowledge as an ongoing, natural by-product of the care experience, and seamlessly refine and deliver best practices for continuous improvement in health and health care” —IOM

• Needs not currently met:• Comprehensive collation of all clinical, social, demographic, behavioral, ... data that

are now captured in the health care system• Routine capture of novel data sources:

• genomes, gene expression, etc.• environmental factors (e.g., metagenomics)• physiological response to life situations

• (related to fitness and wellness)• Technical infrastructure

• Storage and analysis of truly “big data”• Incentives and demonstrations of utility

What is Evidence-Based Medicine?

• Randomized Controlled Clinical Trials• E.g., is drug A more effective than drug B for condition X?• Narrow selection of patient cases and controls• Careful collection of systematically organized data• Statistical analysis of outcomes• => Statistically significant conclusions

• But:• Heterogeneity: Most cases to which RCT results are applied do not fit trial criteria• Short Follow-Up: Trials run for limited times, but use is longer• Small Samples: Some effects are rare but devastating

• Instead: consider every patient’s experience as a source of knowledge by which to improve health care

Vioxx and Heart Attacks

time series of MI-related hospitalizations by Poisson regressionwhere the two drugs were included as indicator variables. Themodel for rofecoxib revealed a positive relationship with counts ofMIs (rate ratio 1.14; 95 percent confidence interval, 1.09 to 1.19;P,0.001), explaining 28.7 percent of the deviance. The model forcelecoxib also revealed a significant positive relationship withcounts of MIs (rate ratio 1.09; 95 percent confidence interval, 1.01to 1.18; P= 0.022), explaining 5.2 percent of the deviance.Finally, we measured the yearly impact of estimated celecoxib

and rofecoxib use on hospitalizations due to MIs as a dose-response relationship using national prescription data (Table 1).Prescriptions for celecoxib and rofecoxib peaked in 2001 with25.41 and 27.06 million prescriptions dispensed nationally,respectively. Incidence of MI-related hospitalizations also peakedin 2001 with 62.3 hospitalizations per 10,000 (Figure 2). Yearlycounts of MIs were significantly predicted by prescriptions forrofecoxib (0.7% increase in MIs per million prescriptions; 95percent confidence interval, 0.2% to 1.3%) and marginally forcelecoxib (0.6% increase in MIs per million prescriptions, 95percent confidence interval, 20.005% to 1.8%;). For every millionprescriptions of both rofecoxib and celecoxib, we found a 0.5percent increase in MI (95 percent confidence interval, 0.1 to 0.9),explaining approximately 50.3 percent of the deviance in yearlyvariation of MI-related hospitalizations.We also find a significant negative association between mean

patient age and prescriptions for celecoxib and rofecoxib(Spearman correlation, 20.67, P,0.05). Over the years whenrofecoxib was not available, the mean age was 66.5 compared toa mean age of 65.4 over the period of market availability. Figure 3

shows the nadir in age corresponding to the peak of rofecoxib andcelecoxib prescriptions.

Anchor DiagnosesTo demonstrate that our findings were not simply based onchanging healthcare utilization, we investigated temporal trendsfor hospitalizations for three other diagnoses. Though showinga positive linear increase, abdominal aortic aneurysm (Spearmancorrelation, 0.17; P = 0.66), appendicitis (Spearman correlation,0.11; P= 0.78), and pneumonia and influenza (Spearmancorrelation, 0.071; P= 0.86) were not associated with trends inMI. Further, only trends in inpatient visits for MI weresignificantly correlated with national prescription rates forrofecoxib and celecoxib (Spearman correlation 0.93; P,0.001).

DISCUSSIONOur findings demonstrate striking longitudinal trends in thepopulation rate of MI and strongly suggest a population-levelimpact of rofecoxib and celecoxib on MI. Temporal trends ininpatient stays due to MIs are tightly coupled to the rise and fall ofprescription of COX-2 inhibitors, with an 18.5 percent increase ininpatient stays for MI during the time when rofecoxib was on themarket. These results are consistent with prior reports demon-strating cardiovascular effects of rofecoxib and celecoxib.[9,10,11,12,13,14,15,17,27] Further, mean age at MI appears tohave been lowered by use of these medications.Using population surveillance methods [28,29,30] we confirm

a sharp rise in adverse cardiovascular events within eight months

Figure 1. Cumulative sum (CUSUM) chart of monthly incidence of hospitalizations due to myocardial infarction from January 1, 1997 to March30, 2006. Assigning 1997 and 1998 as a baseline period, a target mean of 47.1 hospitalizations due to myocardial infarctions per 100,000 anda standard deviation of 2.8 per 100,000.The threshold for CUSUM was calculated as 1.18, yielding an average-run-length of 50 (dashed red line). Anaberration, two standard deviations from baseline, was initially detected in January of 2000, with a CUSUM value of 1.37 and 51.5 myocardialinfarction-related hospitalization per 100,000 stays (solid line).doi:10.1371/journal.pone.0000840.g001

COX-2 Inhibitors and MI

PLoS ONE | www.plosone.org 3 September 2007 | Issue 9 | e840

“Trends in inpatient stay due to MI were tightly coupled to the rise and fall of prescriptions of COX-2 inhibitors, with an 18.5% increase in inpatient stays for MI when both rofecoxib and celecoxib were on the market (P<0.001). For every million prescriptions of rofecoxib and celecoxib, there was a 0.5% increase in MI (95%CI 0.1 to 0.9) explaining 50.3% of the deviance in yearly variation of MI-related hospitalizations.”

Brownstein JS, Sordo M, Kohane IS, Mandl KD (2007) The Tell-Tale Heart: Population-Based Surveillance Reveals an Association of Rofecoxib and Celecoxib with Myocardial Infarction. PLoS ONE 2(9): e840.

Adoption of Electronic Health Records

• Rapid adoption in hospitals, likely driven by 2009 “Stimulus Bill”• “Basic EHR systems include …:

patient demographics, patient problem lists, patient medication histories, clinical notes, electronic orders for prescriptions, laboratory results viewing, and imaging results viewing.” —ONC(Data from http://dashboard.healthit.gov/HITAdoption/?view=0)

• Positive factors for adoption: large size, urban location and HMO penetrationHealth information technology adoption in U.S. Acute care hospitals. Zhang NJ, et al. J Med Syst. 2013 Apr;37(2):9907

• … some reasons for suspicion about adoption rates

https://dashboard.healthit.gov/evaluations/data-briefs/non-federal-acute-care-hospital-ehr-adoption-2008-2015.php

National Adoption of Hospital EHR’s

15

https://dashboard.healthit.gov/evaluations/data-briefs/non-federal-acute-care-hospital-ehr-adoption-2008-2015.php

Adoption of Electronic Health Records

Vendors of HIT for Hospitals

17https://dashboard.healthit.gov/quickstats/pages/FIG-Vendors-of-EHRs-to-Participating-Hospitals.php

EHR Adoption in Doctors’ Practices

18https://dashboard.healthit.gov/quickstats/pages/FIG-Health-Care-Professionals-EHR-Incentive-Programs.php

Vendors of HIT for Doctors

19

Reducing Readmissions

• Starting Oct 1, 2012, hospitals are penalized by Medicare if a patient is re-admitted for the same condition within 30 days• Initial focus conditions: Acute Myocardial Infarction, Heart Failure, Pneumonia• Risk adjustments for demographics, co-morbidities, patient frailty

• Predictive models are worth $$$• E.g., Potentially avoidable 30-day hospital readmissions in medical patients:

derivation and validation of a prediction model. Donzé, et al., JAMA Intern Med. 2013 Apr 22;173(8):632-8• Retrospective study of 10,731 discharges from Partners HealthCare• 2,398 (22.3%) readmitted within 30 days, 879 (8.5%) judged avoidable• Eliminated planned re-admissions and those for other conditions• Logistic regression model predicts avoidable re-admission with AUC=0.71,

good calibration• Important variables: hemoglobin at discharge, oncology service, Na+ level at

discharge, procedures performed, non-elective admission, number of admissions in past 12 months, length of stay

118 papers on “predict readmission” since

2010 in Pubmed

“Meaningful Use” Adoption Medical Informatics

• Intersection of medicine and computing• Plus theory and experience specific to this

combination• =Medical Computing, ~Health Informatics

• Science• Applied science• Engineering

Types of Bio-Medical Informatics

• Cellular level: Bioinformatics, Systems Biology• Patient level: Clinical Informatics, Health I., Medical I., …• Population level: Public Health I.• Imaging Informatics

Bio-Medical Informatics

• Phenotype = Genotype + Environment• In humans, we rely on “natural experiments”• Measurements

– Genotype: sequencing, gene chips, proteomics, etc.– Environment: longitudinal surveys, etc.– Phenotype: clinical records, assembled to longitudinal data

Outline �(today)

• What is biomedical informatics?• BMI is defined by goals and methods of health

care• The genomic revolution• The science of health care– Genotype, phenotype, environment– From associations to mechanisms

• What is health?• Practice of health care• Challenges

Outline�(semester)

• Clinical and Genomic Data• Methods of modeling• Combining clinical and genetic data• “Translational medicine”• Engineering the health care system• Decision support to improve health care• Personalized medicine• Public health• The developing world•Your projects