Mining of electronic health records (EHRs) has recently gained importance. However, most efforts are restricted to analyzing drugs, diseases and their associations. In biomedical research, network analysis has provided the conceptual framework to interpret protein-protein interactions or gene-disease association networks via large-scale network maps. We analyze associations between drugs, diseases, devices and procedures mined from EHRs using network analysis to extract hidden “modules of care” for hypothesis generation. In particular, we annotated the textual notes of the EHRs of one million patients in the Stanford Clinical Data Warehouse with disease, drug, procedure and device terms using ontologies such as SNOMED-CT or RxNorm. We then used standard co-occurrence statistics to establish associations between these clinical concepts and to construct networks. Hidden modules of care - clusters of diseases, drugs, procedures, devices – useful for hypothesis generation are extracted through network analysis approaches and visualized using Cytoscape. We present a study for comparative effectiveness of Cilostazol vs. a control group in peripheral artery disease (PAD) patients (see Figure 1) and compare our results derived from the network analysis against standard methods such as regression analysis. We believe that network analysis allows us to uncover hidden (“latent”) modules of care not detected through standard approaches, which do not account for the connectivity of the clinical events and entities.

1. Network(analysis(of(unstructured(
EHR(data(for(compara8ve(
effec8veness(research(
Anna$Bauer)Mehren$
ShahLab,$Stanford$Center$for$Biomedical$Informa:cs$

2. Background(>(PAD(
• Peripheral$artery$disease$(PAD):$obstruc:on$of$infra)renal$
abdominal$aorta$and$lower$extremity$arteries$$
• Symptoms:$IntermiEent$claudica:on$$
• Risk$factors:$Smoking,$Diabetes,$Obesity,$Hypertension,$Age$
• Treatments:$Lifestyle$changes,$Surgical$and$endovascular$
interven:ons,$Pharmacotherapy$$
!$Cilostazol$
7/17/12 NetBio SIG, 2012 2

3. Background(>(Cilostazol(
• Reversible$selec:ve$inhibitor$of$phosphodiesterase$(PDE)$type$III$
• Long)term$oral$milrinone$therapy$associated$with$life)
threatening$cardiovascular$events$in$conges:ve$heart$failure$
(CHF)$pa:ents$
$
$$$$$Black$box$warning:$$
$
$ Cilostazol is contraindicated in patients with CHF
$
$$$$$Hypothesis:$
Cilostazol is not associated with increased risk of
major adverse cardiovascular events (MACE)
7/17/12 NetBio SIG, 2012 3

4. STRIDE( NCBO(
• 1.6$million$pa:ents$ • We$create$and$maintain$a$
library$of$more$than$300$
• 15$million$visits$ biomedical$ontologies$
• 25$million$coded$ICD9$ • We$build$tools$and$web$
diagnoses$ services$to$enable$the$use$
of$ontologies$
• 9.5$million$clinical$free$
text$notes$ • We$use$ontologies$for$
• pathology$ annota:ng$EHRs$to$allow$
• radiology$ clinically$relevant$studies$
• transcrip:on$reports$
7/17/12 NetBio SIG, 2012 4

5. Networks(for(compara8ve(effec8veness(
• Networks$have$been$successfully$applied$in$biology$
• Can$we$use$networks$for$clinical$studies?$
• We$want$to$use$networks$for$
• Cohort$building$
Roque et al, PLoS Comput Biol. 2011; 7(8): e1002141.
• Analysis$of$clinical$outcome$
$
7/17/12 NetBioSIG 2012 5

6. From(clinical(notes(to(pa8ent(feature(matrix(
BioPortal – knowledge graph Text clinical note
Creating clean lexicons
Drugs Diseases Annotation Workflow
Term$–$1$
:$ Term recognition tool NegEx
:$ NCBO Annotator Patterns
Term$–$n$
Devices Procedures Terms Recognized
Patient feature matrix
• Negation detection
Further$Analysis$
Select cohort of
• FH detection
Interest
Negation, FH detected
7/17/12 NetBio SIG, 2012 6

7. Dimension(reduc8on(using(ontologies(
7/17/12 NetBio SIG, 2012 7

8. Cohort(building(–(restrict(by(follow(up(8me( Follow up time in
peripheral artery disease patients
3000 Patient timeline
PAD Last note
2500
tPAD tlast t
2000
Follow up time
Frequency
1500
1000
5757 PAD patients
500
0
0 365 1000 2000 3000 4000 5000 6000
follow up time in 30 day intervals
7/17/12 NetBio SIG, 2012 8

9. Cohort(building(–(set(8me(
Patient timeline
PAD CIL Last note
Cilostazol t
patient t PAD t CIL= t0 t last
Matching Outcome
Median = 22 days
PAD Last note
Other PAD t
patient t PAD t0 t last
22 days
before after
7/17/12 NetBio SIG, 2012 9

10. Matching(
Pa8ent>pa8ent(similarity(network( Propensity>score(matching(
1159(
1. Choose$variables$for$matching$
Pa:ents$
5776(
A∩B
J(A, B) = 1−
A∪B
Nearest(neighbor(
Pa:ents$ 5776( Matching((1:1)(
0 0.6 0.8 0.6 0.5 0.8 2. Compute$PS$based$on$variables$(logis:c$
Cilostazol( 0 0.8
0
0.9
0.7
0.8
0.9
0.3
0.9
regression)$
Pa:ents$
0 0.7 0.8
3. Nearest$neighbor$matching$(1:1)$
0 0.8
0
5776(
Control(
Concepts$
Concepts$ 1 0 1 0 0 1 17(
Pa:ents$
1 0 1 0 0 1 1159( 0 0 1 1 1 0
1 1 0 0 0 1
Pa:ents$
0 0 1 1 1 0
1 0 1 0 0 1
1 1 0 0 0 1
0 0 1 1 1 1
1 0 1 0 0 1
0 0 1 0 0 1
0 0 1 1 1 1
446( 0 0 1 0 0 1
446(
7/17/12 NetBio SIG, 2012 10

11. Matching(>(comparison(
!
Variable Before Matching Patient-patient similarity Propensity score !
network matching
!
Treatment Control p-value Control p-value Control p-value !
Demographics
(n= 223) (n= 5534) (n= 223) (n= 223) !
Age (at indication onset), mean (sd) 71.22 70.43 0.295 72.05 0.41196 70.87 0.74704
(11.02) (12.46) (10.62) (11.51)
Gender (female), n (%) 37.22 45.94 0.0090723 36.65 0.84432 35.87 0.76777
Race , (%)
Asian 8.52 7.41 0.56047 6.33 0.36671 10.31 0.50525
Black 2.69 3.71 0.36423 3.61 0.588 0.90 0.15684
Unknown 22.87 26.17 0.25365 22.17 0.82063 20.63 0.57404
White 65.47 62.22 0.31854 67.87 0.54642 67.27 0.68645
Comorbidities
Coronary artery disease, n (%) 5.38 6.47 0.48345 4.98 0.83089 6.28 0.69516
Congestive heart failure, n (%) 25.56 22.84 0.36255 20.36 0.21361 30.49 0.36255
Hypertension, n (%) 10.76 11.31 0.79597 9.50 0.75135 10.31 0.87893
Co-prescriptions
Beta blocking agents, n (%) 75.34 60.77 1.6611e-06 69.68 0.20252 74.89 0.90693
ACE inhibitors, plain, n (%) 78.03 69.57 0.0032697 67.87 0.013595 78.92 0.81386
Platelet aggregation inhibitors excl. 91.93 79.00 8.1983e-11 89.59 0.41347 95.51 0.072929
heparin, n (%)
Vasodilators, n (%) 32.29 26.36 0.064893 31.67 0.83902 37.22 0.28753
History of
Cardiac arrhythmia, n (%) 32.29 32.17 0.96957 23.08 0.03383 33.18 0.84006
Stroke, n (%) 17.94 18.31 0.88877 15.84 0.61129 21.52 0.33903
Myocardial infarction, n (%) 17.94 15.87 0.43015 13.58 0.23919 19.73 0.63765
Vascular surgical procedures, n (%) 74.44 47.71 < 2.22e-16 65.61 0.048937 74.44 1
Bypass surgery, n (%) 41.70 26.56 1.0506e-05 36.20 0.24269 40.36 0.75073
7/17/12 NetBio SIG, 2012 13

12. Cohort(building(–(set(8me(
Patient timeline
PAD CIL Last note
Cilostazol t
patient t PAD t CIL= t0 t last
Matching Outcome
Median = 22 days
PAD Last note
Other PAD t
patient t PAD t0 t last
22 days
before after
7/17/12 NetBio SIG, 2012 14

13. Outcome(comparison(
Concept>concept(network( Dispropor8onality(analysis(
$ $
$ $
&β β β β #
$
10 11 13 1d
$ !
$β 21
β 20
β 23
β ! 2d
$β β β β !
31 32 30 3d
$ !
$ !
$ !
$
$β !
% β β β !
"
Compute edges
d1 d2 d3 d0
cofreqcid1,cid 2
ascid1,cid 2 = log(
freq1cid1 * freq2 cid 2
) cilostazol
n - 0.5 2.5 4 10 Logistic regression
- 0.5 2.5 4 10
- 1.3 2 0 Odds ratios − cilostazol vs. control
- 1.3 2 0
- 0 0 1 0 0 1 1 Atrial fibrillation ●
similarity
●
- 0 0 psm
0 1 1 0 1
- 1 Bypass surgery ●
- 1 0 1 0 0 1 ●
cephalexin cane - Cardiac arrythmia ●
0 1 1 0 0 ●
-
ultrasound Coronary stents ●
doppler
imaging 1 0 1 0 0
control
ultrasonography
●
amoxicillin
doppler studies 1 1 1 1 0 Death ●
angioplasty
atherectomy ●
revascularization
wheelchair
bypass graft Electric countershock ●
vascular
cilostazol ●
surgical
procedures
diagnostic
hydralazine congestive heart imaging
Myocardial infarction ●
pneumonia surgical revision
failure ●
heart failure
bypass Stroke ●
nifedipine
testosterone ●
amiodarone pravastatin vascular
diseases
carotid Vasulcar surgical procedures ●
pantoprazole insulin glargineendarterectomy
●
obesity
transplantation ramipril fentanyl Ventricular fibrillation ●
zolpidem
trimethoprim ●
decompressive
incision Ventricular tachicardia ●
coronary sulfamethoxazole
diazepam angiography ●
fluoroscopic
heart angiography
transplantation
tacrolimus
temazepam
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
7/17/12 NetBio SIG, 2012 15

14. Concept>concept(network(
cephalexin cane
doppler ultrasound
ultrasonography imaging
amoxicillin
doppler studies
angioplasty
atherectomy
wheelchair Vascular surgical procedures
revascularization
bypass graft
vascular
cilostazol
surgical
procedures
diagnostic
hydralazine congestive heart imaging
pneumonia surgical revision
failure
Heart failure bypass
heart failure
nifedipine
testosterone
amiodarone pravastatin vascular
diseases
carotid
pantoprazole insulin glargineendarterectomy
obesity
transplantation ramipril fentanyl
zolpidem
trimethoprim
decompressive
Heart incision
coronary sulfamethoxazole
diazepam angiography transplantation
fluoroscopic
heart angiography
transplantation
tacrolimus
temazepam
7/17/12 NetBio SIG, 2012 16

15. Dispropor8onality(analysis(
Odds ratios − cilostazol vs. control
Atrial fibrillation ●
similarity
●
psm
Bypass surgery ●
●
Cardiac arrythmia ●
●
Coronary stents ●
●
Death ●
●
Electric countershock ●
●
Myocardial infarction ●
●
Stroke ●
●
Vasulcar surgical procedures ●
●
Ventricular fibrillation ●
●
Ventricular tachicardia ●
●
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
7/17/12 NetBio SIG, 2012 17

16. Conclusions(
• Technical$
• Pa:ent)pa:ent$network$based$on$similarity$
in$annota:ons$can$be$used$for$cohort$
building$
cephalexin cane
• Concept)concept$network$based$on$co)
doppler ultrasound
ultrasonography imaging
amoxicillin
doppler studies
angioplasty
atherectomy
revascularization
wheelchair
occurrence$in$clinical$notes$can$be$used$for$
bypass graft
vascular
cilostazol
surgical
procedures
diagnostic
hydralazine congestive heart imaging
pneumonia surgical revision
failure
bypass
heart failure
nifedipine
analyzing$clinical$outcome$
testosterone
amiodarone pravastatin vascular
diseases
carotid
pantoprazole insulin glargineendarterectomy
obesity
transplantation ramipril fentanyl
zolpidem
trimethoprim
decompressive
incision
coronary sulfamethoxazole
diazepam angiography
fluoroscopic
heart angiography
transplantation
tacrolimus
temazepam
• Medical$
Cilostazol is contraindicated in patients with CHF
• No$significant$differences$in$MACE$
comparing$cilostazol$vs.$control$$ Cilostazol is not associated with increased risk of
major adverse cardiovascular events (MACE)
$
7/17/12 NetBio SIG, 2012 18

17. Future(work( angioplasty
wheelchair
cilostazol
• Technical$ pneumonia
hydralazine congestive heart
failure
heart failure
heart failure
• Improve$aggrega:on$of$concepts$ amiodarone pravastatin
carotid
pantoprazole insulin glargineendarterectomy
• Might$improve$cohort$building$
obesity
transplantation
zolpidem
trimethoprim
• Useful$for$outcome$analysis$based$on$networks$ coronary
angiography
sulfamethoxazole
• Medical$
• Repeat$study$in$different$hospital$
• Profile$PAD$pa:ents$using$all$annota:ons$using$concept)
concept$networks$for$hypothesis$genera:on$
7/17/12 NetBio SIG, 2012 19

18. Acknowledgements(
• ShahLab$
• Nigam$Shah$
• Paea$LePendu$ • NCBO$Team$
• Rave$Harpaz$ • Mark$Musen$
• Srinivasan$Iyer$ • NIH$funding$
• Amogh$Vasekar$
$
• Kenneth$Jung$
• STRIDE$Team$
• Tanya$Podchiyska$
• Medical$collaborator$
• Todd$Ferris$
• Nicholas$Leeper$
$
• John$Cooke$
7/17/12 NetBio SIG, 2012 20