Embed Size (px)
Citation preview
Suchi Saria Assistant Professor of Computer Science, Statistics, Biostatistics, and Health Policy
Malone Center for Engineering in Healthcare,
Johns Hopkins University
@suchisaria
Can Machines Amplify Expert Humans’ to Provide Care?
Home Monitoring for Parkinson’s Disease?
Home Monitoring for Parkinson’s Disease?
Voice Test
Balance TestGait Test
Location Monitoring
Motor Monitoring
Voice Test Gait TestDexterity Test Rest Tremor Test
“
”Intraday severity fluctuations captured using a mobile phone for a patient with 11
years of PD who worsened over 6 mths.
Voice Test
Balance Test
Gait Test
Motor Monitoring
Scleroderma: Systemic autoimmune diseaseChallenging to treat because of heterogeneity
in presentation and disease progression.
●
●
●
●
●●●
●●
●
40
60
80
100
120
0 5 10 15Years Since First Symptom
PFVC
●
●
●
●
●●●
●●
●
40
60
80
100
120
0 5 10 15Years Since First Symptom
PFVC
• Should we administer immunosuppressants, which can be toxic?
Mar
ker f
or lu
ng d
eclin
e (p
FVC
)
Affects 300K individuals; 80 other autoimmune diseases — lupus, multiple sclerosis, diabetes, Crohn’s — many of which are systemic & highly multiphenotypic.
●●●
●
●
●
●●● ●
●●
●●●
●●●
●● ●●
● ● ●
●
●
●● ●
●
●●
●
50
70
90
0 5 10 15 20Years Seen
PFVC
● ● ● ●
● ● ●
● ●
●
●●●
●
●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
RP
●
●
●
●
●
●
●●●
●
●●
●
●0
10
20
30
40
0 5 10 15 20Years Seen
TSS
● ● ●
● ● ● ● ● ● ●●●●●
●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
GI
●
●
●
●
●
●
●●
●●
●
●
●●● ●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●●
● ●
●
●
25
50
75
100
0 5 10 15 20Years Seen
PDLC
O
● ● ● ● ● ● ● ●●●●●●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
Hear
t
●●●
●
●
●●●● ●
●●● ●
●
●●●
●● ●●
● ●●
●
●
●●●
●
●
●
●
●
50
70
90
0 5 10 15 20Years Seen
PFEV
1
● ● ●
●
●
●
● ● ● ●
●●
●
●
●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
Mus
cle
●●
●
●
0
25
50
75
100
125
0 5 10 15 20Years Seen
RVSP
● ● ● ● ● ● ● ● ●●●●●●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
Kidn
ey
●
●
●
●●
●●
●●●●●
● ●● ● ●
●●
●
50
70
90
0 5 10 15 20Years Seen
PFVC
● ●● ●●●●● ● ●● ●●●●● ●● ●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
RP
●
●● ●●
●
●
● ● ●
● ●●●●● ●● ●●0
10
20
30
40
0 5 10 15 20Years Seen
TSS ●
●● ●
●●●● ● ●● ●●●●● ●● ●●
High
0
25
50
75
100
125
0 5 10 15 20Years Seen
GI
●●
●●●
●●●●● ● ●
● ● ●● ● ●
25
50
75
100
0 5 10 15 20Years Seen
PDLC
O
● ●
●
●●●●● ● ●● ●● ● ●● ●●
Low
0
25
50
75
100
125
0 5 10 15 20Years Seen
Hear
t
●
●
●
●
●●
●●●●
●●
●● ● ● ●
●●
50
70
90
0 5 10 15 20Years Seen
PFEV
1
●● ●●●●● ● ●● ●●●●● ●● ●●
High
0
25
50
75
100
125
0 5 10 15 20Years Seen
Mus
cle
● ● ●●
●
● ●
0
25
50
75
100
125
0 5 10 15 20Years Seen
RVSP
●
●● ●●●●● ● ●● ●●●●● ●● ●
High
0
25
50
75
100
125
0 5 10 15 20Years Seen
Kidn
ey
zi
fi
M
↵
yij tij
Ni
~�g G
~bi
⌃b
~⇢i ~xi⇢
⇡g
B⇢
zi
fi
M
↵
yij tij
Ni
~�g G
~bi
⌃b
~⇢i ~xi⇢
⇡g
B⇢
zi
fi
M
↵
yij tij
Ni
~�g G
~bi
⌃b
~⇢i ~xi⇢
⇡g
B⇢
Conditional random field (CRF) to model pairwise dependencies
Coupled Latent Variable Model to predict latent disease trajectory
Schulam, Saria. JMLR 2016
Adverse Event Onset
Is the patient at risk of a septic shock?
Early Warning Systems for Potentially Preventable Conditions?
Example: Septicemia is the 11th leading cause of death in the US
3.9 times higher mortality rate 2.4 times longer mean Length of Stay (LOS)
2.7 times higher mean cost
R. L. Fuller, et al., Health Care Financing Review, vol. 30, pp. 17-32, 2009
Scalable Joint Modeling for Reliable Event Prediction
Event Data
LongitudinalData
Septic Shock
At higher sensitivity, increase in the Positive Predictive Value from 6% to 50%
Joint Model of Trajectory and Event Data
Soleimani, Hensman, Saria. TPAMI 2017
Patients w/ shock identified a median time of 25 hours prior to shock onset.
Henry et al., Science TM 2015
Making Inferences available to providers
• Every hospital EMR is slightly different: data wrangling technologies to account for differences
• Monitoring every patient: flexible scalable inference and fault tolerant distributed computation
• Enabling providers to reason w/ model: joint human-machine reasoning
• Many other open questions:
• How do we communicate when to trust and when not to trust the resulting computation?
• How do we understand and estimate the different sources of error?
Discrete Events:
Laboratory
Interventions: Medicines, Procedures
Continuous p
hysiologic
measurements
Progress notes
Imaging Data
Administrative Claims
Genomic data
Sensors & Devices
Electronic Health Data
From 2008 to 2014, hospitals with EHRs rose to 75% from 9%, and in doctors’ offices rose to 51% from 17%.
$32+ billion effort went into digitization. Benefits?
