Upload
arcadiasolutions
View
448
Download
0
Embed Size (px)
DESCRIPTION
The WINNING presentation from the Surescripts Technology Challenge -- Arcadia FluFender -- detecting influenza outbreaks in real time from antiviral prescriptions. For more info, go to http://www.arcadiasolutions.com or http://www.twitter.com/ArcadiaHealthIT
Citation preview
1CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Arcadia FluFenderPrepared for the SureScripts Technology Challenge
2CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Viral infectious disease
Spreads via direct transmission, airborne, hand-to-eye/nose/mouth
A healthy adult can be contagious beginning one day before symptoms develop and up to seven days after becoming sick
3CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
5-20% of the population get the flu
in a typical year
The flu tends to be seasonal
In the U.S., Oct. – May, peaking in Jan. or Feb.
4CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Between 1976-2006, estimates of flu-associated deaths in the U.S. range from 3,000 – 49,000 people
Vaccination and early treatment are effective prevention strategies
5CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Why track with prescription antivirals?Indicates the presence of influenza but also of its intensity and the involved population
Antiviral prescriptions are a reliable indicator
Unlike OTC and ILI estimates, ePrescriptions contain verified data about the patient that reliably inform models
Antiviral prescriptions are a precise indicator
FDA requires that AVs be prescribed within 48h of symptoms, and only aggressive (non-complicated) cases
Symptoms generally commence 24-48h after contagion
6CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelFluFender performs short-term forecasting based on known clinician and patient behavior
Rx Rx
𝑅𝑥 𝑡 =
𝑖≠𝑘
𝛾
𝛿𝑖𝑏
𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
+𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
𝑅𝑥 𝑡, 𝑖 = Prescriptions at time 𝑡 in area 𝑖𝛾 = effect of close proximity𝜑 = likelihood of transmission
𝛿𝑖 = spatial separation between 𝑖 and 𝑘𝑏 = strength of distance effect
1.01d
1.01d
2 3 4
1
t=0
1-4d
3 4
𝜑
2
𝜑
2
1
1 Patient is infected with the flu virus.
7CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelFluFender performs short-term forecasting based on known clinician and patient behavior
Rx Rx
𝑅𝑥 𝑡 =
𝑖≠𝑘
𝛾
𝛿𝑖𝑏
𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
+𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
𝑅𝑥 𝑡, 𝑖 = Prescriptions at time 𝑡 in area 𝑖𝛾 = effect of close proximity𝜑 = likelihood of transmission
𝛿𝑖 = spatial separation between 𝑖 and 𝑘𝑏 = strength of distance effect
1.01d
1.01d
2 3 4
1
t=0
1-4d
3 4
𝜑
2
𝜑
2
1
1 Patient is infected with the flu virus.
2
2
Patient becomes highly symptomatic 1-4 days later.
8CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelFluFender performs short-term forecasting based on known clinician and patient behavior
Rx Rx
𝑅𝑥 𝑡 =
𝑖≠𝑘
𝛾
𝛿𝑖𝑏
𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
+𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
𝑅𝑥 𝑡, 𝑖 = Prescriptions at time 𝑡 in area 𝑖𝛾 = effect of close proximity𝜑 = likelihood of transmission
𝛿𝑖 = spatial separation between 𝑖 and 𝑘𝑏 = strength of distance effect
1.01d
1.01d
2 3 4
1
t=0
1-4d
3 4
𝜑
2
𝜑
2
1
1 Patient is infected with the flu virus.
2
2
Patient becomes highly symptomatic 1-4 days later.
3 Patient seeks medical attention.3
9CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelFluFender performs short-term forecasting based on known clinician and patient behavior
Rx Rx
𝑅𝑥 𝑡 =
𝑖≠𝑘
𝛾
𝛿𝑖𝑏
𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
+𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
𝑅𝑥 𝑡, 𝑖 = Prescriptions at time 𝑡 in area 𝑖𝛾 = effect of close proximity𝜑 = likelihood of transmission
𝛿𝑖 = spatial separation between 𝑖 and 𝑘𝑏 = strength of distance effect
1.01d
1.01d
2 3 4
1
t=0
1-4d
3 4
𝜑
2
𝜑
2
1
1 Patient is infected with the flu virus.
2
2
Patient becomes highly symptomatic 1-4 days later.
3 Patient seeks medical attention.3
4 Patient can be prescribed Tamiflu up to 48 hours following onset of symptoms.
4
10CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelFluFender performs short-term forecasting based on known clinician and patient behavior
Rx Rx
𝑅𝑥 𝑡 =
𝑖≠𝑘
𝛾
𝛿𝑖𝑏
𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
+𝜑
2𝑅𝑥 𝑡 − 2, 𝑖 + 𝑅𝑥 𝑡 − 3, 𝑖
𝑅𝑥 𝑡, 𝑖 = Prescriptions at time 𝑡 in area 𝑖𝛾 = effect of close proximity𝜑 = likelihood of transmission
𝛿𝑖 = spatial separation between 𝑖 and 𝑘𝑏 = strength of distance effect
1.01d
1.01d
2 3 4
1
t=0
1-4d
3 4
𝜑
2
𝜑
2
1
1 Patient is infected with the flu virus.
2
2
Patient becomes highly symptomatic 1-4 days later.
3 Patient seeks medical attention.3
4 Patient can be prescribed Tamiflu up to 48 hours following onset of symptoms.
5 What no one noticed was that the patient was already contagious one day before onset of symptoms.
Therefore, the most likely window for transmission is 2-3 days prior to the day
a Tamiflu prescription is transmitted.
4
5
11CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Why FluFender?Arcadia Analytics FluFender offers a platform for users and healthcare professionals to both track and forecast flu and antiviral requirements
12CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender AnalysisPrepared for SureScripts Technology Challenge
Based on dataset 11/13/2013
13CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
516,032Number of SCRIPT
10.6 Records
14CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
16,033
516,032Number of SCRIPT
10.6 Records
Number of Antiviral Rx
(Tamiflu, Relenza, etc.)
15CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
2
16,033
516,032Number of SCRIPT
10.6 Records
Number of Antiviral Rx
(Tamiflu, Relenza, etc.)
Number of Temporal Clusters
16CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
2
16,033
516,032
15/4
Number of SCRIPT 10.6 Records
Number of Antiviral Rx
(Tamiflu, Relenza, etc.)
Number of Temporal Clusters
Number of Spatial Clusters
17CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
2
16,033
516,032
15/4
Age/Sex
Number of SCRIPT 10.6 Records
Number of Antiviral Rx
(Tamiflu, Relenza, etc.)
Number of Temporal Clusters
Number of Spatial Clusters
Geo-temporally Relevant Flu
Cases
18CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Starting In North Carolina
19CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Moving to South Carolina
20CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Growth and Peak in South Carolina
21CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Peaks in Florida
22CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Outbreak A – FloridaFluFender detected an outbreak with narrow geospatial movement in Orlando, FL with an outbreak radius of 51mi
Demographics Male (50%), Female (50%)
2-15yo - 25% 65+yo -70%
Population Status: Vulnerable Large percentage elderly (65+) and young (2-15)
Outbreak Status : Ongoing Date Range – 12/07/2013 – 12/30/2013
Predictions (3 Day Forecast) Likely continuation: 23 Rx in existing area Likely addition: 10 Rx in new areas (see report for
new Rx)
23CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
What can FluFender tell us?Forecasts antiviral requirements over a short timespan, assisting with potential drug distribution
15 clusters in November
– Upper New England• 274 prescriptions in November
• Almost exclusively 45 years of age and older, mostly in 65+
– Anchorage, Alaska• 231 prescriptions in November
• Almost exclusively ages 2-15, with large percentage in 0-2
4 Clusters in December
– Southeast United States (centered in Florida)• Large percentage elderly (65+) and young (2-15)
24CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
WHY US?
“Dreaming, after all, is a form of planning.” –Gloria Steinem
25CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
APPENDIX
26CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Future Potential for FluFender
• Integrate Flu Vaccine provider locations• User Auto-Location• Local flu outbreak email updates• Add “Did You Feel It?” style survey that
relates directly to health outcomes
• Share flu tracking over social networks• Enable user-embedded FluFender maps
on blogs and websites• Timescale sliders for easier user
experienceUse
r F
ea
ture
sT
ech
no
log
y &
S
ecu
rity
Pu
bli
c H
ea
lth
“Dreaming, after all, is a form of planning.” –Gloria Steinem
• Enhanced cell-blocking on low Rx counts for improved Rx privacy
• Improved predictive geocoding• Improve graph options and layouts• Use of Rx ICD-9 codes to identify
vulnerable populations
• Improved predictive algorithm through additional “training” datasets from previous outbreaks
• Integration with Health Information Exchanges for improved population management
• E-mail prescription alerts to vulnerable populations
• Additional Rx and OTR medications in predictive model
• Alignment with existing public health data sources for improved forecasting
• Standardized public health reporting modules for local authorities and Federal agencies (e.g. HHS, CDC)
27CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
Future Potential for FluFender
• Integrate Flu Vaccine provider locations• User Auto-Location• Local flu outbreak email updates• Add “Did You Feel It?” style survey that
relates directly to health outcomes
• Share flu tracking over social networks• Enable user-embedded FluFender maps
on blogs and websites• Timescale sliders for easier user
experienceUse
r F
ea
ture
sT
ech
no
log
y &
S
ecu
rity
Pu
bli
c H
ea
lth
“Dreaming, after all, is a form of planning.” –Gloria Steinem
• Enhanced cell-blocking on low Rx counts for improved Rx privacy
• Improved predictive geocoding• Improve graph options and layouts• Use of Rx ICD-9 codes to identify
vulnerable populations
• Improved predictive algorithm through additional “training” datasets from previous outbreaks
• Integration with Health Information Exchanges for improved population management
• E-mail prescription alerts to vulnerable populations
• Additional Rx and OTR medications in predictive model
• Alignment with existing public health data sources for improved forecasting
• Standardized public health reporting modules for local authorities and Federal agencies (e.g. HHS, CDC,
28CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Enhancements
Distinguish between palliative and prophylactic use
Distinguish between vulnerable and regular populations
– Determined based on ages and genders
– Increased accuracy if ICD-9 codes in Rx data were reliable
A small, intense outbreak could look the same as a large, relatively mild outbreak
Currently uses a simple predictive model
– Requires larger data sets and additional testing to determine whether the model extends from one outbreak to another, between years, and between populations.
– Need to build interaction between user-generated data and predictions
29CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological Model
Developing the model We considered the factors surrounding an individual getting a Rx for Tamiflu:
– Prophylactic: Vulnerable (0-2yo, 65+yo, pregnant ♀, ↓immune)– Palliative: Intense symptoms, non-complicated (e.g. no comorbidities)– Patient can only get Rx for Tamiflu within 48h of onset of symptoms– Since contagion starts about 1d before symptoms, that 1d represents a window for transmission of high-intensity flu;
because of the FDA regs, that 1 day is almost exactly 3d before the Rx is written– Also, transmission should be most likely very close to existing cases, and for children and high-density populations (that
have greater group contact)
These facts and assumptions directly form the model for occurrence and transmission used here.
Testing the model We tested the three-day conclusion on SureScripts test data
– The outbreak intensity auto-correlated at 2-3 days, supporting the contagion window model– The results rolled off for smaller outbreaks– Geospatial and learning models contributed very little in initial testing, although the number of cases available were
too small to make any strong conclusions.
The model currently primarily uses Rx data for forecasting; other prediction models have been tested but are not as reliable as the Rx data.
Improving the model The following could help improve the forecasting model: More training data; more information about the
prescriptions; more patient information in the prescription (e.g. ICD-9 codes); incorporation of information about populations into the model; reliable historical context
30CONFIDENTIAL and PROPRIETARY. | ©2013 Arcadia Solutions.www.arcadiasolutions.com
FluFender Epidemiological ModelHow a typical meeting starts at Arcadia Healthcare Solutions