1. Rafael Meza Department of Epidemiology University of Michigan
Experiences Sharing Data and Models from a Multi-Institutional Cancer Modeling Consortium
Rafael Meza
Department of Epidemiology
University of Michigan 1

2. CISNET Cancer Intervention and Surveillance Modeling Network
NCI-sponsored collaborative consortium of simulation modelers in breast, prostate, colorectal, lung, and esophageal cancers formed in 2000
Funded through a U01 (collaborative agreement) mechanism
Use surveillance, epidemiology, clinical data and simulation modeling to guide public health research and priorities

4. CISNET Lung Group Six lung cancer models: Smoking and lung cancer
Erasmus, Georgetown, MGH, Michigan, Stanford and Yale
Smoking and lung cancer
Reconstruction of smoking histories in the US
Impact of tobacco control on lung cancer outcomes and overall mortality
Lung cancer screening
Extrapolation of NLST/PLCO trial results to the US
Assessment of alternative lung cancer screening strategies

5. Why so many models? The more the merrier?
Comparative modeling approach
“Independent” models but with comparable inputs and outputs
Heterogeneity in model assumptions / data sources
Address model misspecification
When consensus can be reached, it greatly enhances the credibility of modeling results by highlighting their reproducibility

6. Why so many models? Comparative modeling approach – cont.
When results are disparate, it can help to pinpoint areas where our knowledge base is insufficient and further research is needed
Comparative modeling is a powerful tool that would be impossible without a collaborative effort within the scientific community
Cited by the International Society Pharmacoeconomics and Outcomes Research Modeling TF as example of good modeling practices

7. Fred Hutchinson Lung Cancer Model – Model F
Meza, ten Haaf et al, Cancer in press

8. Erasmus-MISCAN Model – Model E

9. MGH-HMS Lung Cancer Policy Model – Model M
Meza, ten Haaf et al, Cancer in press

10. Stanford Natural History Model of Lung Cancer – Model S
Meza, ten Haaf et al, Cancer in press

11. apm bpm m0 m1 apc bpc mpm lIA1 lIA2 lIB lII lIIIA lIIIB IA1 IA2 IB II
Michigan Lung Cancer screening model – Model U
By gender and histology (SC,AC,SQ,ONSCLC)
Normal X
Preinitiated
Pre
malignant
apm
bpm m0 m1
Preclinical
apc
bpc
mpm
Preclinical
lIA1
lIA2
lIB
lII
lIIIA
lIIIB
IA1
IA2 IB II
IIIA
IIIB IV
dIA1
dIB
dII
dIIIA
dIIIB
dIV
dIA2
Clinical
Detection
Meza, ten Haaf et al, Cancer in press

12. Computational / Data / Documentation Issues
6 independent models
Computational implementation
Shared know-how
Sharing inputs
Large datasets, eg, large-scale screening clinical trials
Input simulators, eg, smoking history generator
Producing same outputs – to be compared
Processing outputs from multiple models
Result dissemination
Common documentation – model profiler
Model sharing / model real-time “running”

15. Collaboration tools

16. Shared Input Generator
Smoking History Generator
Microsimulation of individual smoking histories
C++ code
Inputs based on analysis of historical data on smoking patterns – 50 years of annual national cross-sectional surveys

18. Holford et al, AJPM 2014

19. Holford et al, AJPM 2014

20. Holford et al, AJPM 2014

22. CISNET public resources
Historical US rates of initiation, cessation, cigarettes per day by gender, age and birth cohort
Interactive graphs from publications

23. Simulating smoking related health outcomes
Some groups use macro (population) models Others use microsimulations:
Generate smoking histories for (large) N individuals for each birth-cohort from using the SHG
Use individual histories as inputs for lung cancer models
Lung cancer incidence/mortality, overall mortality, impact of screening
Large scale computations
model calibrations / intervention simulation

24. High-Performance Computing Experiences
Some groups run their models on their own PCs
Others in Linux/Unix clusters
Most run in single threads, but at least one or two groups have done some parallelization
The Michigan group has compiled the SHG in the high-performance UM Flux cluster and is working on parallelizing the SHG and its tobacco control policies module
Parallelization in a single processor multi-core workstation

25. McMahon et al, Plos One 2014

26. Sharing tools – publication support

27. Tobacco Policy Tool - Dissemination

29. Conclusions
Examples of computational tools to support collaborative process
Sharing inputs, models, outputs
Modeling limited by computational power
DOE collaboration
Basis for new Center for the Assessment of the Public Health Impact of Tobacco Regulations proposal (U54)