Advanced statistics analysis on Finnish Maternity Cohort and Swedish Cervical Screening Cohort (part A). Public Deliverable 6.5

The research question is whether we can use detailed cervical screening
data to predict the risk for cervical cancer and the risk for high-grade
lesions. The broad analytical approach is an adaptation of the screening
models developed by Day and Walter during the 1980s. To our knowledge,
our approach is novel for cervical cancer screening.

We provide (a) SQL and
SAS code for the data extraction from the Swedish Cervical Cancer
Screening Register and (b) R and C++ code for the likelihood construction,
optimisation, and predictions. As a proof of concept, we fit the model to the
cohort of women born in 1960 who were living in Sweden on their fifteenth
birthday. We found some issues with fitting the model due to the lack of
identifiability of the model parameters. We propose some extensions to the
current approach, including scaling up the computations to more birth
cohorts and extending the model to include negative biopsies.

Finally, we
conclude that such a mathematical approach could be compared with
predictions based on machine learning and evaluate whether the machine
learning algorithm gives sufficient weight to different screening histories.