Cervical cancer risk prediction

Problem

Every 3 to 5 years women are invited to participate in cervical cancer screening programs. Early detection of cancer development significantly increases the chances of successful treatment. Hence, to further optimise the screening procedure, we want to see if we can predict whether an individual is at high risk of developing cervical cancer based on their screening history. This allows us to intervene faster with individuals that have a higher risk, while reducing the number of examinations for individuals with lower risk.

Solution

For this project, we had access to two extensive datasets provided by Norwegian and Estonian screening programs. Using these datasets, we aim to predict the future development of cervical cancer for individuals by comparing their screening diagnostics with others. As the data is sequential in nature, with multiple observations following each other over large time intervals, we trained various sequential deep learning approaches.

Results

The following results were obtained:

• Trained multiple variants of recurrent neural networks (RNN, LSTM and GRU) on screening history data.

• Investigated the model performance across populations by comparing the prediction results between Norwegian and Estonian screening data.

For more info, link to paper.

Severin Elvatun, Daan Knoors, Mari Nygård, Anneli Uusküla, Andres Võrk, Jan F. Nygård (2024). “Cross-population evaluation of cervical cancer risk prediction algorithms”, International Journal of Medical Informatics.