Towards a data-driven system for personalized cervical cancer risk stratification

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Cervical cancer screening programs in Nordic countries have reduced incidence and mortality.
• Current one-size-fits-all screening strategies may lead to over-screening in low-risk populations.
• Personalized risk estimation is needed to optimize screening and identify high-grade lesions.

Purpose of the Study:

• To develop and compare machine learning risk estimators for personalized cervical cancer prediction.
• To address challenges of rare disease, imbalanced data, and non-stationary screening habits.
• To treat cervical cancer risk prediction as a longitudinal forecasting problem.

Main Methods:

• Utilized longitudinal screening histories from the Cancer Registry of Norway.
• Extended existing frameworks using incremental learning for longitudinal predictions.
• Adapted imbalanced classification strategies for non-stationary data.

Main Results:

• Initial models showed bias towards normal results due to imbalanced data.
• Adapted strategies improved identification of at-risk individuals.
• Estimated absolute risk curves closely reflected observed risks in hold-out data.

Conclusions:

• Longitudinal machine learning models show promise for cervical cancer risk stratification.
• Personalized screening recommendations can be improved by these advanced prediction models.
• This approach offers a more targeted strategy than current population-level screening.