Adulthood BMI trajectories And Cancer using clustering (ABACus): A consortium approach to test generalisability and develop standardised methodology in unsupervised machine learning approaches

Abstract

Background: Excess body fatness, commonly estimated by ‘one-off’ body mass index (BMI), is associated with increased risk of 13 cancer types. Modelling longitudinal BMI may be more informative for cancer associations and using unsupervised machine learning clustering techniques (e.g. Latent Class Trajectory Modelling - LCTM) may highlight unique groups at risk. Methods: Using 5 large cohort studies (AARP, ARIC, ELSA, PLCO and WHI), longitudinal BMI profiles were obtained for various subsets of the population. I built upon an existing methodological framework to develop gender specific LCTMs. LCTM memberships were then modelled with associated obesity-related cancer (ORC) and non-ORC risk. All developed LCTM models were compared to identify common classes across all the cohorts. Results 1 (Chapter 3): Latent profiles were developed in AARP and PLCO and a 3-class solution was chosen for both cohorts in men and women. The LCTM development process was detailed for the AARP cohort and a 5-class natural spline solution chosen for both sexes. Results 2 (Chapter 4): Different types of data were modelled in an LCTM for specific age recall versus decade recall from the PLCO cohort and recall versus prospectively collected from the WHI cohort. Final shape and number of clusters were drastically affected by the data used. Results 3 (Chapter 5): All trajectories derived from recall data were linked to ORC and non-ORC risk. Heavier trajectories consistently had a higher ORC risk than those who stayed a healthy weight over life course. This corresponded with ORC risk associated with the WHO BMI categories. Results 4 (Chapter6): In addition to prospective data LCTMs, a joint LCTM (longitudinal exposure with time-to-event data included) was developed, and ORC risk was found to be less important as a grouping factor in these models. Results 5 (Chapter 7): Waist circumference was included in the LCTMs for the ARIC and ELSA cohorts. These trajectories were not plausible and did not display associations with ORC risk. Results 6 (Chapter 8): Finally, all derived models were compared to determine whether any patterns were common across all cohorts. In men and women three patterns in recall data were consistent across cohorts. Conclusion: Novel findings in this thesis were latent class variables that appear to incorporate the risk of other lifestyle factors. Notable methodological achievements included: directly comparing how types of data affect the reported LCTM; improvements on the model development process; new methods for generalising trajectory-based patient subgroups.

Date of Award	1 Aug 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Andrew Renehan (Supervisor) & Nophar Geifman (Supervisor)