HeRC seminar: 'Efficient High-Dimensional Disease Outcome Prediction in Heterogeneous
Time: 12:00 - 1pm
Venue: The Congregation, Vaughan House, Health eResearch Centre
Sorry, this event has now ended.
The increasing availability of both detailed clinical information and in some cases, high-dimensional molecular data has led to a renewed interest in precision medicine. One of the hopes is to be able to predict disease outcomes based on a patient's molecular and clinical profile. This requires new statistical and machine learning methodology to build high-dimensional predictive models that can deal with the inherent heterogeneity of many diseases. For example, in the neurodegenerative disease Amyotrophic Lateral Sclerosis (ALS), it is well known that some patients exhibit rapid progression, while others progress very slowly. However, the causes underlying these differences are unknown.
We present a principled method for modelling prediction in heterogeneous settings, based on information sharing across penalised linear regression models. This allows us to build prediction models that reflect the heterogeneity of the population, while at the same time leveraging the commonalities between groups of homogeneous samples. We present two related approaches, using l1 and l2 fusion penalties on the model-specific parameters, and show in extensive simulation studies that these approaches outperform both naive pooling and complete separation of models in realistic scenarios.
We then apply our method to three datasets: 1. gene expression and drug sensitivity data from The Cancer Cell Line Encyclopedia (Barretina et al. 2012), 2. clinical trial data from ProACT, a database of ALS patients, and 3. GWAS mutation data from the ADNI Alzheimer's database. Using these datasets, we demonstrate that out method improves on other popular prediction approaches, as well as being highly computationally efficient. Additionally, we show that our model can be used to identify biomarkers for disease progression, and to identify common features between otherwise disparate population groups.
Booking: All welcome. No need to book, just turn up.