Machine learning for molecular classification
A coauthored University of North Texas poster, shown at the 2014 MCBIOS conference, compares neural networks, support-vector machines, and random forests for identifying cancer types from gene-expression data. It also varies each algorithm’s settings, examining how model choices affect classification in that research task.
The work presents machine learning as a way to analyze patterns in molecular measurements. Microarray data record the activity of very many genes at once, so choosing the right classifier and the right settings is central to finding the signal that separates one cancer type from another, and the poster’s comparison of algorithms addresses exactly that choice.
Data science for behavioral health
Digital-phenotyping studies use behavioral measurements collected through smartphones to model changes in mental-health symptoms over time. A schizophrenia study evaluates prediction, change detection, and counterfactual explanations, which show what would have to differ for a model’s answer to change, in a simulated continuous-monitoring workflow; a separate study examines factors associated with outcomes in smartphone-delivered interventions for depression.
The studies apply statistical and machine-learning methods to longitudinal behavioral and treatment data. Because symptoms and treatment response unfold over weeks and months, these longitudinal methods make it possible to follow change within an individual and to explain predictions in terms clinicians can examine; each publication page describes its methods and findings.
Multimodal health research
The University of Oxford’s Deep and Frequent Phenotyping project, in which Dr. Costilla Reyes took part from 2016 to 2018, examined how PET (positron emission tomography) imaging, structural MRI (magnetic resonance imaging), retinal imaging, electrophysiology, and peripheral molecular markers could support research into preclinical and early Alzheimer’s disease. The project explored how frequent measurement could serve experimental medicine and clinical-trial research.
These examples show how data science connects measurement design with classification, prediction, and evaluation across distinct molecular, behavioral, and multimodal research questions.