Data & Systems Intelligence Meetups are monthly gatherings of LASIGE members with interests in Data Science, Data Mining, Machine Learning, Artificial intelligence and related topics.

Title: Towards a reliable prediction of conversion from Mild Cognitive Impairment to Alzheimer’s Disease

Abstract: Machine learning is at the core of significant advances in the medical and
healthcare domain. In the particular case of Alzheimer’s disease (AD), researchers have
sought for robust supervised learning models to predict whether a patient with Mild
Cognitive Impairment (MCI) is likely to convert to dementia in the future. These prognostic
models may then be used to guide clinical decisions in real-world situations concerning
patients’ treatment, participation in cognitive rehabilitation programs, and selection for
clinical trials with novel drugs. Despite the efforts, the clinical application of such models
has been hampered by 1) not knowing the time to conversion, 2) the black-box nature of
prognostic models which limits clinicians’ understanding of the emergent outcome, and 3)
the lack of a reliable assessment of the uncertainty of predictions at a patient-based level. In
this presentation, we will talk about some approaches to address these issues, with a
particular focus on the latter. In particular, we will discuss the effectiveness of different
methods to target uncertainty of predictions at a patient-based level, using commonly used
classifiers. We tested two well-known calibration methods (Platt Scaling and Isotonic
Regression), confidence-based predictors (Conformal Predictors), and probabilistic
predictors (Venn-ABERS predictors). The best combination of methods and classifiers to
assess the uncertainty of predictions were then combined in an ensemble-based approach,
with the aim of quantitatively and qualitatively improve individual predictions.

Presenter: Telma Pereira (LASIGE)