Quetelet seminars by Max Taubert and Robin Vandaele

For whom
Students
When
24-02-2021 from 12:00 to 13:00
Where
webinar
Language
English
Organizer
Jan De Neve
Contact
Jan.DeNeve@UGent.be
Website
https://forms.gle/mKfxrmAiYcNvC2Wg8
Add to my calendar

Quetelet seminars by Max Taubert and Robin Vandaele

The seminars are given by the 2020 Quetelet Prize winners: Max Taubert (UHasselt) and Robin Vandaele (UGent and a MaStat graduate of last year who won the prize for his MaStat thesis).

The seminar takes place on February 24, 12h00-13h00. Attendance is free, but registration is required through this link: https://forms.gle/mKfxrmAiYcNvC2Wg8 You will receive a link to the Zoom webinar session.

Abstracts:

Modeling the time-course and interrelationship of routine laboratory data to improve pharmacokinetic/-dynamic evaluations in intensive care patients (Max Taubert)

Clinical laboratory data plays an important role in the model-based optimisation of pharmacotherapies since it may serve as an easily measurable surrogate of pharmacokinetic and -dynamic parameters. However, laboratory data is often incomplete and measured infrequently. Thus, the aim of this study was to develop a multivariate model that allows to
properly integrate longitudinal laboratory data into future pharmacokinetic and -dynamic evaluations. Ten laboratory parameters were analysed based on a linear mixed effects model, with a spatial correlation structure describing changes in laboratory parameters over time. Bivariate models of pair-wise laboratory parameter combinations were used to describe relationships between parameters. The predictive performance of all models was evaluated in terms of the mean squared prediction error in a validation dataset. Based on 43,802 observations from 1,097 patients, auto-correlations were found to improve predictions for all laboratory parameters, while bivariate models revealed a cross-correlation for 4 and correlated means for 15 out of 45 parameter pairs. In conclusion, a linear mixed effects model with spatial auto-correlation and pair-wise correlated means provided a reasonable description of clinical laboratory data, while computationally demanding cross-correlation structures provided only a marginal benefit.


Topological Data Analysis for Evaluating Cell Trajectory Data Representations (Robin Vandaele)

Cell trajectory inference (CTI) methods aim to infer biological cell differentiation processes from gene or protein expression data. These processes resemble graph-structured models, and are of crucial importance in immunology or cancer research. Unfortunately, the highdimensionality of the data conceals much of its required geometric information, due to various phenomena that fall under the curse of dimensionality. For this reason, CTI methods commonly start from an initial dimensionality reduction, such as PCA, before inferring the final graph-structured model. As a consequence, the quality of this initial ‘representation’ of the data highly influences the quality of the model. In this talk, we present novel methods to evaluate and compare different representations for cell trajectory data independently from CTI methods, using tools from the rising field of Topological Data Analysis.