Medical Times Series Mining
Een diagnose en behandelingsplan wordt meestal samengesteld aan de hand van medische onderzoeken, symptomen, anamneses en zelfrapportage door de patiënt - vastgelegd in het medisch centrum of bij de huisarts. Dergelijke verzamelde gegevens kunnen in de loop van de tijd een veel duidelijker en persoonlijker beeld geven van het ziekteverloop en van de ervaringen van de patiënt. Daarvoor zijn dan wel uitgebreide dataminingtechnieken nodig, om taken te verrichten die nog maar deels opgelost werden door de wetenschappelijke gemeenschap. In deze tweedaagse masterclass leer je hoe die technieken werken, uitgelegd door toponderzoekers en ervaren lesgevers uit het veld!
Praktische info:
Leertraject
Inschrijven?
- Inschrijvingen: tot 20 feb. 2023
- Prijs: Professionals profit sector €300 euro - Professionals non-profit sector €200 - Researchers €110
Opgelet, KU Leuven-onderzoekers moeten 'invoice' kiezen bij betaling (niet betalen via credit card).
Koffiepauzes, broodjeslunches en een afsluitende drink met de experts zijn inbegrepen in de prijs
georganiseerd door:
Two-day Masterclass with an international and varied program, and the possibility to meet and discuss with your peers!
Course will focus on:
Machine learning methods and hands-on exercises using temporal data from the medical domain. Emphasis will be given on both model building as well as on providing trustworthy explanations for the predictions. Data sources to be used include Electronic Health Records, mHealth recordings, and Intensive Care Unit measurements.
Presentation of methods for model induction and model explanation on temporal patient data. Focus will be given on both univariate and multivariate time series data of different granularities, containing gaps, and missing values.
Teachers:
- Prof. Myra Spiliopoulou, Otto-von-Guericke-University Magdeburg
- Prof. Panagiotis Papapetrou, Stockholm University, Sweden
- Phd. Ioanna Miliou, Stockholm University
Program
23 February
KU Leuven ESAT, Kasteelpark Arenberg 10, room 01.57
9h30-12h30:
I. Time series classification and counterfactual explanations, mainly using time series extracted from Electronic Health Records.
12h30 - 13h30: Lunch
13h30 - 16h30:
II. Missingness in time series, inducing predictors and predictive patterns, mainly on time series of mHealth recordings.
24 February
KU Leuven ESAT, Kasteelpark Arenberg 10, room 01.60
9h30-12h30:
III. Time series forecasting methods using time series obtained from the Intensive Care Unit
Hands-on exercises on 2 forecasting methods:
- statistical methods such as AR, MA, ARIMA
- machine learning models, such as MLP and LSTM
12h30 - 14h: Walk to Health House & Lunch
14h00-17h00:
IV. Health House
- Sharing moment: short presentations by young researchers
Sleep stage classification using deep learning - Elisabeth Heremans (KU Leuven) >>>
Analysing pseudotemporal single-cell data - Louise Deconinck (UGent) >>>
Auditory attention decoding: how to predict to whom someone wants to listen from brain signals - Nicolas Heintz (KU Leuven) >>>
A software package for efficient patient trajectory analysis – project ATHENA - Valerie Vandeweerd (Janssen) & dr. Charlotte Herzeel (Imec) >>>
- Thematic Guided Tour "The future of health and care"
- Networking/final drink
Learning outcomes
Participants:
- Become familiar with state-of-the-art methods on time series classification and explore medical time series use-cases
- Become familiar with state-of-the-art methods on time series forecasting and prediction and explore medical time series use-cases
- Apply the learned methods on example time series datasets and critically assess their output and findings
- Become familiar with different ways of modeling high-dimensional timestamped medical data, and understand also how each different way affects the way the learning problem is formulated
- Understand the semantics of missingness in multivariate time series and distinguish between cases where imputation makes sense and cases where imputation may need to be avoided
- Acquire insights on how temporal patterns unfold themselves on multivariate time series.
The future of health and care
This Health House storyline takes you on an innovative and interactive experience through the future of medicine and care. From a 3D movie theatre about the wonderful world of genetics and DNA we move into a Kinect room where we show the applications of 3D printing on a human body, a unique digital anatomy table where we can perform various surgeries and analyse real-life scans on a "digital" body, and a Virtual Reality room where you can dive into the human brain and learn all about epilepsy and how to treat it. Our technology storyline also provides many examples from the region of Flemish Brabant: Zaventem, Spentys, but also from UZ Leuven, KU Leuven and imec.
Sleep stage classification using deep learning
Elisabeth Heremans (KU Leuven)
Sleep stage classification is a fundamental step in sleep assessment to diagnose sleep-wake disturbances. It requires the analysis of 30-second segments of electroencephalography (EEG) data to determine the corresponding sleep stage. In clinical environments, sleep staging is mainly performed manually by medical experts following developed guidelines. The procedure is time-consuming, labor-intensive and prone to human error. In recent years, deep learning methods have managed to achieve close-to-human performances in sleep staging, with the goal of automating this task. However, today, these methods are not yet being adopted in clinical practice. In this talk, we will address the challenges that stand between automated sleep staging methods and clinical practice and discuss methods we have researched to tackle these.
Analysing pseudotemporal single-cell data
Louise Deconinck (UGent)
The analysis of single-cell transcriptomics data often presents a unique challenge as it is not possible to track individual cells over time. Instead of this temporal information, usually only a snapshot of cells within a tissue is available. In order to gain insight into the development of biological processes, trajectory inference methods were developed to order these cells along a pseudotemporal axis representing a continuous developmental process. In this talk, I will discuss the particular challenges associated with using this type of data, the prevalent ways in which this data is analysed and highlight ongoing research efforts in expanding these methods to accommodate multiple samples and more complex biological processes.
Auditory attention decoding: how to predict to whom someone wants to listen from brain signals
Nicolas Heintz (KU Leuven)
Suffering from hearing loss has a large implication on your social life. Even with modern hearing aids, it is still very difficult for a hearing impaired person to take part in a conversation when a lot of other people are talking in the background. Tragically, these situations also tend to be when communication is the most key; think of family gatherings, receptions, meeting with friends... Hearing aids fail in such scenarios because they don't know to whom the user wishes to listen and thus don't know which speaker to enhance. Instead, they become unstable, enhance the wrong speaker and eventually, most users are forced to switch off their hearing aid, effectively isolating them from the conversation. A promising solution to this problem is to decode from from detected brain activity to which speaker the user wishes to listen, called auditory attention decoding (AAD). The idea is to measure the electrical activity of specific regions of the brain using miniature concealed devices, and then use this data to predict who the attended speaker is using several signal processing techniques. In this talk, I will explain the core ideas behind AAD, show how some key signal processing techniques can be used and provide an intuitive understanding why brain decoding is possible.
A software package for efficient patient trajectory analysis – project ATHENA
Valerie Vandeweerd (Janssen) & Charlotte Herzeel (Imec)
We present PTRA, a software package for explorative analysis of disease development.
PTRA provides the tools for extracting statistically relevant trajectories from the medical event histories of a patient population. These trajectories can additionally be clustered for visual inspection and identifying key events in disease development. The algorithms of PTRA are based on a statistical method developed previously by Jensen et al, but we contribute several modifications and extensions to implement a practical tool. This includes a new clustering strategy, filter mechanisms for controlling analysis to specific cohorts and for controlling trajectory output, a parallel implementation that executes on a single server rather than an HPC cluster, etc. We illustrate our tool by discussing trajectories extracted from the TriNetX database for analyzing bladder cancer development.
Cancellation policy
Cancellation free of charge (10% administrative cost) will be possible till February 13th included. After that date cancellation free of charge will only be possible with a valid reason. If no valid reason presented no reimboursement will apply.