Dresden 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 40: Data Analytics, Extreme Events, and Nonlinear Stochastic Systems (joint session DY/SOE)

DY 40.6: Vortrag

Mittwoch, 18. März 2020, 16:45–17:00, ZEU 118

Machine Learning on temperature fluctuations in health and disease — •Jens Karschau, Sona Michlíková, Daniel Kotik, Sebastian Starke, Steffen Löck, and Damian McLeod — OncoRay, HZDR, TU Dresden, Dresden, Germany

Rendering disease diagnoses from measurements is a highly complex task. Clinicians train for many years in order to identify pathological events from patient data. Exemplarily, medical expert knowledge recognises subtle differences between normal and tumor-looking features. Today, machine learning (ML) allows us to support not only the clinical decision maker during classification; it also has potential to promptly warn self-monitored individuals. We developed an RNN model that learns on time series temperature data of up to 120 days to detect cancer features in mice. It successfully bins particular days into either tumor vs. no-tumor days. Using out-of-sample data from the same or a different cohort, the model successfully classifies with an accuracy and AUC of up to 0.80. The dynamic time warping dissimilarity measure applied to different days indicates that oscillation patterns contain distinctive features that the RNN model learns. We hypothesise that the model learns features based on oscillatory behaviour at the  150 min time scale: the so-called ’ultradian’ rhythm.
The double benefit from our method is: (a) it uses non-invasive measurements to classify the disease state and (b) it could be deployed for applications in future on-line monitoring of data from wearable devices. Our next efforts are testing human data to deliver actionable insights in disease control and decision support.