Characterization of Physiological States using Machine Learning

PhD position (2024-10 / 2027-09)

Supervision:

• Thesis director: Dr. Laurent Perrinet, Institut de Neurosciences de la Timone (INT)
• Thesis co-director: Thomas Rakotomamonjy (with Sébastien Angelliaume and Nicolas Lantos), Département traitement de l’information et systèmes, ONERA - The French Aerospace Lab - Centre de Salon de Provence

Funding: 3-year contract from ONERA and Région PACA
Doctoral School: ED658 SV - Sciences du Vivant n°658

Project Description

The continuous increase in automation and embedded artificial intelligence in techno-industrial systems makes them both increasingly capable and complex, particularly in the aeronautical field: autopilots, multi-drone supervision systems, etc. The human operator’s role gradually shifts from “low-level” regulation to system supervision. The performance of human-machine systems depends on numerous factors, including:

• The intrinsic efficiency of autonomous components
• The current cognitive state of the operator

There is growing interest in online characterization of operator states such as stress or cognitive fatigue, both during:

• Preliminary system design phases (following operator-centered design approaches)
• Real-time adaptation of system operation to human cognitive states

Methodology

The measurement of brain electrical activity through electroencephalography (EEG) constitutes a widely used tool in neuroengineering, providing insight into cognitive states despite current limitations in operational contexts.

This thesis aims to explore Machine Learning methods for characterizing and recognizing signatures of cognitive states (fatigue, stress, etc.) in aeronautical contexts, based on:

1. The formalism of Riemannian geometry coupled with Machine Learning
2. ONERA’s varied and labeled databases of physiological signals

Research Objectives

The work will focus on developing machine learning classification methodologies to:

• Characterize an individual’s physiological state (inter-subject problem)
• Track the evolution of an individual’s physiological state over time (intra-subject problem)

The ultimate goal is to develop robust metrics to characterize states (or variations) without prior knowledge of initial physiological states, using limited volumes of data characterized by low signal-to-noise ratios.

Keywords

Machine learning, Electroencephalography (EEG), Automated learning, Cognitive states