A5007707507- Functional Brain Connectivity Studies
- Neural dynamics and brain function
- EEG and Brain-Computer Interfaces
- Neural Networks and Applications
- Explainable Artificial Intelligence (XAI)
- Face Recognition and Perception
- Advanced Neuroimaging Techniques and Applications
- Machine Learning in Healthcare
- Gaussian Processes and Bayesian Inference
- Neural and Behavioral Psychology Studies
- Advanced MRI Techniques and Applications
- Digital Mental Health Interventions
- Sleep and Wakefulness Research
- Anesthesia and Sedative Agents
- Blind Source Separation Techniques
- Parkinson's Disease Mechanisms and Treatments
- Multisensory perception and integration
- Cell Image Analysis Techniques
- Mobile Health and mHealth Applications
- Time Series Analysis and Forecasting
- Refrigeration and Air Conditioning Technologies
- Telemedicine and Telehealth Implementation
- Neurological disorders and treatments
- Face recognition and analysis
- Visual perception and processing mechanisms
Stanford University
2013-2024
Google (United Kingdom)
2021-2023
DeepMind (United Kingdom)
2023
Google (United States)
2021
University of Liège
2009-2020
University College London
2015-2019
Cognitive Research (United States)
2016
Palo Alto University
2014
Walloon Excellence in Lifesciences and Biotechnology
2013
Cyclotron (Netherlands)
2009-2012
In the past years, mass univariate statistical analyses of neuroimaging data have been complemented by use multivariate pattern analyses, especially based on machine learning models. While these allow an increased sensitivity for detection spatially distributed effects compared to techniques, they lack established and accessible software framework. The goal this work was build a toolbox comprising all necessary functionalities data, “Pattern Recognition Neuroimaging Toolbox” (PRoNTo) is...
ML models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An pipeline is underspecified it can return many predictors with equivalently strong held-out performance the training domain. Underspecification common modern pipelines, such those based on deep learning. Predictors returned by pipelines treated equivalent their domain performance, but we show here that behave very differently...
Evidence for intrinsic functional connectivity (FC) within the human brain is largely from neuroimaging studies of hemodynamic activity. Data are lacking anatomically precise electrophysiological recordings in most widely studied nodes networks. Here we used a combination fMRI and electrocorticography (ECoG) five neurosurgical patients with electrodes canonical “default” (medial prefrontal posteromedial cortex), “dorsal attention” (frontal eye fields superior parietal lobule),...
Multivariate classification is used in neuroimaging studies to infer brain activation or medical applications diagnosis. Their results are often assessed through either a binomial permutation test. Here, we simulated of generated random data assess the influence cross-validation scheme on significance results. Distributions built from with did not follow distribution. The test therefore adapted. On contrary, was unaffected by scheme. further illustrated real-data brain–computer interface...
Machine learning (ML) approaches have demonstrated promising results in a wide range of healthcare applications. Data plays crucial role developing ML-based systems that directly affect people's lives. Many the ethical issues surrounding use ML stem from structural inequalities underlying way we collect, use, and handle data. Developing guidelines to improve documentation practices regarding creation, maintenance datasets is therefore critical importance. In this work, introduce Healthsheet,...
Abstract Machine learning (ML) holds great promise for improving healthcare, but it is critical to ensure that its use will not propagate or amplify health disparities. An important step characterize the (un)fairness of ML models—their tendency perform differently across subgroups population—and understand underlying mechanisms. One potential driver algorithmic unfairness, shortcut learning, arises when models base predictions on improper correlations in training data. Diagnosing this...
Most available pattern recognition methods in neuroimaging address binary classification problems. Here, we used relevance vector machine (RVM) combination with booststrap resampling ('bagging') for non-hierarchical multiclass classification. The method was tested on 120 cerebral 18fluorodeoxyglucose (FDG) positron emission tomography (PET) scans performed patients who exhibited parkinsonian clinical features 3.5 years average but that were outside the prevailing perception Parkinson's...
Recent studies suggest common neural substrates involved in verbal and visual working memory (WM), interpreted as reflecting shared attention-based, short-term retention mechanisms. We used a machine-learning approach to determine more directly the extent which patterns characterize WM WM. Verbal was assessed via standard delayed probe recognition task for letter sequences of variable length. Visual array involving maintenance amounts information focus attention. trained classifier...
Significance Humans have the unique ability to perform exact mental arithmetic, which derives from association of symbols (e.g., “3”) with discrete quantities. Using direct intracranial recordings, we measured electrophysiological activity neuronal populations in lateral parietal cortex (LPC) and ventral temporal (VTC) that are known be important for numerical processing as subjects performed various experiments. We observed functional heterogeneity within each region at millimeter...
Pattern recognition models have been increasingly applied to neuroimaging data over the last two decades. These applications ranged from cognitive neuroscience clinical problems. A common limitation of these approaches is that they do not incorporate previous knowledge about brain structure and function into models. Previous can be embedded pattern by imposing a grouping based on anatomically or functionally defined regions. In this work, we present novel approach uses group sparsity model...
We started writing the “fMRI artefact rejection and sleep scoring toolbox”, or “FA<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi mathvariant="double-struck">S</mml:mi></mml:math>T”, to process our EEG-fMRI data, that is, simultaneous recording of electroencephalographic functional magnetic resonance imaging data acquired while a subject is asleep. FA<mml:math mathvariant="double-struck">S</mml:mi></mml:math>T tackles three crucial issues typical this kind data: (1)...
Abstract Introduction The mildly invasive 18F‐fluorodeoxyglucose positron emission tomography ( FDG ‐ PET ) is a well‐established imaging technique to measure ‘resting state’ cerebral metabolism. This made it possible assess changes in metabolic activity clinical applications, such as the study of severe brain injury and disorders consciousness. Objective We assessed possibility creating functional MRI maps, which could estimate relative levels maps. If no absolute measures can be extracted,...
Abstract We measured the fast temporal dynamics of face processing simultaneously across human cortex (TC) using intracranial recordings in eight participants. found sites with selective responses to faces clustered ventral TC, which responded increasingly strongly marine animal, bird, mammal, and faces. Both face-selective face-active but non-selective showed a posterior anterior gradient response time selectivity. A sparse model focusing on information from performed as well as, or better...
It is becoming increasingly clear that pathophysiological processes underlying psychiatric disorders categories are heterogeneous on many levels, including symptoms, disease course, comorbidity and biological underpinnings. This heterogeneity poses challenges for identifying markers associated with dimensions of symptoms behaviour could provide targets to guide treatment choice novel treatment. In response, the research domain criteria (RDoC) ( Insel et al., 2010 ) was developed advocate a...
Machine learning models have been successfully applied to neuroimaging data make predictions about behavioral and cognitive states of interest. While these multivariate methods greatly advanced the field neuroimaging, their application electrophysiological has less common especially in analysis human intracranial electroencephalography (iEEG, also known as electrocorticography or ECoG) data, which contains a rich spectrum signals recorded from relatively high number recording sites.In...
Diagnosing and mitigating changes in model fairness under distribution shift is an important component of the safe deployment machine learning healthcare settings. Importantly, success any mitigation strategy strongly depends on structure shift. Despite this, there has been little discussion how to empirically assess a that one encountering practice. In this work, we adopt causal framing motivate conditional independence tests as key tool for characterizing shifts. Using our approach two...
Summary This study characterizes hypnagogic hallucinations reported during a polygraphically recorded 90‐min daytime nap following or preceding practice of the computer game Tetris. In experimental group ( N = 16), participants played Tetris in morning for 2 h three consecutive days, while first control 13, controlling effect experience) did not play any game, and second 14, anticipation) after nap. During afternoon naps, were repetitively awakened 15, 45, 75, 120 180 s onset S1, asked to...
Recently, machine learning models have been applied to neuroimaging data, allowing make predictions about a variable of interest based on the pattern activation or anatomy over set voxels. These recognition methods present undeniable assets classical (univariate) techniques, by providing for unseen as well weights each voxel in model. However, obtained weight map cannot be thresholded perform regionally specific inference, leading difficult localization interest. In this work, we provide...
Predicting a particular cognitive state from specific pattern of fMRI voxel values is still methodological challenge. Decoding brain activity usually performed in highly controlled experimental paradigms characterized by series distinct states induced temporally constrained design. In more realistic conditions, the number, sequence and duration mental are unpredictably generated individual, resulting complex imbalanced data sets. This study tests classification activity, acquired on 16...
Memories are consolidated during sleep by two apparently antagonistic processes: (1) reinforcement of memory-specific cortical interactions and (2) homeostatic reduction in synaptic efficiency. Using fMRI, we assessed whether episodic memories processed either or both mechanisms, comparing recollection before after sleep. We probed LTP influences these processes contrasting groups individuals prospectively recruited based on BDNF rs6265 (Val66Met) polymorphism. Between immediate retrieval...
Recurrent Neural Networks (RNNs) are often used for sequential modeling of adverse outcomes in electronic health records (EHRs) due to their ability encode past clinical states. These deep, recurrent architectures have displayed increased performance compared other approaches a number tasks, fueling the interest deploying deep models settings. One key elements ensuring safe model deployment and building user trust is explainability. Testing with Concept Activation Vectors (TCAV) has recently...
Multitask learning (MTL) using electronic health records allows concurrent prediction of multiple endpoints. MTL has shown promise in improving model performance and training efficiency; however, it often suffers from negative transfer - impaired if tasks are not appropriately selected. We introduce a sequential subnetwork routing (SeqSNR) architecture that uses soft parameter sharing to find related encourage cross-learning between them.Using the MIMIC-III (Medical Information Mart for...
Since machine learning models have been applied to neuroimaging data, researchers drawn conclusions from the derived weight maps. In particular, maps of classifiers between two conditions are often described as a proxy for underlying signal differences conditions. Recent studies however suggested that such could not reliably recover source neural signals and even led false positives (FP). this work, we used semi-simulated data ElectroCorticoGraphy (ECoG) investigate how signal-to-noise ratio...