Publications

Machine and Deep Learning Models for Hypoxemia Severity Triage in CBRNE Emergencies
Santino Nanini
Mariem Abid
Yassir Mamouni
Arnaud Wiedemann
Philippe Jouvet
Stéphane Bourassa
BACKGROUND/OBJECTIVES: This study develops machine learning (ML) models to predict hypoxemia severity during emergency triage, particularly … (see more)in Chemical, Biological, Radiological, Nuclear, and Explosive (CBRNE) scenarios, using physiological data from medical-grade sensors. METHODS: Tree-based models (TBMs) such as XGBoost, LightGBM, CatBoost, Random Forests (RFs), Voting Classifier ensembles, and sequential models (LSTM, GRU) were trained on the MIMIC-III and IV datasets. A preprocessing pipeline addressed missing data, class imbalances, and synthetic data flagged with masks. Models were evaluated using a 5-min prediction window with minute-level interpolations for timely interventions. RESULTS: TBMs outperformed sequential models in speed, interpretability, and reliability, making them better suited for real-time decision-making. Feature importance analysis identified six key physiological variables from the enhanced NEWS2+ score and emphasized the value of mask and score features for transparency. Voting Classifier ensembles showed slight metric gains but did not outperform individually optimized models, facing a precision-sensitivity tradeoff and slightly lower F1-scores for key severity levels. CONCLUSIONS: TBMs were effective for real-time hypoxemia prediction, while sequential models, though better at temporal handling, were computationally costly. This study highlights ML's potential to improve triage systems and reduce alarm fatigue, with future plans to incorporate multi-hospital datasets for broader applicability.
2024-12-07
Diagnostics (published)
doi.org
SAFT: Towards Out-of-Distribution Generalization in Fine-Tuning
Bac Nguyen
Stefan Uhlich
Fabien Cardinaux
Lukas Mauch
Marzieh Edraki
Aaron Courville
Handling distribution shifts from training data, known as out-of-distribution (OOD) generalization, poses a significant challenge in the fie… (see more)ld of machine learning. While a pre-trained vision-language model like CLIP has demonstrated remarkable zero-shot performance, further adaptation of the model to downstream tasks leads to undesirable degradation for OOD data. In this work, we introduce Sparse Adaptation for Fine-Tuning (SAFT), a method that prevents fine-tuning from forgetting the general knowledge in the pre-trained model. SAFT only updates a small subset of important parameters whose gradient magnitude is large, while keeping the other parameters frozen. SAFT is straightforward to implement and conceptually simple. Extensive experiments show that with only 0.1% of the model parameters, SAFT can significantly improve the performance of CLIP. It consistently outperforms baseline methods across several benchmarks. On the few-shot learning benchmark of ImageNet and its variants, SAFT gives a gain of 5.15% on average over the conventional fine-tuning method in OOD settings.
2024-12-06
Lecture Notes in Computer Science (published)
doi.org
arxiv.org
The Responsible Foundation Model Development Cheatsheet: A Review of Tools&Resources
Shayne Longpre
Stella Biderman
Alon Albalak
Hailey Schoelkopf
Daniel McDuff
Sayash Kapoor
Kevin Klyman
Kyle Lo
Gabriel Ilharco
Nay San
Maribeth Rauh
Aviya Skowron
Bertie Vidgen
Laura Weidinger
Arvind Narayanan
Victor Sanh
David Ifeoluwa Adelani
Percy Liang
Rishi Bommasani
Peter Henderson … (see 3 more)
Sasha Luccioni
Yacine Jernite
Luca Soldaini
2024-12-06
TMLR (accepted)
doi.org
openreview.net
Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation
Shivalika Singh
Angelika Romanou
Cl'ementine Fourrier
David Ifeoluwa Adelani
Jian Gang Ngui
Daniel Vila-Suero
Peerat Limkonchotiwat
Kelly Marchisio
Wei Qi Leong
Yosephine Susanto
Raymond Ng
Shayne Longpre
Wei-Yin Ko
Madeline Smith
Antoine Bosselut
Alice Oh
André F. T. Martins
Leshem Choshen
Daphne Ippolito
Enzo Ferrante … (see 3 more)
Marzieh Fadaee
Beyza Ermis
Sara Hooker
2024-12-03
ArXiv (preprint)
doi.org
arxiv.org
L-AVATeD: The lidar and visual walking terrain dataset
David Whipps
Patrick Ippersiel
Philippe C. Dixon
Worldwide, a significant number of individuals depend on assistive robotic technologies, such as sophisticated lower-limb prostheses and com… (see more)prehensive exoskeletons to address challenges related to gait and mobility (Laschowski et al. (2020); Hu et al. (2018)). These innovative devices significantly improve the quality of life for people with various disabilities, enabling them to navigate real-world environments more effectively and with greater independence. Despite significant advancements in recent decades, these systems often struggle with prompt and accurate responses to changes in the local environment Hu et al. (2018). Typically, users either control these systems directly or use semi-autonomous modes, frequently needing to manually switch locomotion modes to adapt to different terrains (Laschowski et al. (2022)). This requirement imposes an additional cognitive burden, potentially leading to distraction or injury. Recent research highlights the potential for substantial enhancements in exoskeleton control systems by transitioning away from user-initiated to automated locomotion mode changes (such as shifting from walking on flat surfaces to climbing stairs) to more autonomous control systems. A critical aspect of such systems is their ability to precisely classify the local environment Laschowski et al. (2020)).Conventionally, the study and diagnosis of gait conditions is conducted in laboratory settings, where clinicians can control numerous variables and observe patients on a known walking surface. While this ensures accuracy, it limits the evaluation of gait patterns in varied natural environments and walking terrains Adamczyk et al. (2023).Emerging studies, (e.g. Dixon et al. (2018)) have demonstrated that gait is materially affected by the walking surface. Specifically, walking surfaces have been shown to impact lower-extremity muscle activity (Voloshina and Ferris (2015)), joint kinematics (Dixon et al. (2021)), inter-joint coordination and variability (Ippersiel et al. (2021), Ippersiel et al. (2022)), and joint kinetics (Lay et al. (2006)). Freeing researchers from the constraints imposed by the laboratory requires moving away from measurement systems designed to operate in a fixed environment, towards more portable hardware solutions that can function in ecological contexts. One issue that remains, however, is the ability of portable systems to accurately classify walking surfaces.Portable gait analysis systems with multiple participant-mounted inertial measurement units (IMUs) can themselves be used to classify walking terrain. Shah et al. (2022) used 6 sensors to distinguish 9 surface types in a group of 30 young healthy adults using a machine learning algorithm. It remains unclear however if terrain classification accuracy from IMU data would be affected by patient pathology. That is, algorithms based on IMU data alone may incorrectly assess patient pathology or severity of impairment due to particularities of terrain on which walking is performed. The data IMUs produce is valuable, but data collection outside of the lab could cause them to suffer from exactly the problem they are trying to solve; namely the lack of a known, controlled walking surface. While these mobile systems capture patient gait characteristics, they must also leverage additional sensor capabilities to simultaneously deliver an accurate, ground-truth classification of the local walking terrain.Advancements in Deep Learning, particularly in Convolutional Neural Network (CNN) architectures, have shown remarkable success in image classification Krizhevsky et al. (2012). Various studies (Laschowski et al. (2022); Diaz et al. (2018); Diaz et al. (2018)) have successfully used deep learning and visual data for accurate terrain identification, but these techniques require expansive data sets for their training. Combining multiple data modalities, such as depth sensors (Zhang et al. (2011); Zhang et al. (2019)) and IMUs (Shah et al. (2022)) show promise, though more comprehensive studies (e.g. Zhang et al. (2011)), and publicly available datasets remain scarce. While databases hosting visual images or depth data are available, few if any exist in this domain that provide both modalities of data captured of a single scene (Laschowski et al. (2020)). Further, none could be found that combine these modalities of data with simultaneous measurement of the device orientation sensors.The proliferation of interpreted (stereo camera) or directly measured (LiDAR) depth data into a modern smartphone's sensor capabilities presents a new opportunity for terrain classification. Gyroscopes, magnetometer (compass), and accelerometer sensors are now included in even modest smartphones and allow capture of device orientation and inertia at the same instant as image and depth data. This synergy of visual, depth, and orientation data can potentially enhance environmental recognition accuracy. This paper presents a method to capture a high-resolution, multi-modal dataset in real-time without expensive, professional grade equipment, and an novel dataset that can be used in the aforementioned domains. We expect that providing access to a dataset that hosts multiple modalities of simultaneously captured data will prove invaluable to many groups, from engineers developing high level control systems for exoskeletons, to researchers studying gait. Whether using visual, depth, or other modalities of data, In this data report, we provide a detailed description of the dataset, our hardware and methods for collecting the data, and post-processing steps taken to improve the utility and accessibility of the data. We conclude with suggestions for how other researchers may use this dataset. Analyses of the database for walking terrain classification will be presented in future work.To accurately reflect surfaces common in the built environment inside and surrounding typical North-American academic and healthcare institutions, 9 terrain classes were chosen for this investigation: banked-left, banked-right, irregular, flat-even, grass, sloped-up, sloped-down, stairs-up and stairs-down.While the class names were designed to be as descriptive as possible, some explanation is warranted.The banked-[left,right] labels were applied in cases where the terrain declined significantly, perpendicular to the direction of motion. In cases where another class might also apply (e.g. grass, irregular), these labels were given precedence.The flat-even class was a base-case; indicating any terrain that was generally smooth and solid, and neither sloped, banked nor grassy. Samples may include any material (e.g. concrete, tile) or color (there are many examples exhibiting bright colours and patterns).Irregular surfaces were defined as those that had no slope up/down or left/right and had enough irregularity that they might be expected to materially affect gait. Examples include cobblestone, gravel, and rough mud.Surfaces were labelled as grass if they were generally flat, similar to irregular in that they should be neither sloped nor banked, and consisted primarily of short grasses found on typical found in North American lawns.sloped-[up,down] were defined as any surface (including grassy ones) which had a significant incline or decline in the immediate direction of motion.stairs- [up,down] were the easiest to label, and consisted of stairs of any material, indoors or out.Examples of visual and LiDAR image from each class can be seen in figure 1.Apple's (Apple Inc., Cupertino, USA) iPhone (iOS) was chosen as the platform for collecting our walking terrain dataset due to their built-in LiDAR sensor as well as extensive, well-documented APIs for capturing and manipulating depth data. Data were captured on three physical devices (1 iPhone 12 Pro and 2 iPhone 14 Pros). While it's possible to extract depth data from mobile devices which interpret depth data using stereoscopic methods Wang et al. (2019), we chose devices which contain a built-in LiDAR scanner, for accuracy and consistency.Visual (RGB) images were captured in landscape orientation using the front-facing cameras of each of the iPhone 12 and 14. Data were captured at the native resolution of the sensor: 4032??3024 pixels, RGB, 8-bit per channel and were stored using the Apple High Efficiency Image File Format (HEIF).The specifications for the LiDAR scanner are not available directly from the manufacturer, but Luetzenburg et al. (2021) performed and in-depth analysis of the iPhone 12/14 LiDAR hardware (the devices share the same sensor) for applications in geosciences. Their conclusions coincide with the output LiDAR depth map dimensions we recorded, at 768x576 pixels. In this dataset, each LiDAR "pixel" uses a full 32-bits to store depth data.It should be noted that while differing device models were used to collect samples, both the visual and LiDAR data were identical in spatial and depth resolution. Data were collected over six months during the spring and summer of 2023, by three young healthy adult members of our research lab. Participants were fitted with a chest-mounted mobile-phone harness.The harness allowed for hands free data capture and provided some consistency in data capture across participants. iPhones were clipped to the harness horizontally (i.e. in "landscape orientation") using the built-in mount, with an initial angle between 30-50 degrees downward from the horizon. This orientation provided a wide view of the local walking surface within about a meter in front of the participant and up to roughly 5 meters away, depending on the local terrain. The variation in mounting angle was similar in magnitude to the small up and down variation of camera field of view introduced by simple act of walking. This small amount of noise acts, in effect, as a natural regularizer for the dataset, and will help e.g. a Convolutional Neural Network trained on it to better generalize on unseen data.A custom iOS application was written to simplify simultaneous capture and labelling of visual, LiDAR, and device orientation data at 1Hz. This capture frequency provided a balance between volume of data recorded while preventing too many captures of the same visual scene (i.e., walking terrain). In an individual capture session, participants would select the terrain type (based on their visual interpretation upcoming terrain and the definitions above) in the capture application, tap begin data capture, and terminate capture before the terrain class changed. The data would then be automatically labelled and stored on the device. Data were imported from each device into a central repository and individually reviewed (by D.W.) for labelling errors. All members who participated in data collection were briefed on use of the system prior to data collection. Data were labelled in sequence, with a numeric prefix indicating the order of capture [000-999], and a unique suffix in the form of a universally unique identifier (UUID). Every image pair has the same file name, apart from an additional suffix " depth" on the LiDAR disparity map and differing file extension.The gravity vector data was captured into individual comma-separated-values files, each with the matching UUID suffix. 7,968 RGB/LiDAR image pairs were captured along with the device orientation gravity vector, but due to availability of suitable terrain, class data is imbalanced (Table 1). Class imbalance, while not ideal, can be easily handled using any number of techniques when actually making use of the data. While training a Convolutional Neural Network for example, oversampling of the minority class, adding class weights to the loss function, or using ensemble methods such as bagging and boosting are some common solutions.With the goal of making the dataset more manageable (the raw dataset is almost 45GBs in size), RGB images were pre-processed by reducing their size from 4032x3024 pixels to 512x384 pixels. Resizing used Apple's Scriptable Image Processing System (SIPS) and simultaneously converted images to a more standard JPEG format from their native HEIF format.Processing the LiDAR data required special care, as it is captured in a format not readily consumed by typical image libraries. LiDAR data were captured and stored natively as 32-bit disparity maps (disparity = 1 / distance) and saved to TIFF files. These files were down sampled and normalized using OpenCV Itseez (2015), converted to 8-bit grayscale, and exported to JPEG. At 768x576 pixels, the spatial resolution of the built-in LiDAR scanner is much lower than the visual camera, and so these images were not pre-scaled.A device orientation vector was also recorded at the instant both the visual and LiDAR data were captured. This orientation vector is recorded relative to gravity and the iPhone device axes. It is normalized in each direction in units of the accepted acceleration due to gravity at Earth's surface, i.e. 9.8m/s 2 . These vectors are difficult to use in their raw (x, y, z) vector format, and were translated to more understandable pitch and roll values using the 2-element arc-tangent function:pitch = atan2(z, x)(1)and,roll = atan2(z, y)(2)where pitch is defined as the angle down from the horizon, and roll as deviation either clockwise or anti-clockwise from the horizon line itself.A comprehensive understanding of the local walking context is crucial both for human gait analysis and the advanced control systems of robotic prostheses. Relying solely on laboratory analysis limits researchers' capacity to accurately evaluate clinical gait in patients and hinders the optimal functioning of robotic prostheses in varied ecological contexts. Consequently, conducting studies across a spectrum of walking terrains is essential to address these limitations. A ground-truth understanding of walking terrain is traditionally identified using simple visual data and Deep Neural Networks, in particular ConvolutionalNeural Networks (CNNs). While CNNs are ideal for this classification task, they require extensive training data which may not be readily available.Multiple studies ((Ophoff et al., 2019), (Melotti et al., 2018), (AlDahoul et al., 2021)) have revealed that classifiers trained using multi-modal data (and in particular a combination of depth and visual data) can outperform simple visual classification in object-detection and image classification (Schwarz et al., 2015). Previous datasets in the context of walking terrain have provided only a single modality of data L-AVATeD is notable for being the only open-source database of its kind to provide not just two, but three modalities of data captured simultaneously for a given scene of walking terrain. While not as large as the largest available walking terrain datasets reviewed by Laschowski et al. (2020), at almost 8,000 samples L-AVATeD matches the median dataset size. Further, the spatial resolution of the published RGB images in L-AVATeD (at 4032x3024 pixels) is more than 9 times higher than the highest resolution presented. Images at resolutions of this magnitude may prove unwieldy for most neural networks, especially the low-resource-optimized architectures available for use in mobile and edge-computation hardware. It remains important however to preserve as much signal as possible to not limit future research, and so full-resolution images are provided.Depth data, captured via LiDAR were in fact saved as disparity maps (i.e. 1/distance). Notably, signal decreases as distance increases. At 768x576 "pixels" of native spatial resolution, these too are the most information-dense depth measurements in any of the environment recognition systems reviewed in Laschowski et al. (2020). To ensure we did not introduce any algorithmic bias to the depth data, the capture program was instructed to replace missing signal (depth "pixels") with a placeholder N aN value rather than automatically interpolating such points. This allows researchers to handle the missing data however they see fit, for example: replacing with zeros, average depth, or an interpolation scheme. The raw depth maps are available as TIFF files. To make the depth data easily accessible, a copy is provided created using a very simple algorithm of replacing missing data with zeros, normalizing, and then compressing the depth dimension into 8 bits before saving the file as a JPEG. This format provides a representation suitable for use in most deep neural network software packages. It should be noted however that this simple algorithm is effectively a lossy compression of the data, and while useful, should not be considered as the most information-rich representation. Examples of these lossy depth-data representations can be seen in Figure 1 (bottom row).Device orientation data were captured as a gravity vector. The raw data are available as comma-separated values, indexed with the sample UUID. This raw format is not easily interpreted by humans, and so each was converted to a more easily understood pitch (camera angle up and down relative to the horizon) and roll (camera rotation clockwise and anti-clockwise relative to the horizon). Per-class pitch and roll statistics are provided in Table 1. Preliminary post-hoc analyses of these statistics do not immediately reveal significant signal, but used in combination with the RGB and LiDAR data counterparts, in a deep learning context in particular could prove fruitful. For example, researchers might use this data to "de-rotate" an image using the inverse device rotation angle before passing it into the classifier, potentially improving classifications where horizon angle of the scene may be important.
2024-12-03
Frontiers in Robotics and AI (published)
doi.org
Beta cells are essential drivers of pancreatic ductal adenocarcinoma development
Cathy C. Garcia
Aarthi Venkat
Daniel C. McQuaid
Sherry Agabiti
Alex Tong
Rebecca L. Cardone
Rebecca Starble
Akin Sogunro
Jeremy B. Jacox
Christian F. Ruiz
Richard G. Kibbey
Smita Krishnaswamy
Mandar Deepak Muzumdar
Pancreatic endocrine-exocrine crosstalk plays a key role in normal physiology and disease. For instance, endocrine islet beta (β) cell secr… (see more)etion of insulin or cholecystokinin (CCK) promotes progression of pancreatic adenocarcinoma (PDAC), an exocrine cell-derived tumor. However, the cellular and molecular mechanisms that govern endocrine-exocrine signaling in tumorigenesis remain incompletely understood. We find that β cell ablation impedes PDAC development in mice, arguing that the endocrine pancreas is critical for exocrine tumorigenesis. Conversely, obesity induces β cell hormone dysregulation, alters CCK-dependent peri-islet exocrine cell transcriptional states, and enhances islet proximal tumor formation. Single-cell RNA-sequencing, in silico latent-space archetypal and trajectory analysis, and genetic lineage tracing in vivo reveal that obesity stimulates postnatal immature β cell expansion and adaptation towards a pro-tumorigenic CCK+ state via JNK/cJun stress-responsive signaling. These results define endocrine-exocrine signaling as a driver of PDAC development and uncover new avenues to target the endocrine pancreas to subvert exocrine tumorigenesis.
2024-12-02
bioRxiv (preprint)
doi.org
MiRGraph: A hybrid deep learning approach to identify microRNA-target interactions by integrating heterogeneous regulatory network and genomic sequences
Pei Liu
Yang Liu
Jiawei Luo
Yuemei Li
2024-12-02
2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (published)
doi.org
Neural Implicit Reduced Fluid Simulation
Yuanyuan Tao
Ivan Puhachov
Derek Nowrouzezahrai
Paul Kry
High-fidelity simulation of fluid dynamics is challenging because of the high dimensional state data needed to capture fine details and the … (see more)large computational cost associated with advancing the system in time. We present neural implicit reduced fluid simulation (NIRFS), a reduced fluid simulation technique that combines an implicit neural representation of fluid shapes and a neural ordinary differential equation to model the dynamics of fluid in the reduced latent space. The latent trajectories are computed at very little cost in comparison to simulations for training, while preserving fine physical details. We show that this approach can work well, capturing the shapes and dynamics involved in a variety of scenarios with constrained initial conditions, e.g., droplet-droplet collisions, crown splashes, and fluid slosh in a container. In each scenario, we learn the latent implicit representation of fluid shapes with a deep-network signed distance function, as well as the energy function and parameters of a damped Hamiltonian system, which helps guarantee desirable properties of the latent dynamics. To ensure that latent shape representations form smooth and physically meaningful trajectories, we simultaneously learn the latent representation and dynamics. We evaluate novel simulations for conservation of volume and momentum conservation, discuss design decisions, and demonstrate an application of our method to fluid control.
2024-12-02
ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (published)
doi.org
Insect Identification in the Wild: The AMI Dataset
Aditya Jain
F. Cunha
M. J. Bunsen
Juan Sebastián Cañas
L. Pasi
N. Pinoy
Flemming Helsing
JoAnne Russo
Marc Botham
Michael Sabourin
Jonathan Fréchette
Alexandre Anctil
Yacksecari Lopez
Eduardo Navarro
Filonila Perez Pimentel
Ana Cecilia Zamora
José Alejandro Ramirez Silva
Jonathan Gagnon
Tom August
K. Bjerge … (see 8 more)
Alba Gomez Segura
Marc Bélisle
Yves Basset
K. P. McFarland
David Roy
Toke Thomas Høye
Maxim Larrivée
David Rolnick
Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems… (see more) and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups.
2024-12-01
Lecture Notes in Computer Science (published)
doi.org
arxiv.org
ProGRes: Prompted Generative Rescoring on ASR n-Best
Ada Defne Tur
Adel Moumen
Mirco Ravanaelli
2024-12-01
2024 IEEE Spoken Language Technology Workshop (SLT) (published)
doi.org
arxiv.org
Combining supervised learning and local search for the multicommodity capacitated fixed-charge network design problem
Charly Robinson La Rocca
Jean-François Cordeau
Emma Frejinger
The multicommodity capacitated fixed-charge network design problem has been extensively studied in the literature due to its wide range of a… (see more)pplications. Despite the fact that many sophisticated solution methods exist today, finding high-quality solutions to large-scale instances remains challenging. In this paper, we explore how a data-driven approach can help improve upon the state of the art. By leveraging machine learning models, we attempt to reveal patterns hidden in the data that might be difficult to capture with traditional optimization methods. For scalability, we propose a prediction method where the machine learning model is called at the level of each arc of the graph. We take advantage of off-the-shelf models trained via supervised learning to predict near-optimal solutions. Our experimental results include an algorithm design analysis that compares various integration strategies of predictions within local search algorithms. We benchmark the ML-based approach with respect to the state-of-the-art heuristic for this problem. The findings indicate that our method can outperform the leading heuristic on sets of instances sampled from a uniform distribution.
2024-11-30
Transportation Research Part E: Logistics and Transportation Review (published)
doi.org
arxiv.org
Decomposing the Brain in Autism: Linking Behavioral Domains to Neuroanatomical Variation and Genomic Underpinnings.
Hanna Seelemeyer
Caroline Gurr
Johanna Leyhausen
Lisa M. Berg
Charlotte M. Pretzsch
Tim Schäfer
Bassem Hermila
Christine M. Freitag
Eva Loth
Beth Oakley
Luke Mason
Jan K. Buitelaar
Christian Beckmann
Dorothea L. Floris
Tony Charman
Tobias Banaschewski
Emily Jones
Thomas Bourgeron
Jumana Ahmad
Sara Ambrosino … (see 58 more)
Bonnie Auyeung
Simon Baron-Cohen
Sarah Baumeister
Sven Bölte
Carsten Bours
Michael Brammer
Daniel Brandeis
Claudia Brogna
Yvette de Bruijn
Bhismadev Chakrabarti
Ineke Cornelissen
Daisy Crawley
Flavio Dell’Acqua
Guillaume Dumas
Sarah Durston
Christine Ecker
Jessica Faulkner
Vincent Frouin
Pilar Garcés
David Goyard
Lindsay Ham
Hannah Hayward
Joerg F. Hipp
Rosemary Holt
Mark Johnson
Emily J. H. Jones
Prantik Kundu
Meng-Chuan Lai
Xavier Liogier D’ardhuy
Michael V. Lombardo
David J. Lythgoe
René Mandl
Andre Marquand
Maarten Mennes
Andreas Meyer-Lindenberg
Carolin Moessnang
Nico Bast
Larry O’Dwyer
Marianne Oldehinkel
Bob Oranje
Gahan Pandina
Antonio Persico
Barbara Ruggeri
Declan G.M. Murphy
Amber N. V. Ruigrok
Jessica Sabet
Roberto Sacco
Antonia San José Cáceres
Emily Simonoff
Will Spooren
Julian Tillmann
Roberto Toro
Heike Tost
Jack Waldman
Steve C. R. Williams
Caroline Wooldridge
Marcel P. Zwiers
Declan Murphy
2024-11-30
Biological Psychiatry: Cognitive Neuroscience and Neuroimaging (published)
doi.org