Detecting shortcut learning for fair medical AI using shortcut testing.

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 is to characterize the (un)fairness of ML models-their tendency to perform differently across subgroups of the population-and to understand its underlying mechanisms. One potential driver of algorithmic unfairness, shortcut learning, arises when ML models base predictions on improper correlations in the training data. Diagnosing this phenomenon is difficult as sensitive attributes may be causally linked with disease. Using multitask learning, we propose a method to directly test for the presence of shortcut learning in clinical ML systems and demonstrate its application to clinical tasks in radiology and dermatology. Finally, our approach reveals instances when shortcutting is not responsible for unfairness, highlighting the need for a holistic approach to fairness mitigation in medical AI.

Multitask prediction of organ dysfunction in the intensive care unit using sequential subnetwork routing.

OBJECTIVE: 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 learning if tasks are not appropriately selected. We introduce a sequential subnetwork routing (SeqSNR) architecture that uses soft parameter sharing to find related tasks and encourage cross-learning between them. MATERIALS AND METHODS: Using the MIMIC-III (Medical Information Mart for Intensive Care-III) dataset, we train deep neural network models to predict the onset of 6 endpoints including specific organ dysfunctions and general clinical outcomes: acute kidney injury, continuous renal replacement therapy, mechanical ventilation, vasoactive medications, mortality, and length of stay. We compare single-task (ST) models with naive multitask and SeqSNR in terms of discriminative performance and label efficiency. RESULTS: SeqSNR showed a modest yet statistically significant performance boost across 4 of 6 tasks compared with ST and naive multitasking. When the size of the training dataset was reduced for a given task (label efficiency), SeqSNR outperformed ST for all cases showing an average area under the precision-recall curve boost of 2.1%, 2.9%, and 2.1% for tasks using 1%, 5%, and 10% of labels, respectively. CONCLUSIONS: The SeqSNR architecture shows superior label efficiency compared with ST and naive multitasking, suggesting utility in scenarios in which endpoint labels are difficult to ascertain.

Fast temporal dynamics and causal relevance of face processing in the human temporal cortex.

We measured the fast temporal dynamics of face processing simultaneously across the human temporal cortex (TC) using intracranial recordings in eight participants. We found sites with selective responses to faces clustered in the ventral TC, which responded increasingly strongly to marine animal, bird, mammal, and human faces. Both face-selective and face-active but non-selective sites showed a posterior to anterior gradient in response time and selectivity. A sparse model focusing on information from the human face-selective sites performed as well as, or better than, anatomically distributed models when discriminating faces from non-faces stimuli. Additionally, we identified the posterior fusiform site (pFUS) as causally the most relevant node for inducing distortion of conscious face processing by direct electrical stimulation. These findings support anatomically discrete but temporally distributed response profiles in the human brain and provide a new common ground for unifying the seemingly contradictory modular and distributed modes of face processing.

Predicting anxiety from wholebrain activity patterns to emotional faces in young adults: a machine learning approach.

BACKGROUND: 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 biological markers associated with dimensions of symptoms and behaviour that could provide targets to guide treatment choice and novel treatment. In response, the research domain criteria (RDoC) (Insel et al., 2010) was developed to advocate a dimensional approach which omits any disease definitions, disorder thresholds, or cut-points for various levels of psychopathology to understanding the pathophysiological processes underlying psychiatry disorders. In the present study we aimed to apply pattern regression analysis to identify brain signatures during dynamic emotional face processing that are predictive of anxiety and depression symptoms in a continuum that ranges from normal to pathological levels, cutting across categorically-defined diagnoses. METHODS: The sample was composed of one-hundred and fifty-four young adults (mean age=21.6 and s.d.=2.0, 103 females) consisting of eighty-two young adults seeking treatment for psychological distress that cut across categorically-defined diagnoses and 72 matched healthy young adults. Participants performed a dynamic face task involving fearful, angry and happy faces (and geometric shapes) while undergoing functional Magnetic Resonance Imaging (fMRI). Pattern regression analyses consisted of Gaussian Process Regression (GPR) implemented in the Pattern Recognition for Neuroimaging toolbox (PRoNTo). Predicted and actual clinical scores were compared using Pearson's correlation coefficient (r) and normalized mean squared error (MSE) to evaluate the models' performance. Permutation test was applied to estimate significance levels. RESULTS: GPR identified patterns of neural activity to dynamic emotional face processing predictive of self-report anxiety in the whole sample, which covered a continuum that ranged from healthy to different levels of distress, including subthreshold to fully-syndromal psychiatric diagnoses. Results were significant using two different cross validation strategies (two-fold: r=0.28 (p-value=0.001), MSE=4.47 (p-value=0.001) and five fold r=0.28 (p-value=0.002), MSE=4.62 (p-value=0.003). The contributions of individual regions to the predictive model were very small, demonstrating that predictions were based on the overall pattern rather than on a small combination of regions. CONCLUSIONS: These findings represent early evidence that neuroimaging techniques may inform clinical assessment of young adults irrespective of diagnoses by allowing accurate and objective quantitative estimation of psychopathology.

Intracranial Electrophysiology Reveals Reproducible Intrinsic Functional Connectivity within Human Brain Networks.

Evidence for intrinsic functional connectivity (FC) within the human brain is largely from neuroimaging studies of hemodynamic activity. Data are lacking from anatomically precise electrophysiological recordings in the most widely studied nodes of human brain networks. Here we used a combination of fMRI and electrocorticography (ECoG) in five human neurosurgical patients with electrodes in the canonical "default" (medial prefrontal and posteromedial cortex), "dorsal attention" (frontal eye fields and superior parietal lobule), and "frontoparietal control" (inferior parietal lobule and dorsolateral prefrontal cortex) networks. In this unique cohort, simultaneous intracranial recordings within these networks were anatomically matched across different individuals. Within each network and for each individual, we found a positive, and reproducible, spatial correlation for FC measures obtained from resting-state fMRI and separately recorded ECoG in the same brains. This relationship was reliably identified for electrophysiological FC based on slow (<1 Hz) fluctuations of high-frequency broadband (70-170 Hz) power, both during wakeful rest and sleep. A similar FC organization was often recovered when using lower-frequency (1-70 Hz) power, but anatomical specificity and consistency were greatest for the high-frequency broadband range. An interfrequency comparison of fluctuations in FC revealed that high and low-frequency ranges often temporally diverged from one another, suggesting that multiple neurophysiological sources may underlie variations in FC. Together, our work offers a generalizable electrophysiological basis for intrinsic FC and its dynamics across individuals, brain networks, and behavioral states.SIGNIFICANCE STATEMENT The study of human brain networks during wakeful "rest", largely with fMRI, is now a major focus in both cognitive and clinical neuroscience. However, little is known about the neurophysiology of these networks and their dynamics. We studied neural activity during wakeful rest and sleep within neurosurgical patients with directly implanted electrodes. We found that network activity patterns showed striking similarities between fMRI and direct recordings in the same brains. With improved resolution of direct recordings, we also found that networks were best characterized with specific activity frequencies and that different frequencies show different profiles of within-network activity over time. Our work clarifies how networks spontaneously organize themselves across individuals, brain networks, and behavioral states.

Embedding Anatomical or Functional Knowledge in Whole-Brain Multiple Kernel Learning Models.

Pattern recognition models have been increasingly applied to neuroimaging data over the last two decades. These applications have ranged from cognitive neuroscience to clinical problems. A common limitation of these approaches is that they do not incorporate previous knowledge about the brain structure and function into the models. Previous knowledge can be embedded into pattern recognition models by imposing a grouping structure based on anatomically or functionally defined brain regions. In this work, we present a novel approach that uses group sparsity to model the whole brain multivariate pattern as a combination of regional patterns. More specifically, we use a sparse version of Multiple Kernel Learning (MKL) to simultaneously learn the contribution of each brain region, previously defined by an atlas, to the decision function. Our application of MKL provides two beneficial features: (1) it can lead to improved overall generalisation performance when the grouping structure imposed by the atlas is consistent with the data; (2) it can identify a subset of relevant brain regions for the predictive model. In order to investigate the effect of the grouping in the proposed MKL approach we compared the results of three different atlases using three different datasets. The method has been implemented in the new version of the open-source Pattern Recognition for Neuroimaging Toolbox (PRoNTo).

Mapping human temporal and parietal neuronal population activity and functional coupling during mathematical cognition.

Brain areas within the lateral parietal cortex (LPC) and ventral temporal cortex (VTC) have been shown to code for abstract quantity representations and for symbolic numerical representations, respectively. To explore the fast dynamics of activity within each region and the interaction between them, we used electrocorticography recordings from 16 neurosurgical subjects implanted with grids of electrodes over these two regions and tracked the activity within and between the regions as subjects performed three different numerical tasks. Although our results reconfirm the presence of math-selective hubs within the VTC and LPC, we report here a remarkable heterogeneity of neural responses within each region at both millimeter and millisecond scales. Moreover, we show that the heterogeneity of response profiles within each hub mirrors the distinct patterns of functional coupling between them. Our results support the existence of multiple bidirectional functional loops operating between discrete populations of neurons within the VTC and LPC during the visual processing of numerals and the performance of arithmetic functions. These findings reveal information about the dynamics of numerical processing in the brain and also provide insight into the fine-grained functional architecture and connectivity within the human brain.

Correlation between resting state fMRI total neuronal activity and PET metabolism in healthy controls and patients with disorders of consciousness.

INTRODUCTION: The mildly invasive 18F-fluorodeoxyglucose positron emission tomography (FDG-PET) is a well-established imaging technique to measure 'resting state' cerebral metabolism. This technique made it possible to assess changes in metabolic activity in clinical applications, such as the study of severe brain injury and disorders of consciousness. OBJECTIVE: We assessed the possibility of creating functional MRI activity maps, which could estimate the relative levels of activity in FDG-PET cerebral metabolic maps. If no metabolic absolute measures can be extracted, our approach may still be of clinical use in centers without access to FDG-PET. It also overcomes the problem of recognizing individual networks of independent component selection in functional magnetic resonance imaging (fMRI) resting state analysis. METHODS: We extracted resting state fMRI functional connectivity maps using independent component analysis and combined only components of neuronal origin. To assess neuronality of components a classification based on support vector machine (SVM) was used. We compared the generated maps with the FDG-PET maps in 16 healthy controls, 11 vegetative state/unresponsive wakefulness syndrome patients and four locked-in patients. RESULTS: The results show a significant similarity with ρ = 0.75 ± 0.05 for healthy controls and ρ = 0.58 ± 0.09 for vegetative state/unresponsive wakefulness syndrome patients between the FDG-PET and the fMRI based maps. FDG-PET, fMRI neuronal maps, and the conjunction analysis show decreases in frontoparietal and medial regions in vegetative patients with respect to controls. Subsequent analysis in locked-in syndrome patients produced also consistent maps with healthy controls. CONCLUSIONS: The constructed resting state fMRI functional connectivity map points toward the possibility for fMRI resting state to estimate relative levels of activity in a metabolic map.

Decoding intracranial EEG data with multiple kernel learning method.

BACKGROUND: Machine learning models have been successfully applied to neuroimaging data to make predictions about behavioral and cognitive states of interest. While these multivariate methods have greatly advanced the field of neuroimaging, their application to electrophysiological data has been less common especially in the analysis of human intracranial electroencephalography (iEEG, also known as electrocorticography or ECoG) data, which contains a rich spectrum of signals recorded from a relatively high number of recording sites. NEW METHOD: In the present work, we introduce a novel approach to determine the contribution of different bandwidths of EEG signal in different recording sites across different experimental conditions using the Multiple Kernel Learning (MKL) method. COMPARISON WITH EXISTING METHOD: To validate and compare the usefulness of our approach, we applied this method to an ECoG dataset that was previously analysed and published with univariate methods. RESULTS: Our findings proved the usefulness of the MKL method in detecting changes in the power of various frequency bands during a given task and selecting automatically the most contributory signal in the most contributory site(s) of recording. CONCLUSIONS: With a single computation, the contribution of each frequency band in each recording site in the estimated multivariate model can be highlighted, which then allows formulation of hypotheses that can be tested a posteriori with univariate methods if needed.

Cross-Modal Decoding of Neural Patterns Associated with Working Memory: Evidence for Attention-Based Accounts of Working Memory.

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 to which common neural patterns characterize retention in verbal WM and visual WM. Verbal WM was assessed via a standard delayed probe recognition task for letter sequences of variable length. Visual WM was assessed via a visual array WM task involving the maintenance of variable amounts of visual information in the focus of attention. We trained a classifier to distinguish neural activation patterns associated with high- and low-visual WM load and tested the ability of this classifier to predict verbal WM load (high-low) from their associated neural activation patterns, and vice versa. We observed significant between-task prediction of load effects during WM maintenance, in posterior parietal and superior frontal regions of the dorsal attention network; in contrast, between-task prediction in sensory processing cortices was restricted to the encoding stage. Furthermore, between-task prediction of load effects was strongest in those participants presenting the highest capacity for the visual WM task. This study provides novel evidence for common, attention-based neural patterns supporting verbal and visual WM.

Biased binomial assessment of cross-validated estimation of classification accuracies illustrated in diagnosis predictions.

Multivariate classification is used in neuroimaging studies to infer brain activation or in medical applications to infer diagnosis. Their results are often assessed through either a binomial or a permutation test. Here, we simulated classification results of generated random data to assess the influence of the cross-validation scheme on the significance of results. Distributions built from classification of random data with cross-validation did not follow the binomial distribution. The binomial test is therefore not adapted. On the contrary, the permutation test was unaffected by the cross-validation scheme. The influence of the cross-validation was further illustrated on real-data from a brain-computer interface experiment in patients with disorders of consciousness and from an fMRI study on patients with Parkinson disease. Three out of 16 patients with disorders of consciousness had significant accuracy on binomial testing, but only one showed significant accuracy using permutation testing. In the fMRI experiment, the mental imagery of gait could discriminate significantly between idiopathic Parkinson's disease patients and healthy subjects according to the permutation test but not according to the binomial test. Hence, binomial testing could lead to biased estimation of significance and false positive or negative results. In our view, permutation testing is thus recommended for clinical application of classification with cross-validation.

Multiclass classification of FDG PET scans for the distinction between Parkinson's disease and atypical parkinsonian syndromes.

Most available pattern recognition methods in neuroimaging address binary classification problems. Here, we used relevance vector machine (RVM) in combination with booststrap resampling ('bagging') for non-hierarchical multiclass classification. The method was tested on 120 cerebral (18)fluorodeoxyglucose (FDG) positron emission tomography (PET) scans performed in patients who exhibited parkinsonian clinical features for 3.5 years on average but that were outside the prevailing perception for Parkinson's disease (PD). A radiological diagnosis of PD was suggested for 30 patients at the time of PET imaging. However, at follow-up several years after PET imaging, 42 of them finally received a clinical diagnosis of PD. The remaining 78 APS patients were diagnosed with multiple system atrophy (MSA, N = 31), progressive supranuclear palsy (PSP, N = 26) and corticobasal syndrome (CBS, N = 21), respectively. With respect to this standard of truth, classification sensitivity, specificity, positive and negative predictive values for PD were 93% 83% 75% and 96%, respectively using binary RVM (PD vs. APS) and 90%, 87%, 79% and 94%, respectively, using multiclass RVM (PD vs. MSA vs. PSP vs. CBS). Multiclass RVM achieved 45%, 55% and 62% classification accuracy for, MSA, PSP and CBS, respectively. Finally, a majority confidence ratio was computed for each scan on the basis of class pairs that were the most frequently assigned by RVM. Altogether, the results suggest that automatic multiclass RVM classification of FDG PET scans achieves adequate performance for the early differentiation between PD and APS on the basis of cerebral FDG uptake patterns when the clinical diagnosis is felt uncertain. This approach cannot be recommended yet as an aid for distinction between the three APS classes under consideration.

Concurrent synaptic and systems memory consolidation during sleep.

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 are processed during sleep by either or both mechanisms, by comparing recollection before and after sleep. We probed whether LTP influences these processes by contrasting two groups of individuals prospectively recruited based on BDNF rs6265 (Val66Met) polymorphism. Between immediate retrieval and delayed testing scheduled after sleep, responses to recollection increased significantly more in Val/Val individuals than in Met carriers in parietal and occipital areas not previously engaged in retrieval, consistent with "systems-level consolidation." Responses also increased differentially between allelic groups in regions already activated before sleep but only in proportion to slow oscillation power, in keeping with "synaptic downscaling." Episodic memories seem processed at both synaptic and systemic levels during sleep by mechanisms involving LTP.

Decoding semi-constrained brain activity from FMRI using support vector machines and gaussian processes.

Predicting a particular cognitive state from a specific pattern of fMRI voxel values is still a methodological challenge. Decoding brain activity is usually performed in highly controlled experimental paradigms characterized by a series of distinct states induced by a temporally constrained experimental design. In more realistic conditions, the number, sequence and duration of mental states are unpredictably generated by the individual, resulting in complex and imbalanced fMRI data sets. This study tests the classification of brain activity, acquired on 16 volunteers using fMRI, during mental imagery, a condition in which the number and duration of mental events were not externally imposed but self-generated. To deal with these issues, two classification techniques were considered (Support Vector Machines, SVM, and Gaussian Processes, GP), as well as different feature extraction methods (General Linear Model, GLM and SVM). These techniques were combined in order to identify the procedures leading to the highest accuracy measures. Our results showed that 12 data sets out of 16 could be significantly modeled by either SVM or GP. Model accuracies tended to be related to the degree of imbalance between classes and to task performance of the volunteers. We also conclude that the GP technique tends to be more robust than SVM to model unbalanced data sets.

Experience-dependent induction of hypnagogic images during daytime naps: a combined behavioural and EEG study.

This study characterizes hypnagogic hallucinations reported during a polygraphically recorded 90-min daytime nap following or preceding practice of the computer game Tetris. In the experimental group (N = 16), participants played Tetris in the morning for 2 h during three consecutive days, while in a first control group (N = 13, controlling the effect of experience) participants did not play any game, and in a second control group (N = 14, controlling the effect of anticipation) participants played Tetris after the nap. During afternoon naps, participants were repetitively awakened 15, 45, 75, 120 or 180 s after the onset of S1, and were asked to report their mental content. Reports content was scored by three judges (inter-rater reliability 85%). In the experimental group, 48 out of 485 (10%) sleep-onset reports were Tetris-related. They mostly consisted of images and sounds with very little emotional content. They exactly reproduced Tetris elements or mixed them with other mnemonic components. By contrast, in the first control group, only one report out of 107 was scored as Tetris-related (1%), and in the second control group only three reports out of 112 were scored as Tetris-related (3%; between-groups comparison; P = 0.006). Hypnagogic hallucinations were more consistently induced by experience than by anticipation (P = 0.039), and they were predominantly observed during the transition of wakefulness to sleep. The observed attributes of experience-related hypnagogic hallucinations are consistent with the particular organization of regional brain activity at sleep onset, characterized by high activity in sensory cortices and in the default-mode network.

fMRI artefact rejection and sleep scoring toolbox.

We started writing the "fMRI artefact rejection and sleep scoring toolbox", or "FAST", to process our sleep EEG-fMRI data, that is, the simultaneous recording of electroencephalographic and functional magnetic resonance imaging data acquired while a subject is asleep. FAST tackles three crucial issues typical of this kind of data: (1) data manipulation (viewing, comparing, chunking, etc.) of long continuous M/EEG recordings, (2) rejection of the fMRI-induced artefact in the EEG signal, and (3) manual sleep-scoring of the M/EEG recording. Currently, the toolbox can efficiently deal with these issues via a GUI, SPM8 batching system or hand-written script. The tools developed are, of course, also useful for other EEG applications, for example, involving simultaneous EEG-fMRI acquisition, continuous EEG eye-balling, and manipulation. Even though the toolbox was originally devised for EEG data, it will also gracefully handle MEG data without any problem. "FAST" is developed in Matlab as an add-on toolbox for SPM8 and, therefore, internally uses its SPM8-meeg data format. "FAST" is available for free, under the

Brain functional integration decreases during propofol-induced loss of consciousness.

Consciousness has been related to the amount of integrated information that the brain is able to generate. In this paper, we tested the hypothesis that the loss of consciousness caused by propofol anesthesia is associated with a significant reduction in the capacity of the brain to integrate information. To assess the functional structure of the whole brain, functional integration and partial correlations were computed from fMRI data acquired from 18 healthy volunteers during resting wakefulness and propofol-induced deep sedation. Total integration was significantly reduced from wakefulness to deep sedation in the whole brain as well as within and between its constituent networks (or systems). Integration was systematically reduced within each system (i.e., brain or networks), as well as between networks. However, the ventral attentional network maintained interactions with most other networks during deep sedation. Partial correlations further suggested that functional connectivity was particularly affected between parietal areas and frontal or temporal regions during deep sedation. Our findings suggest that the breakdown in brain integration is the neural correlate of the loss of consciousness induced by propofol. They stress the important role played by parietal and frontal areas in the generation of consciousness.