Comparing the performance of statistical, machine learning, and deep learning algorithms to predict time-to-event: A simulation study for conversion to mild cognitive impairment.

Mild Cognitive Impairment (MCI) is a condition characterized by a decline in cognitive abilities, specifically in memory, language, and attention, that is beyond what is expected due to normal aging. Detection of MCI is crucial for providing appropriate interventions and slowing down the progression of dementia. There are several automated predictive algorithms for prediction using time-to-event data, but it is not clear which is best to predict the time to conversion to MCI. There is also confusion if algorithms with fewer training weights are less accurate. We compared three algorithms, from smaller to large numbers of training weights: a statistical predictive model (Cox proportional hazards model, CoxPH), a machine learning model (Random Survival Forest, RSF), and a deep learning model (DeepSurv). To compare the algorithms under different scenarios, we created a simulated dataset based on the Alzheimer NACC dataset. We found that the CoxPH model was among the best-performing models, in all simulated scenarios. In a larger sample size (n = 6,000), the deep learning algorithm (DeepSurv) exhibited comparable accuracy (73.1%) to the CoxPH model (73%). In the past, ignoring heterogeneity in the CoxPH model led to the conclusion that deep learning methods are superior. We found that when using the CoxPH model with heterogeneity, its accuracy is comparable to that of DeepSurv and RSF. Furthermore, when unobserved heterogeneity is present, such as missing features in the training, all three models showed a similar drop in accuracy. This simulation study suggests that in some applications an algorithm with a smaller number of training weights is not disadvantaged in terms of accuracy. Since algorithms with fewer weights are inherently easier to explain, this study can help artificial intelligence research develop a principled approach to comparing statistical, machine learning, and deep learning algorithms for time-to-event predictions.

Removing the cardiac field artifact from the EEG using neural network regression.

When EEG recordings are used to reveal interactions between central-nervous and cardiovascular processes, the cardiac field artifact (CFA) poses a major challenge. Because the electric field generated by cardiac activity is also captured by scalp electrodes, the CFA arises as a heavy contaminant whenever EEG data are analyzed time-locked to cardio-electric events. A typical example is measuring stimulus-evoked potentials elicited at different phases of the cardiac cycle. Here, we present a nonlinear regression method deploying neural networks that allows to remove the CFA from the EEG signal in such scenarios. We train neural network models to predict R-peak centered EEG episodes based on the ECG and additional CFA-related information. In a second step, these trained models are used to predict and consequently remove the CFA in EEG episodes containing visual stimulation occurring time-locked to the ECG. We show that removing these predictions from the signal effectively removes the CFA without affecting the intertrial phase coherence of stimulus-evoked activity. In addition, we provide the results of an extensive grid search suggesting a set of appropriate model hyperparameters. The proposed method offers a replicable way of removing the CFA on the single-trial level, without affecting stimulus-related variance occurring time-locked to cardiac events. Disentangling the cardiac field artifact (CFA) from the EEG signal is a major challenge when investigating the neurocognitive impact of cardioafferent traffic by means of the EEG. When stimuli are presented time-locked to the cardiac cycle, both sources of variance are systematically confounded. Here, we propose a regression-based approach deploying neural network models to remove the CFA from the EEG. This approach effectively removes the CFA on a single-trial level and is purely data-driven, providing replicable results.

Mental chronometry in big noisy data.

Temporal measures (latencies) in the event-related potentials of the EEG (ERPs) are a valuable tool for estimating the timing of mental processes, one which takes full advantage of the high temporal resolution of the EEG. Especially in larger scale studies using a multitude of individual EEG-based tasks, the quality of latency measures often suffers from high and low frequency noise residuals due to the resulting low trial counts (because of compressed tasks) and because of the limited feasibility of visual inspection of the large-scale data. In the present study, we systematically evaluated two different approaches to latency estimation (peak latencies and fractional area latencies) with respect to their data quality and the application of noise reduction by jackknifing methods. Additionally, we tested the recently introduced method of Standardized Measurement Error (SME) to prune the dataset. We demonstrate that fractional area latency in pruned and jackknifed data may amplify within-subjects effect sizes dramatically in the analyzed data set. Between-subjects effects were less affected by the applied procedures, but remained stable regardless of procedure.

Decoding of cognitive processes involved in the continuous performance task.

Decoding of electroencephalogram brain representations is a powerful data driven technique to assess the stream of cognitive information processing. It could promote a more thorough understanding of cognitive control networks. For many years, the continuous performance task has been utilized to investigate impaired proactive and reactive cognitive functions. So far, mainly task performance and univariate electroencephalogram were involved in such investigations. In this study, we benefit from multi-variate pattern analysis of continuous performance task variations to provide a more complete spatio-temporal outline of information processing flow involved in sustained and transient attention and response preparation. Besides effects that are well in line with previous EEG research but could be described in more spatial and temporal detail by the used methods, our results could suggest the presence of a higher order feedback control system when expectations are violated. Such a feedback control is related to modulations of behavior both intra- and inter-individually.

Not scene learning, but attentional processing is superior in team sport athletes and action video game players.

We tested if high-level athletes or action video game players have superior context learning skills. Incidental context learning was tested in a spatial contextual cueing paradigm. We found comparable contextual cueing of visual search in repeated displays in high-level amateur handball players, dedicated action video game players and normal controls. In contrast, both handball players and action video game players showed faster search than controls, measured as search time per display item, independent of display repetition. Thus, our data do not indicate superior context learning skills in athletes or action video game players. Rather, both groups showed more efficient visual search in abstract displays that were not related to sport-specific situations.

Cross-task perceptual learning of object recognition in simulated retinal implant perception.

The perception gained by retina implants (RI) is limited, which asks for a learning regime to improve patients' visual perception. Here we simulated RI vision and investigated if object recognition in RI patients can be improved and maintained through training. Importantly, we asked if the trained object recognition can be generalized to a new task context, and to new viewpoints of the trained objects. For this purpose, we adopted two training tasks, a labelling task where participants had to choose the correct label out of other distracting labels for the presented object, and a reverse labelling task where participants had to choose the correct object out of other distracting objects to match the presented label. Our results showed that, despite of the task order, recognition performance was improved in both tasks and lasted at least for a week. The improved object recognition, however, can be transferred only from the labelling task to the reverse labelling task but not vice versa. Additionally, the trained object recognition can be transferred to new viewpoints of the trained objects only in the labelling task but not in the reverse labelling task. Training with the labelling task is therefore recommended for RI patients to achieve persistent and flexible visual perception.

Reward modulation of contextual cueing: Repeated context overshadows repeated target location.

Contextual cueing can be enhanced by reward. However, there is a debate if reward is associated with the repeated target-distractor configurations or with the repeated target locations that occur in both repeated and new displays. Based on neuroimaging evidence, we hypothesized that reward becomes associated with the target location only in new displays, but not in repeated displays, where the repeated target location is overshadowed by the more salient repeated target-distractor configuration. To test this hypothesis, we varied the reward value associated with the same target location in repeated and new displays. The results confirmed the overshadowing hypothesis in that search facilitation in repeated target-distractor configurations was modulated by the variable value associated with the target location. This effect was observed mainly in early learning.

Feedback to distal dendrites links fMRI signals to neural receptive fields in a spiking network model of the visual cortex.

The blood oxygenation level-dependent (BOLD) response has been strongly associated with neuronal activity in the brain. However, some neuronal tuning properties are consistently different from the BOLD response. We studied the spatial extent of neural and hemodynamic responses in the primary visual cortex, where the BOLD responses spread and interact over much longer distances than the small receptive fields of individual neurons would predict. Our model shows that a feedforward-feedback loop between V1 and a higher visual area can account for the observed spread of the BOLD response. In particular, anisotropic landing of inputs to compartmental neurons were necessary to account for the BOLD signal spread, while retaining realistic spiking responses. Our work shows that simple dendrites can separate tuning at the synapses and at the action potential output, thus bridging the BOLD signal to the neural receptive fields with high fidelity.

Modeling fMRI signals can provide insights into neural processing in the cerebral cortex.

Every stimulus or task activates multiple areas in the mammalian cortex. These distributed activations can be measured with functional magnetic resonance imaging (fMRI), which has the best spatial resolution among the noninvasive brain imaging methods. Unfortunately, the relationship between the fMRI activations and distributed cortical processing has remained unclear, both because the coupling between neural and fMRI activations has remained poorly understood and because fMRI voxels are too large to directly sense the local neural events. To get an idea of the local processing given the macroscopic data, we need models to simulate the neural activity and to provide output that can be compared with fMRI data. Such models can describe neural mechanisms as mathematical functions between input and output in a specific system, with little correspondence to physiological mechanisms. Alternatively, models can be biomimetic, including biological details with straightforward correspondence to experimental data. After careful balancing between complexity, computational efficiency, and realism, a biomimetic simulation should be able to provide insight into how biological structures or functions contribute to actual data processing as well as to promote theory-driven neuroscience experiments. This review analyzes the requirements for validating system-level computational models with fMRI. In particular, we study mesoscopic biomimetic models, which include a limited set of details from real-life networks and enable system-level simulations of neural mass action. In addition, we discuss how recent developments in neurophysiology and biophysics may significantly advance the modelling of fMRI signals.

Contextual Modulation is Related to Efficiency in a Spiking Network Model of Visual Cortex.

In the visual cortex, stimuli outside the classical receptive field (CRF) modulate the neural firing rate, without driving the neuron by themselves. In the primary visual cortex (V1), such contextual modulation can be parametrized with an area summation function (ASF): increasing stimulus size causes first an increase and then a decrease of firing rate before reaching an asymptote. Earlier work has reported increase of sparseness when CRF stimulation is extended to its surroundings. However, there has been no clear connection between the ASF and network efficiency. Here we aimed to investigate possible link between ASF and network efficiency. In this study, we simulated the responses of a biomimetic spiking neural network model of the visual cortex to a set of natural images. We varied the network parameters, and compared the V1 excitatory neuron spike responses to the corresponding responses predicted from earlier single neuron data from primate visual cortex. The network efficiency was quantified with firing rate (which has direct association to neural energy consumption), entropy per spike and population sparseness. All three measures together provided a clear association between the network efficiency and the ASF. The association was clear when varying the horizontal connectivity within V1, which influenced both the efficiency and the distance to ASF, DAS. Given the limitations of our biophysical model, this association is qualitative, but nevertheless suggests that an ASF-like receptive field structure can cause efficient population response.

Visual interactions conform to pattern decorrelation in multiple cortical areas.

Neural responses to visual stimuli are strongest in the classical receptive field, but they are also modulated by stimuli in a much wider region. In the primary visual cortex, physiological data and models suggest that such contextual modulation is mediated by recurrent interactions between cortical areas. Outside the primary visual cortex, imaging data has shown qualitatively similar interactions. However, whether the mechanisms underlying these effects are similar in different areas has remained unclear. Here, we found that the blood oxygenation level dependent (BOLD) signal spreads over considerable cortical distances in the primary visual cortex, further than the classical receptive field. This indicates that the synaptic activity induced by a given stimulus occurs in a surprisingly extensive network. Correspondingly, we found suppressive and facilitative interactions far from the maximum retinotopic response. Next, we characterized the relationship between contextual modulation and correlation between two spatial activation patterns. Regardless of the functional area or retinotopic eccentricity, higher correlation between the center and surround response patterns was associated with stronger suppressive interaction. In individual voxels, suppressive interaction was predominant when the center and surround stimuli produced BOLD signals with the same sign. Facilitative interaction dominated in the voxels with opposite BOLD signal signs. Our data was in unison with recently published cortical decorrelation model, and was validated against alternative models, separately in different eccentricities and functional areas. Our study provides evidence that spatial interactions among neural populations involve decorrelation of macroscopic neural activation patterns, and suggests that the basic design of the cerebral cortex houses a robust decorrelation mechanism for afferent synaptic input.