Using large language models for safety-related table summarization in clinical study reports.

Objectives: The generation of structured documents for clinical trials is a promising application of large language models (LLMs). We share opportunities, insights, and challenges from a competitive challenge that used LLMs for automating clinical trial documentation. Materials and Methods: As part of a challenge initiated by Pfizer (organizer), several teams (participant) created a pilot for generating summaries of safety tables for clinical study reports (CSRs). Our evaluation framework used automated metrics and expert reviews to assess the quality of AI-generated documents. Results: The comparative analysis revealed differences in performance across solutions, particularly in factual accuracy and lean writing. Most participants employed prompt engineering with generative pre-trained transformer (GPT) models. Discussion: We discuss areas for improvement, including better ingestion of tables, addition of context and fine-tuning. Conclusion: The challenge results demonstrate the potential of LLMs in automating table summarization in CSRs while also revealing the importance of human involvement and continued research to optimize this technology.

Close-range vocal interaction in the common marmoset (Callithrix jacchus).

Vocal communication in animals often involves taking turns vocalizing. In humans, turn-taking is a fundamental rule in conversation. Among non-human primates, the common marmoset is known to engage in antiphonal calling using phee calls and trill calls. Calls of the trill type are the most common, yet difficult to study, because they are not very loud and uttered in conditions when animals are in close proximity to one another. Here we recorded trill calls in captive pair-housed marmosets using wearable microphones, while the animals were together with their partner or separated, but within trill call range. Trills were exchanged mainly with the partner and not with other animals in the room. Animals placed outside the home cage increased their trill call rate and uttered more trills in response to their partner compared to strangers. The fundamental frequency, F0, of trills increased when animals were placed outside the cage. Our results indicate that trill calls can be monitored using wearable audio equipment and that minor changes in social context affect trill call interactions and spectral properties of trill calls.

Atypical behaviour and connectivity in SHANK3-mutant macaques.

Mutation or disruption of the SH3 and ankyrin repeat domains 3 (SHANK3) gene represents a highly penetrant, monogenic risk factor for autism spectrum disorder, and is a cause of Phelan-McDermid syndrome. Recent advances in gene editing have enabled the creation of genetically engineered non-human-primate models, which might better approximate the behavioural and neural phenotypes of autism spectrum disorder than do rodent models, and may lead to more effective treatments. Here we report CRISPR-Cas9-mediated generation of germline-transmissible mutations of SHANK3 in cynomolgus macaques (Macaca fascicularis) and their F1 offspring. Genotyping of somatic cells as well as brain biopsies confirmed mutations in the SHANK3 gene and reduced levels of SHANK3 protein in these macaques. Analysis of data from functional magnetic resonance imaging revealed altered local and global connectivity patterns that were indicative of circuit abnormalities. The founder mutants exhibited sleep disturbances, motor deficits and increased repetitive behaviours, as well as social and learning impairments. Together, these results parallel some aspects of the dysfunctions in the SHANK3 gene and circuits, as well as the behavioural phenotypes, that characterize autism spectrum disorder and Phelan-McDermid syndrome.

Deep convolutional network for animal sound classification and source attribution using dual audio recordings.

This paper introduces an end-to-end feedforward convolutional neural network that is able to reliably classify the source and type of animal calls in a noisy environment using two streams of audio data after being trained on a dataset of modest size and imperfect labels. The data consists of audio recordings from captive marmoset monkeys housed in pairs, with several other cages nearby. The network in this paper can classify both the call type and which animal made it with a single pass through a single network using raw spectrogram images as input. The network vastly increases data analysis capacity for researchers interested in studying marmoset vocalizations, and allows data collection in the home cage, in group housed animals.

Opportunities and challenges in modeling human brain disorders in transgenic primates.

Molecular genetic tools have had a profound impact on neuroscience, but until recently their application has largely been confined to a few model species, most notably mouse, zebrafish, Drosophila melanogaster and Caenorhabditis elegans. With the development of new genome engineering technologies such as CRISPR, it is becoming increasingly feasible to apply these molecular tools in a wider range of species, including nonhuman primates. This will lead to many opportunities for brain research, but it will also pose challenges. Here we identify some of these opportunities and challenges in light of recent and foreseeable technological advances and offer some suggestions. Our main focus is on the creation of new primate disease models for understanding the pathological mechanisms of brain disorders and for developing new approaches to effective treatment. However, we also emphasize that primate genetic models have great potential to address many fundamental questions about brain function, providing an essential foundation for future progress in disease research.

Effect of distracting faces on visual selective attention in the monkey.

In primates, visual stimuli with social and emotional content tend to attract attention. Attention might be captured through rapid, automatic, subcortical processing or guided by slower, more voluntary cortical processing. Here we examined whether irrelevant faces with varied emotional expressions interfere with a covert attention task in macaque monkeys. In the task, the monkeys monitored a target grating in the periphery for a subtle color change while ignoring distracters that included faces appearing elsewhere on the screen. The onset time of distracter faces before the target change, as well as their spatial proximity to the target, was varied from trial to trial. The presence of faces, especially faces with emotional expressions interfered with the task, indicating a competition for attentional resources between the task and the face stimuli. However, this interference was significant only when faces were presented for greater than 200 ms. Emotional faces also affected saccade velocity and reduced pupillary reflex. Our results indicate that the attraction of attention by emotional faces in the monkey takes a considerable amount of processing time, possibly involving cortical-subcortical interactions. Intranasal application of the hormone oxytocin ameliorated the interfering effects of faces. Together these results provide evidence for slow modulation of attention by emotional distracters, which likely involves oxytocinergic brain circuits.

Laminar differences in gamma and alpha coherence in the ventral stream.

Attention to a stimulus enhances both neuronal responses and gamma frequency synchrony in visual area V4, both of which should increase the impact of attended information on downstream neurons. To determine whether gamma synchrony is common throughout the ventral stream, we recorded from neurons in the superficial and deep layers of V1, V2, and V4 in two rhesus monkeys. We found an unexpected striking difference in gamma synchrony in the superficial vs. deep layers. In all three areas, spike-field coherence in the gamma (40-60 Hz) frequency range was largely confined to the superficial layers, whereas the deep layers showed maximal coherence at low frequencies (6-16 Hz), which included the alpha range. In the superficial layers of V2 and V4, gamma synchrony was enhanced by attention, whereas in the deep layers, alpha synchrony was reduced by attention. Unlike these major differences in synchrony, attentional effects on firing rates and noise correlation did not differ substantially between the superficial and deep layers. The results suggest that synchrony plays very different roles in feedback and feedforward projections.

A backward progression of attentional effects in the ventral stream.

The visual processing of behaviorally relevant stimuli is enhanced through top-down attentional feedback. One possibility is that feedback targets early visual areas first and the attentional enhancement builds up at progressively later stages of the visual hierarchy. An alternative possibility is that the feedback targets the higher-order areas first and the attentional effects are communicated "backward" to early visual areas. Here, we compared the magnitude and latency of attentional enhancement of firing rates in V1, V2, and V4 in the same animals performing the same task. We found a reverse order of attentional effects, such that attentional enhancement was larger and earlier in V4 and smaller and later in V1, with intermediate results in V2. These results suggest that attentional mechanisms operate via feedback from higher-order areas to lower-order ones.

Relationship between change detection and pre-change [corrected] activity in visual area V1.

Humans are poor at detecting changes to visual scenes occurring during brief disruptions. It is unclear whether this 'change blindness' results from failure to process the relevant item before the change, or failure to compare/recall the item after the change. We recorded pre-change multi-unit activity in area V1 of monkeys performing a change detection task. The animals were rewarded for making a saccade to the changing figure. Figure-ground related activity was observed, even when no correct saccade was made. However, for the changing figure, pre-change activity was stronger in correct trials than in incorrect trials. We conclude that small differences in pre-change figure-ground segregation have predictive value in whether the change will be successfully detected.

The role of figure-ground segregation in change blindness.

Partial report methods have shown that a large-capacity representation exists for a few hundred milliseconds after a picture has disappeared. However, change blindness studies indicate that very limited information remains available when a changed version of the image is presented subsequently. What happens to the large-capacity representation? New input after the first image may interfere, but this is likely to depend on the characteristics of the new input. In our first experiment, we show that a display containing homogeneous image elements between changing images does not render the large-capacity representation unavailable. Interference occurs when these new elements define objects. On that basis we introduce a new method to produce change blindness: The second experiment shows that change blindness can be induced by redefining figure and background, without an interval between the displays. The local features (line segments) that defined figures and background were swapped, while the contours of the figures remained where they were. Normally, changes are easily detected when there is no interval. However, our paradigm results in massive change blindness. We propose that in a change blindness experiment, there is a large-capacity representation of the original image when it is followed by a homogeneous interval display, but that change blindness occurs whenever the changed image forces resegregation of figures from the background.

Set size effects in the macaque striate cortex.

Attentive processing is often described as a competition for resources among stimuli by mutual suppression. This is supported by findings that activity in extrastriate cortex is suppressed when several stimuli are presented simultaneously, compared to a single stimulus. In this study, we randomly varied the number of simultaneously presented figures (set size) in an attention-demanding change detection task, while we recorded multiunit activity in striate cortex (V1) in monkeys. After figure-background segregation, activity was suppressed as set size increased. This effect was stronger and started earlier among cells stimulated by the background than those stimulated by the figures themselves. As a consequence, contextual modulation, a correlate of figure-background segregation, increased with set size, approximately 100 msec after its initial generation. The results indicate that suppression of responses under increasing attentional demands differentially affects figure and background responses in area V1.

Large capacity storage of integrated objects before change blindness.

Normal people have a strikingly low ability to detect changes in a visual scene. This has been taken as evidence that the brain represents only a few objects at a time, namely those currently in the focus of attention. In the present study, subjects were asked to detect changes in the orientation of rectangular figures in a textured display across a 1600 ms gray interval. In the first experiment, change detection improved when the location of a possible change was cued during the interval. The cue remained effective during the entire interval, but after the interval, it was ineffective, suggesting that an initially large representation was overwritten by the post-change display. To control for an effect of light intensity during the interval on the decay of the representation, we compared performance with a gray or a white interval screen in a second experiment. We found no difference between these conditions. In the third experiment, attention was occasionally misdirected during the interval by first cueing the wrong figure, before cueing the correct figure. This did not compromise performance compared to a single cue, indicating that when an item is attentionally selected, the representation of yet unchosen items remains available. In the fourth experiment, the cue was shown to be effective when changes in figure size and orientation were randomly mixed. At the time the cue appeared, subjects could not know whether size or orientation would change, therefore these results suggest that the representation contains features in their 'bound' state. Together, these findings indicate that change blindness involves overwriting of a large capacity representation by the post-change display.