Machine Learning Core

National Institute Of Mental Health

Investigators

Linked publications, trials & patents

Abstract

------------------- Research activity ------------------- Several of our ongoing collaboration projects led to articles published this year. Below we summarize the goals of the projects where our group had a significant technical contribution, arranged by investigator we were collaborating with: • Investigators Lauren Atlas and Joyce Chung characterized the dynamic interaction of psychiatric vulnerability, loneliness, and isolation during the first year of the COVID-19 pandemic. • Investigators Sofia Beas and Mario Penzo identified the roles of different, parallel thalamo-striatal projections in shaping instrumental actions in response to changing motivational states. • Investigators Ana Inácio and Soohyun Lee characterized the brain-wide network of functionally distinct neurons connecting to a somatosensory area. • Investigators Zoe Laky and Reut Naim characterized how, over time, changes in executive function in youth are associated with externalizing problems. This paper is a pre-registration of the analysis approach we developed with the investigators. • Investigators Tarun Madangopal and Bruce Hope had two related project goals. The first was to determine whether prelimbic cortex contains distinct neural populations, encoding both response execution and inhibition, which are active during rewarded acquisition and extinction of a response. The second was to determined whether they were re-activated during re-instatement of the response, or whether this was driven by a different population. • Investigators Alexander Maitland and Grant Glatfelter evaluated the feasibility of automatically detecting a specific behavioral response -- correlated with drug effects -- from video data at various resolutions and frame rates. • Investigators: Matthias Nau and Chris Baker tested the hypothesis that, when recalling past events, patterns of gaze position and neural activity resemble those observed during the original experience and are linked through a common process that underlies the reinstatement of past experiences during memory retrieval. • Investigators Dipta Saha and Melissa Brotman determined whether ecological momentary assessment data could be used to predict temper outbursts in youths. • Investigators Javier Gonzalez-Castillo and Peter Bandettini, in a collaboration lead by others, were interested in improving paradigm-free hemodynamic deconvolution of functional MRI data, used in identifying the timing of neural events. We also published papers that were the outcome of projects initiated by our group. These aim at developing new methods and technologies, in anticipation of, or in response to, the broad needs of the Intramural Research Program. These projects had the following goals: • Develop text classification models to detect impact, symptoms, and outcomes of depression in adolescents, from their own language. The ultimate goal is to be able to extract information not captured by standard questionnaires, from patient-written narratives. • Develop a new statistical framework for analysis of trial-level temporal dynamics in fiber photometry experiments. While still being extended, this framework is now in use by several groups, in the IRP and outside, and has already led to other publications. • Develop a new statistical framework for doing causal inference on the behavioral effects of optogenetic stimulation in closed loop settings. This framework handles the potential confounds of stimulating as a function of the outcome one wants to promote, and can also estimate many kinds of response: additive, ceiling effects, refractory, etc. • Carry out a comparison of existing sentiment analysis tools and large language models (LLMs) for the purposes of quantifying sentiment -- positive/negative/neutral -- in survey text. This was used for an IRP project analyzing COVID-19 survey data, described in the last report, and determined our switch to AI-based tools. • Develop a methodology to combine text and images in generating summaries of biomedical literature. • Develop a general architectural modification for deep neural networks to induce modularity and sparsity. This was an extension of a project carried out by our intern. Several other projects with NIH Principal Investigators (PIs) led to article submissions as preprints currently under review: • Investigators David Jangraw and Daniel Pine wanted to develop and evaluate an automated classification system for labeling Exposure Process Coding System quality codes -- specifically exposure and encourage events -- during in-person exposure therapy sessions. This was implemented using automatic speech recognition (ASR) and natural language processing techniques. • Investigators Briana Machens and Sofia Beas wanted to understand how the brain integrates internal physiological states with external sensory cues to guide behavior. • Investigators Javier Gonzalez-Castillo and Peter Bandettini wanted to analyze how the contents of thought during resting state functional MRI scans affect estimates of functional connectivity. • Investigators Ido Maor and Geoffrey Schoenbaum wanted to understand how the orbitofrontal cortex changes the schemas it learns to represent tasks, as demands change and conflict with previously acquired knowledge. • Investigators David Kupferschmidt and Joshua Gordon wanted to understand the interactions between neural populations involved in spatial working memory, specifically the ventral hippocampus and the medial prefrontal cortex. • Investigators Lauren Henry and Melissa Brotman wanted to develop robust models to automatically detect the presence of crying in audio recordings from infants. Finally, we have articles in preparation for submission on the following projects: • PI - Armin Raznahan: Interpretable meta-factorization of clinical questionnaires to identify general dimensions of psychopathology • PI - Bruno Averbeck: Analysis of neural data collected during a decision-making task paradigm ------------------- Ongoing Research ------------------- In addition to the projects listed above, we have ongoing research or service projects with researchers in the groups of Drs. Atlas, Afraz, Averbeck, Baker, Bandettini, Brotman, Lee, Lopez-Guzman, Martin, Marenco, Pine, Raznahan, Schmidt, Tejeda, and Zarate at NIMH, Aponte, Hope, Janes, and Schoenbaum at NIDA. As these are ongoing, they have not yet led to articles. We try to mix high-risk/high-impact projects with the potential to produce tools usable by several groups, with those satisfying specific needs of the PIs. These usually consume weeks to months of person-effort, depending on whether there exists a tool to satisfy the request, and the degree to which we have to handle the analysis process (versus advising a researcher in the group). ------------------- Methods under Development ------------------- We have a number of internal projects for developing new methods. There are four methods for which we have either preprints under review, code released, or both: 1) non-negative factorization method for questionnaire data, with additional constraints to further interpretability 2) extension of the statistical method for testing differences in time courses of photometry signals, as a function of covariates of interest (e.g., condition), accounting for confounds and individual variability; the extension allows covariates to change over the course of a trial, instead of only being allowed to take a single, scalar value per trial, as requested by multiple groups 3) statistical method to identify causal effects in closed loop designs in behavioral optogenetic experiments, or micro-randomized trial experimental designs 4) statistical method for longitudinal functional regression using a one-step functional Generalized Estimation Equation (GEE) algorithm (applicable to many different data types, e.g., calcium imaging, neural recordings, etc.) and we are also developing others, specifically: 1) an AI-based framework to extract structured information from free text (e.g., clinical notes), while displaying *where* in the text the AI identified the information, and allowing a human operator to check/modify the AI-extracted information. 2) an AI-based approach to producing labels that describe what dimensions represent, in a dimensional model produced using the VICE tool we developed in the past (e.g., for a model of concrete items, this approach might describe one dimension as representing animacy, another edibility, etc.). ------------------- Consultation/Service activity ------------------- We carried out ad-hoc consultations on machine learning methods and data analysis (which can take hours to days, if they require reading articles or finding/testing code), or co-advising of postbac and postdoc trainees (days-weeks), for researchers in NIMH, NIDA, and NINDS. Francisco Pereira served in the IRP AI Task Force, specifically as part of the subgroup writing a report for the director of IRP research, and the subgroup organizing an upcoming workshop on the development of AI models at NIH.