Machine Learning Team

National Institute Of Mental Health

Investigators

Linked publications & trials

Abstract

This is the fifth annual report for the Machine Learning Team. The team comprised research scientists Francisco Pereira, Charles Zheng, Patrick McClure (until March of 2022), Ka Chun Lam, Yuan Zhao, and Juan Antonio Lossio-Ventura; and postbac IRTA Nicole Kuznetsov. Dylan Nielson, a staff scientist, joined the team in January of 2022. We work very closely with the Data Science and Sharing Team (DSST), both on specific joint projects and informally across many others. Research activity: We continued IRP research projects initiated in the previous year. These were the projects that led to published articles: 1) Investigators: Xiaoyu Ma and Zheng Li. Goal: To develop an approach for testing whether the encoding of information by prefrontal cortex neurons changes between experimental conditions, subject to confounding factors. 2) Investigators: Lucrezia Liuzzi, Hanna Keren, and Argyris Stringaris. Goal: To develop models to predict mood during a gambling task, based on participant characteristics, trial parameters, and experiment history. This project was extended to magnetoencephalography (MEG) data. In addition, we had publications with external collaborators, or publications by team members that were prepared or completed after their joining the team, on the following topics: 1) Recovery of human object knowledge from word embeddings 2) Identifying silver linings during the pandemic from longitudinal survey data 3) Standardizing opioid prescriptions to morphine milligram equivalents from electronic health records 4) Mood and behaviors of adolescents with depression in a longitudinal study before and during the COVID-19 pandemic 5) Origins of anhedonia in childhood and adolescence 6) Clinical utility of family history of depression for prognosis of adolescent depression severity and duration assessed with predictive modeling Several other projects led to article submissions currently under review: 1) Martin Hebart (now an external collaborator) and Lukas Muttenthaler (Martins student) discovered interpretable mental representations of objects from a large database of behavioral judgments, which can be used to predict human behavior on a variety of other tasks. We developed an improved version of the method, using approximate Bayesian inference, which should perform better in smaller datasets (such as those collected by the NIMH IRPs Christopher Baker and Alex Martin, who are using the method). 2) MLT + DSST. We developed a new method for extracting brain maps from a deep neural network trained to decode which task is being performed by a subject. The maps show where the information used by the network is present. The article introduces two different quantitative evaluation procedures for gauging the quality of the maps and compares our method with existing ones. 3) We helped Dave Jangraw and Argyris Stringaris analyze imaging and behavioral data from many different experiments to show that rest periods lower the mood of participants, an effect we call resting-state dysphoria. 4) We helped Joyce Chung and the COVID-19 survey group develop a predictor of mental health status for participants, based on data from the subsample of participants that had also participated in prior NIH studies. The first article under review is a preprint describing an analysis to relate predicted mental health status to clinical symptoms and pandemic-related psychological and behavioral responses during lockdown. The second article under review is a preprint describing the analysis of free text produced by participants, including plots of emotional valence over time and summaries of the main topics present in such text. We have articles in preparation for submission on the following projects: 1) PI: Armin Raznahan (NIMH). Analysis of questionnaire data and its relation to brain imaging. 2) PI: Bruce Hope (NIDA). Analysis of neural data collected during an extinction paradigm. 3) PI: Soohyun Lee (NIMH). Analysis of neural data during spontaneous activity. 4) PI: Argyris Stringaris (NIMH). Analysis of a mood training paradigm. 5) PI: Joyce Chung (NIMH). Prediction model for temporal transition in various outcomes in a COVID-19 survey. 6) PI: Bruno Averbeck (NIMH). Analysis of neural data collected during a bandit task paradigm Ongoing Research In addition to all of the above, we have ongoing research or service projects with researchers in the groups of Drs. Atlas, Baker, Bandettini, Brotman, Lisanby, Martin, Merikangas, Penzo, Pine, Tejeda, Zarate at NIMH, and Schoenbaum and Curtis at NIDA, none of which have yet led to articles. We try to mix high-risk/high-impact projects 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). Building re-usable prediction models for neuroimaging We have continued our joint project with the DSST with the goal of building re-usable deep neural network components (DNN) for many different feature types derived from MR imaging modalities. We took advantage of the experience gained in the first iteration of the project to re-engineer most components. Since then, we have reprocessed a variety of large datasets with an improved pipeline and converted substantial parts of their phenotype information into a standardized format. This is intrinsically useful to IRP researchers, as each dataset would otherwise be processed differently, and require extensive curation to extract the phenotype information (often stored in complex ways). We are now in the process of improving the models that predict that phenotype information from imaging for individual datasets, as well as our ability to transfer them to new, smaller clinical datasets. Methods in Development We have a number of internal projects for developing new methods which, if shown to be effective, we will bring to the attention of PIs who might have an application for them. These include customized factor analysis methods (with domain-specific constraints) and extraction of structured information from free text data. Consultation/Service activity We carried out ad-hoc consultations on machine learning methods (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). In addition to the PIs mentioned above, we provided consulting to people in the NIMH IRP groups led by Drs. Conway, Cox, McMahon, Murray, Schmidt, White, Zarate, as well as Leggio (NIDA), and Haller (NINDS). Francisco Pereira participated in the IRP seminar committee, and the committee to select the new head of the SSCC. Education, Training, and Presentations Given the general fatigue with videoconferencing, and poor attendance at talks, we suspended the Machine Learning in Brain Imaging, Neuroscience and Psychology seminar, in collaboration with Javier Gomez-Castillo. We developed an Introduction to Data Science course, which was taught at FAES by Patrick McClure. This will serve as the basis for a course on data preparation that we are designing, with a focus on postbac IRTAs in labs that work with us.