Research Spotlight: Dr. Pakhomov

Research in Dr. Pakhomov’s lab is focused on development and validation of novel computational methods and technology for collection and analysis of unstructured time-series data. For example: continuous speech and language, heart rate, and galvanic skin response in order to understand how behavioral and physiological characteristics related to changes in cognition and mental states are affected by neurodegenerative and mental health disorders, effects of medications, and exposure to stress. Currently, Dr. Pakhomov’s lab is engaged in three NIH funded projects. 

Project 1: “FEASIBILITY OF USING WEARABLE TECHNOLOGY FOR JUST-IN-TIME PREDICTION OF SMOKING LAPSES” (funded by an R21 from NIDA)

Tell us about your research. 

The long-term goal of the proposed line of research is to develop technology-based interventions that increase smoking cessation rates by delivering just-in-time support to minimize the likelihood that smoking triggers lead to smoking lapses. Towards this objective we propose to use continuous physiological and environmental time-series data obtained from sensors embedded in widely used consumer wearable technology to construct a predictive model for detecting antecedents of smoking events (e.g., stressors, smoking cues, etc). Demonstrating that imminent smoking can be predicted would lead to the development and testing of just-in-time interventions that can be delivered via customized messaging on devices such as smartphones or smartwatches.

Did you have partners in your research?

Dr. Pakhomov and Dr. Kotlyar (Dept. of Experimental and Clinical Pharmacology) are Co-Principal Investigators on this project. 

What did you find?

The project started in August of 2020. We are currently in the preparation stage in which we are developing the technology (Apple and Android apps) needed to obtain physiological data remotely from participants in the study.

Why do your findings matter?

Current methods for smoking cessation result in low long-term cessation rates necessitating novel approaches to improve cessation success. The proposed project will use data from consumer wearable devices to identify physiological patterns that occur prior to smoking episodes in order to enable just-in-time delivery of interventions aimed at preventing relapse during smoking cessation attempts. The successful accomplishment of this project will lay the foundation for testing a broad range of behavioral and therapeutic interventions aimed at minimizing the likelihood that smoking triggers lead to smoking lapses and thereby improving the chances of successful smoking cessation.

What do you hope people will take away from your research?

We hope to provide a suite of tools and statistical models that will enable monitoring and just-in-time prediction of when a person attempting to quit smoking is at high risk for relapse. The next steps will be to apply these models in large interventional studies (clinical trials) to see if these approaches can help improve smoking cessation rates.

Links to relevant papers:

https://journals.plos.org/plosone/article/comments?id=10.1371/journal.pone.0229942
https://www.isca-speech.org/archive/interspeech_2011/i11_2949.html
https://psycnet.apa.org/record/2016-39458-001

Project 2: “COMPUTERIZED ASSESSMENT OF LINGUISTIC INDICATORS OF LUCIDITY IN ALZHEIMER'S DISEASE DEMENTIA” (funded by R21 from NIA)

Tell us about your research.

The focus of the proposed project is to enable automated detection and analysis of episodes of unexpected lucidity in individuals with late-stage dementia in which the individual long thought to have succumbed to dementia and lost most of his or her cognitive abilities temporarily regains the ability to communicate in a clear and coherent fashion. Currently, the evidence for the existence of these episodes is mostly anecdotal, stemming from reports by caregivers and healthcare professionals. According to these reports, clear speech and language are the most prominent features of episodes of cognitive lucidity. The very low frequency and unexpected nature of these episodes make it challenging to capture objective evidence in the form of audio or video recordings of these events needed to enable systematic and comprehensive investigations. In this feasibility project, we will develop technology to address two challenging issues: a) accurate conversion of continuous speech to text, and b) automated analysis of the text to measure the degree of coherence. Without robust solutions for these problems, our ability to detect and fully capture and analyze coherent speech in a long-term monitoring setting will remain limited. We will address these problems by developing and testing a robust automatic speech recognition solution based on deep learning technology that can operate autonomously (without sending data to external servers). We will also adapt existing and develop new measures of semantic coherence that are able to work on imperfect transcripts resulting from automatic speech recognition.

Did you have partners in your research?

This project is conducted in partnership with the University of Washington, WA (Dr. Trevor Cohen – Co-Principal investigator).

What did you find?

This project started in September 2020. We are currently in the preparation stages in which we are collecting and annotating data from existing corpora of speech collected from persons with various stages of dementia.

Why do your findings matter?

This project seeks to develop a validated approach to automatically monitoring people with advanced dementia who are thought to have lost their cognitive abilities for potential episodes in which they unexpectedly and temporarily regain their ability to communicate coherently. If this project is successful, our contribution will help in investigating the phenomenon of paradoxical lucidity to confirm its existence and characterize its features. This can potentially result in a fundamental change in our current understanding of how human brain functions.

What do you hope people will take away from your research?

We hope to develop and evaluate a system to record the speech produced in advanced dementia, convert it to text and measure the degree of coherence of the language produced with the intention to identify atypically lucid episodes. This system would then be used in larger long-term observational studies of individuals with late-stage dementia.
 
Links to relevant papers:

https://www.sciencedirect.com/science/article/pii/S1552526019300950https://www.aclweb.org/anthology/2020.acl-main.176/https://pubmed.ncbi.nlm.nih.gov/29502483/

Project 3: “OPEN HEALTH NATURAL LANGUAGE PROCESSING COLLABORATORY” (funded by NCATS)

Tell us about your research.

One of the major barriers in leveraging Electronic Health Record (EHR) data for clinical and translational science is the prevalent use of unstructured or semi-structured clinical narratives for documenting clinical information. Natural Language Processing (NLP), which extracts structured information from narratives, has received great attention and has played a critical role in enabling secondary use of EHRs for clinical and translational research. Current successful NLP use cases often require a strong informatics team (with NLP experts) to work with clinicians to supply their domain knowledge and build customized NLP engines iteratively. This requires close collaboration between NLP experts and clinicians, not feasible at institutions with limited informatics support. Additionally, the usability, portability, and generalizability of the NLP systems are still limited, partially due to the lack of access to EHRs across institutions to train the systems. The limited availability of EHR data limits the training available to improve the workforce competence in clinical NLP. We aim to address the above challenges by extending our existing collaboration among multiple CTSA hubs on open health natural language processing (OHNLP) to share distributional information of NLP artifacts (i.e., words, n-grams, phrases, sentences, concept mentions, concepts, and text segments) acquired from real EHRs across multiple institutions. 

Did you have partners in your research?

This research is being conducted in collaboration with several members of the OHNLP consortium and is led by the Mayo Clinic (Dr. Hongfang Liu). Other consortium members include University of Minnesota, University of Texas, and Columbia University. The University of Minnesota team is partnering with the UMN CTSI and IHI (Dr. Genevieve Melton).

What did you find?

The University of Minnesota team has been focusing on development and validation of a system designed to make it easy for investigators to use a variety of existing open NLP tools and to scale them to real datasets that tend to be very large (e.g., millions of mHealth records). The system also enables combining multiple NLP engines into ensembles. We found that composing ensembles of NLP engines results in improved overall performance on multiple tasks including investigations of information extraction in new clinical language domains such as pre-hospital trauma reports and COVID-19 outcome prediction from notes collected during hospital stay.

Why do your findings matter?

The project aims to broaden the secondary use of electronic health records (EHRs) across the research community by combining innovative privacy-preserving computing techniques and clinical natural language processing. The tools and approaches developed in this project will enable sharing of unstructured clinical information that may not be sufficient for reliable predictive models at any one individual site.

What do you hope people will take away from your research?

We have conducted a large number of dissemination efforts in the past 3 years of this project including several publications and tutorials on how to use the NLP ensembling system developed at the University of Minnesota. We hope that our work will help investigators with limited resources (i.e. those that are not able to hire a team of NLP engineers) to more fully benefit from contributions of the NLP community in their clinical research.

Links to relevant papers, etc.:

http://ohnlp.org/
https://github.com/nlpie/nlp-adapt
https://healthinformatics.umn.edu/research/nlpie-group/nlpie-resources

Tignanelli, C., et al. (2020). Natural language processing of prehospital emergency medical services trauma records allows for automated characterization of treatment appropriateness. Journal of trauma and acute care surgery, 88 (5), 607-614.
Bompelli, A., et al. (2020) Comparing NLP Systems to Extract Entities of Eligibility Criteria in Dietary Supplements Clinical Trials Using NLP-ADAPT. Proc of AIME 2020.
Silverman, G., et al., Named entity recognition in prehospital trauma care. (2019). Proc. MEDINFO 2019.
Finley, et al. Using ensembles of NLP engines without a common type system to improve abbreviation disambiguation. Proc AMIA 2017.