About

Yogatheesan Varatharajah is an Assistant Professor of Computer Science & Engineering at the University of Minnesota and a Visiting Scientist at the Mayo Clinic. He is also affiliated with the Department of Neurology at the Mayo Clinic. His research broadly focuses on leveraging recent advances in machine learning to improve the healthcare system. He works closely with clinical experts to develop novel domain-guided machine learning applications aimed at reducing physician burden, augmenting their capabilities, and enhancing the overall patient experience while ensuring reliability, scalability, and trust. A particular focus of his work has been on improving treatments for neurological diseases by developing novel machine learning methods to model brain activity alterations in diseases such as Alzheimer's and epilepsy. Varatharajah obtained his Ph.D. and M.S. degrees from the Department of Electrical and Computer Engineering at the University of Illinois at Urbana-Champaign under the supervision of Prof. Ravi Iyer. During his graduate studies, he was mentored by Dr. Gregory Worrell at the Mayo Clinic through the Mayo-Clinic-Illinois Alliance. He earned his bachelor's degree in Electronic and Telecommunication Engineering at the University of Moratuwa in Sri Lanka. He also spent a summer at Google and collaborated with researchers at Google Health. His Ph.D. research was supported by a Mayo Clinic-Illinois Alliance Fellowship for Technology-based Healthcare Research and a Rambus Computer Engineering Fellowship. His Ph.D. thesis was awarded the CSL Ph.D. Thesis Award by the Coordinated Science Laboratory in 2021. Before joining the University of Minnesota, Varatharajah was a Research Assistant Professor in the Department of Bioengineering at the University of Illinois at Urbana-Champaign and a faculty affiliate at the Center for AI Innovations at the National Center for Supercomputing Applications. His work integrates machine learning with clinical insights to develop scalable and trustworthy healthcare technologies, particularly focusing on neurological disease diagnosis and treatment optimization.

Research topics

• Neuroscience
• Psychology
• Computer Science
• Medicine
• Artificial Intelligence
• Audiology
• Surgery

Selected publications

• EEG foundation models: a critical review of current progress and future directions

  Journal of Neural Engineering · 2026-02-10 · 2 citations

  articleOpen accessSenior author

  Abstract Premise. Patterns of electrical brain activity recorded via electroencephalography (EEG) offer immense value for scientific and clinical investigations. The inability of supervised EEG encoders to learn robust EEG patterns and their over-reliance on expensive signal annotations have sparked a transition towards general-purpose self-supervised EEG encoders, i.e. EEG foundation models (EEG-FMs), for robust and scalable EEG feature extraction. However, the real-world readiness of early EEG-FMs and the rubrics for long-term research progress remain unclear. Objective. In this work, we conduct a review of ten early EEG-FMs to capture common trends and identify key directions for future development of EEG-FMs. Methods. We comparatively analyze each EEG-FM using three fundamental pillars of foundation modeling, namely the representation of input data, self-supervised modeling, and the evaluation strategy. Based on this analysis, we present a critical synthesis of EEG-FM methodology, empirical findings, and outstanding research gaps. Results. We find that most EEG-FMs adopt a sequence-based modeling scheme that relies on transformer-based backbones and the reconstruction of masked temporal EEG sequences for self-supervision. However, model evaluations remain heterogeneous and largely limited, making it challenging to assess their practical off-the-shelf utility. In addition to adopting standardized and realistic evaluations, future work should demonstrate more substantial scaling effects and make principled and trustworthy choices throughout the EEG representation learning pipeline. Significance. Our review indicates that the development of benchmarks, software tools, technical methodologies, and applications in collaboration with domain experts may advance the translational utility and real-world adoption of EEG-FMs.

  PublisherDOI

• Population-based spectral characteristics of normal interictal scalp EEG inform diagnosis and treatment planning in focal epilepsy

  Scientific Reports · 2025-07-11 · 1 citations

  articleOpen accessSenior author

  Normal routine electroencephalograms (EEGs) can cause delays in the diagnosis and treatment of epilepsy, especially in drug-resistant patients and those without structural abnormalities. There is a need for alternative quantitative approaches that can inform clinical decisions when traditional visual EEG review is inconclusive. We leverage a large population EEG database (N = 13,652 recordings, 12,134 unique patients) and an independent cohort of patients with focal epilepsy (N = 121) to investigate whether normal EEG segments could support the diagnosis of focal epilepsy. We decomposed expertly graded normal EEGs (N = 6,242) using unsupervised tensor decomposition to extract the dominant spatio-spectral patterns present in a clinical population. We then, using the independent cohort of patients with focal epilepsy, evaluated whether pattern loadings of normal interictal EEG segments could classify focal epilepsy, the epileptogenic lobe, presence of lesions, and drug response. We obtained six physiological patterns of EEG spectral power and connectivity with distinct spatio-spectral signatures. Both pattern types together effectively differentiated patients with focal epilepsy from non-epileptic controls (mean AUC 0.78) but failed to classify the epileptogenic lobe. Spectral power-based patterns best classified drug-resistant epilepsy (mean AUC 0.73) and lesional epilepsy (mean AUC 0.67), albeit with high variability across patients. Our findings support that visibly normal patient EEGs contain subtle quantitative differences of clinical relevance. Further development may yield normal EEG-based computational biomarkers that can augment traditional EEG review and epilepsy care.

  PublisherOA PDFDOI

• A Robust Deep Learning Framework for Detecting Bursts in Muscle Sympathetic Nerve Activity

  2025-07-14

  articleSenior author

  Muscle sympathetic nerve activity (MSNA) is a key physiological signal that provides insights into the functioning of the sympathetic nervous system. Characterized by bursts of neural activity, MSNA signals play a crucial role in understanding both normal and pathological states. Accurately detecting these bursts is essential for quantitative analysis and deeper exploration of sympathetic nerve dynamics. However, the tedious task of detecting bursts is currently performed by trained experts, leading to potential burnout and increased risk of error. In this study, we present a novel machine learning-based burst detection method that combines integrated MSNA activity and electrocardiography activity in a convolutional neural network to robustly identify burst peaks. Our approach achieves an average F1 score of 0.87±0.03 in detecting expert-annotated bursts in a dataset including resting autonomic nervous system recordings of 41 healthy female participants when evaluated under a five-fold cross-validation. Our approach outperformed several alternative methods including some previously published automated burst detection approaches.

  PublisherDOI

• The Value of Normal Interictal EEGs in Epilepsy Diagnosis and Treatment Planning: A Retrospective Cohort Study using Population-level Spectral Power and Connectivity Patterns

  medRxiv · 2025-01-05

  preprintOpen accessSenior authorCorresponding

  Introduction: Scalp electroencephalography (EEG) is a cornerstone in the diagnosis and treatment of epilepsy, but routine EEG is often interpreted as normal without identification of epileptiform activity during expert visual review. The absence of interictal epileptiform activity on routine scalp EEGs can cause delays in receiving clinical treatment. These delays can be particularly problematic in the diagnosis and treatment of people with drug-resistant epilepsy (DRE) and those without structural abnormalities on MRI (i.e., MRI negative). Thus, there is a clinical need for alternative quantitative approaches that can inform diagnostic and treatment decisions when visual EEG review is inconclusive. In this study, we leverage a large population-level routine EEG database of people with and without focal epilepsy to investigate whether normal interictal EEG segments contain subtle deviations that could support the diagnosis of focal epilepsy. Data & Methods: We identified multiple epochs representing eyes-closed wakefulness from 19-channel routine EEGs of a large and diverse neurological patient population (N=13,652 recordings, 12,134 unique patients). We then extracted the average spectral power and phase-lag-index-based connectivity within 1-45Hz of each EEG recording using these identified epochs. We decomposed the power spectral density and phase-based connectivity information of all the visually reviewed normal EEGs (N=6,242) using unsupervised tensor decompositions to extract dominant patterns of spectral power and scalp connectivity. We also identified an independent set of routine EEGs of a cohort of patients with focal epilepsy (N= 121) with various diagnostic classifications, including focal epilepsy origin (temporal, frontal), MRI (lesional, non-lesional), and response to anti-seizure medications (responsive vs. drug-resistant epilepsy). We analyzed visually normal interictal epochs from the EEGs using the power-spectral and phase-based connectivity patterns identified above and evaluated their potential in clinically relevant binary classifications. Results: We obtained six patterns with distinct interpretable spatio-spectral signatures corresponding to putative aperiodic, oscillatory, and artifactual activity recorded on the EEG. The loadings for these patterns showed associations with patient age and expert-assigned grades of EEG abnormality. Further analysis using a physiologically relevant subset of these loadings differentiated patients with focal epilepsy from controls without history of focal epilepsy (mean AUC 0.78) but were unable to differentiate between frontal or temporal lobe epilepsy. In temporal lobe epilepsy, loadings of the power spectral patterns best differentiated drug-resistant epilepsy from drug-responsive epilepsy (mean AUC 0.73), as well as lesional epilepsy from non-lesional epilepsy (mean AUC 0.67), albeit with high variability across patients. Significance: Our findings from a large population sample of EEGs suggest that normal interictal EEGs of patients with epilepsy contain subtle differences of predictive value that may improve the overall diagnostic yield of routine and prolonged EEGs. The presented approach for analyzing normal EEGs has the capacity to differentiate several diagnostic classifications of epilepsy, and can quantitatively characterize EEG activity in a scalable, expert-interpretable, and patient-specific fashion. Further technical development and clinical validation may yield normal EEG-derived computational biomarkers that could augment epilepsy diagnosis and assist clinical decision-making in the future.

  PublisherOA PDFDOI

• Improving Out-of-distribution Human Activity Recognition via IMU-Video Cross-modal Representation Learning

  ArXiv.org · 2025-07-17

  preprintOpen accessSenior author

  Human Activity Recognition (HAR) based on wearable inertial sensors plays a critical role in remote health monitoring. In patients with movement disorders, the ability to detect abnormal patient movements in their home environments can enable continuous optimization of treatments and help alert caretakers as needed. Machine learning approaches have been proposed for HAR tasks using Inertial Measurement Unit (IMU) data; however, most rely on application-specific labels and lack generalizability to data collected in different environments or populations. To address this limitation, we propose a new cross-modal self-supervised pretraining approach to learn representations from large-sale unlabeled IMU-video data and demonstrate improved generalizability in HAR tasks on out of distribution (OOD) IMU datasets, including a dataset collected from patients with Parkinson's disease. Specifically, our results indicate that the proposed cross-modal pretraining approach outperforms the current state-of-the-art IMU-video pretraining approach and IMU-only pretraining under zero-shot and few-shot evaluations. Broadly, our study provides evidence that in highly dynamic data modalities, such as IMU signals, cross-modal pretraining may be a useful tool to learn generalizable data representations. Our software is available at https://github.com/scheshmi/IMU-Video-OOD-HAR.

  PublisherOA PDFDOI

• A pilot randomized controlled double-blind trial of intermittent theta burst stimulation (iTBS) repetitive transcranial magnetic stimulation (rTMS) to improve memory in mild cognitive impairment (MCI): a study protocol

  Pilot and Feasibility Studies · 2025-04-01 · 4 citations

  articleOpen access

  BACKGROUND: Mild cognitive impairment (MCI), prevalent among older adults, often precedes Alzheimer's disease (AD) or Alzheimer's disease-related dementias (ADRD), emphasizing the need for effective interventions. Early intervention in MCI is crucial, not only to alleviate symptoms but to potentially delay the progression of cognitive decline. The lack of definitive treatments for MCI has prompted the exploration into alternative non-pharmacological therapeutic approaches. Specifically, noninvasive brain stimulation using repetitive transcranial magnetic stimulation (rTMS) has demonstrated promise in improving cognition in MCI and AD. OBJECTIVES: Our study will test the feasibility of using intermittent theta burst stimulation (iTBS) technique of rTMS in MCI, pilot test the study design, and collect pilot data on the effect of iTBS over three different brain regions on working memory, new learning, and executive function in MCI. Exploratory objectives are to assess the feasibility and usefulness of functional magnetic resonance imaging (fMRI), high-density electroencephalography (HD-EEG), and sleep architecture as potential biomarkers in response to iTBS. METHODS: A pilot randomized double-blind controlled cross-over trial of iTBS on 20 MCI participants randomized to 10 days of active iTBS (left dorsolateral prefrontal cortex or left lateral parietal cortex) or control (vertex). After 4-6-week washout period, they cross over to the alternative treatment arm for another 10 days. Each participant will undergo a total of 20 iTBS sessions. Pre- and post-iTBS assessments include neuropsychological tests, fMRI, HD-EEG, and sleep architecture. DISCUSSION: This innovative study aims to test the feasibility of iTBS as a cognitive enhancement strategy in MCI. If our study is feasible, it could lead to a future larger trial to further test whether iTBS can modulate underlying neurobiology and offer a therapeutic avenue to remediate cognitive decline in MCI or ultimately delay progression to dementia. TRIAL REGISTRATION: ClinicalTrials.gov, NCT05327257. Registered 04 April 2022.

  PublisherOA PDFDOI

• Transcranial direct current simulation as an adjunctive treatment for treatment-resistant depression in hospitalized patients: A feasibility study protocol

  PLoS ONE · 2025-06-10

  articleOpen accessCorresponding

  Transcranial direct current stimulation (tDCS) is clinically effective in treating treatment-resistant depression (TRD), as measured by response, symptom improvement, and disease remission. However, the feasibility and underlying mechanism of tDCS treatment in individuals with TRD during acute psychiatric hospitalization remain poorly characterized. This paper outlines the protocol that aims to investigate the feasibility of implementing a 5-day tDCS treatment in hospitalized patients with TRD and secondarily explore the effects on depression and cognition, and neurophysiological mechanisms underlying tDCS. Current study will enroll ten participants who are diagnosed with TRD and are hospitalized in psychiatric units. Participants will receive a 5-day tDCS treatment protocol, with each treatment session lasting for 30 minutes, delivered twice daily, for a total of 10 stimulations over 5 days. The primary outcomes are the feasibility, acceptability, and tolerability of administering a 5-day tDCS treatment protocol in acutely hospitalized TRD patients. Exploratory outcomes pre- and post-tDCS include measures of depression (Montgomery-Asberg Depression Rating Scale (MADRS)) and cognition (Stroop Test, Revised Hopkins Verbal Learning Test (HVLT-R), Digital Symbol Coding Test (DSCT)), EEG changes in peak alpha frequency (PAF), and cerebral hemodynamic changes by functional near-infrared spectroscopy (fNIRS). This protocol would provide feasibility evidence for tDCS as an add-on to the standard of care treatment of TRD in hospitalized patients. Upon completion of the protocol, the preliminary effects of the 5-day tDCS treatment protocol regarding depression and cognitive symptoms and its neurophysiological mechanisms will be identified to guide the design and delivery of a randomized controlled study. Trial registration: National Institute of Health Clinicaltrials.gov (NCT06236711) and protocol ID: 23-003274.

  PublisherDOI

• Development of a deep learning model to classify choroidal melanoma risk factors based on color fundus photographs

  AJO International · 2025-08-31 · 1 citations

  articleOpen accessCorresponding

  Choroidal melanoma is the most common malignant primary intraocular tumor and can develop either de novo or from a preexisting choroidal nevus, a benign pigmented lesion. Key risk factors for the transformation of choroidal nevus into melanoma include tumor diameter > 5 mm, tumor thickness > 2 mm, orange pigment, subretinal fluid, and low internal reflectivity on ultrasound. However, the assessment of many of these risk factors requires multimodal imaging equipment and skilled subspecialists, only available at tertiary referral centers. In this study, we developed and validated a deep learning approach to identifying these risk factors based solely on fundus images of choroidal nevi. Results indicate acceptable to excellent predictive performance for detection of all five risk factors. These findings suggest that deep learning models may be valuable tools for identifying high-risk choroidal nevi, particularly in resource-limited settings.

  PublisherDOI

• SEEG4D: a tool for 4D visualization of stereoelectroencephalography data

  Frontiers in Neuroinformatics · 2024-09-03 · 1 citations

  articleOpen access

  Epilepsy is a prevalent and serious neurological condition which impacts millions of people worldwide. Stereoelectroencephalography (sEEG) is used in cases of drug resistant epilepsy to aid in surgical resection planning due to its high spatial resolution and ability to visualize seizure onset zones. For accurate localization of the seizure focus, sEEG studies combine pre-implantation magnetic resonance imaging, post-implant computed tomography to visualize electrodes, and temporally recorded sEEG electrophysiological data. Many tools exist to assist in merging multimodal spatial information; however, few allow for an integrated spatiotemporal view of the electrical activity. In the current work, we present SEEG4D, an automated tool to merge spatial and temporal data into a complete, four-dimensional virtual reality (VR) object with temporal electrophysiology that enables the simultaneous viewing of anatomy and seizure activity for seizure localization and presurgical planning. We developed an automated, containerized pipeline to segment tissues and electrode contacts. Contacts are aligned with electrical activity and then animated based on relative power. SEEG4D generates models which can be loaded into VR platforms for viewing and planning with the surgical team. Automated contact segmentation locations are within 1 mm of trained raters and models generated show signal propagation along electrodes. Critically, spatial-temporal information communicated through our models in a VR space have potential to enhance sEEG pre-surgical planning.

  PublisherOA PDFDOI

• Data-driven retrieval of population-level EEG features and their role in neurodegenerative diseases

  Brain Communications · 2024-01-01 · 11 citations

  articleOpen access

  Electrophysiologic disturbances due to neurodegenerative disorders such as Alzheimer's disease and Lewy Body disease are detectable by scalp EEG and can serve as a functional measure of disease severity. Traditional quantitative methods of EEG analysis often require an a-priori selection of clinically meaningful EEG features and are susceptible to bias, limiting the clinical utility of routine EEGs in the diagnosis and management of neurodegenerative disorders. We present a data-driven tensor decomposition approach to extract the top 6 spectral and spatial features representing commonly known sources of EEG activity during eyes-closed wakefulness. As part of their neurologic evaluation at Mayo Clinic, 11 001 patients underwent 12 176 routine, standard 10-20 scalp EEG studies. From these raw EEGs, we developed an algorithm based on posterior alpha activity and eye movement to automatically select awake-eyes-closed epochs and estimated average spectral power density (SPD) between 1 and 45 Hz for each channel. We then created a three-dimensional (3D) tensor (record × channel × frequency) and applied a canonical polyadic decomposition to extract the top six factors. We further identified an independent cohort of patients meeting consensus criteria for mild cognitive impairment (30) or dementia (39) due to Alzheimer's disease and dementia with Lewy Bodies (31) and similarly aged cognitively normal controls (36). We evaluated the ability of the six factors in differentiating these subgroups using a Naïve Bayes classification approach and assessed for linear associations between factor loadings and Kokmen short test of mental status scores, fluorodeoxyglucose (FDG) PET uptake ratios and CSF Alzheimer's Disease biomarker measures. Factors represented biologically meaningful brain activities including posterior alpha rhythm, anterior delta/theta rhythms and centroparietal beta, which correlated with patient age and EEG dysrhythmia grade. These factors were also able to distinguish patients from controls with a moderate to high degree of accuracy (Area Under the Curve (AUC) 0.59-0.91) and Alzheimer's disease dementia from dementia with Lewy Bodies (AUC 0.61). Furthermore, relevant EEG features correlated with cognitive test performance, PET metabolism and CSF AB42 measures in the Alzheimer's subgroup. This study demonstrates that data-driven approaches can extract biologically meaningful features from population-level clinical EEGs without artefact rejection or a-priori selection of channels or frequency bands. With continued development, such data-driven methods may improve the clinical utility of EEG in memory care by assisting in early identification of mild cognitive impairment and differentiating between different neurodegenerative causes of cognitive impairment.

  PublisherOA PDFDOI

Recent grants

• CRII: SCH: Domain-guided Machine Learning for Clinical Decision Support in Epilepsy

  NSF · $175k · 2021–2023

Frequent coauthors

• Gregory A. Worrell

  WinnMed

  52 shared

• Benjamin H. Brinkmann

  Mayo Clinic

  42 shared

• Václav Křemen

  Mayo Clinic

  41 shared

• Brent Berry

  WinnMed

  33 shared

• Ravishankar K. Iyer

  20 shared

• Petr Nejedlý

  Masaryk University

  15 shared

• Krishnakant Saboo

  13 shared

• Jan Cimbálník

  University Hospital Brno

  13 shared

Awards & honors

• 2024: National Science Foundation Faculty Early Career Devel…