Publications

You can also find my articles on my Google Scholar profile.

10 papers

Group by

2026

When Medical VLMs Stop Understanding: MedTEC-Bench for Probing Semantic Specificity

A. H. Akash, A. Devkota, D. Adjeroh, B. Bhattarai, P. K. Gyawali

NeurIPS 2026 Evaluations & Datasets Track 2026

Medical vision–language models (VLMs) are attracting growing interest as a foundation for clinical image interpretation, retrieval, and decision support. However, current evaluations largely emphasize aggregate accuracy, retrieval, or classification performance, offering limited insight into whether these models possess the fine-grained semantic understanding required for safe and reliable use in high-stakes clinical settings. In this work, we introduce MedTEC-Bench, a diagnostic benchmark for probing semantic understanding in medical VLMs across five levels of increasing specificity, from broad modality recognition to fine-grained findings, and negation. MedTEC-Bench spans five medical imaging modalities: chest radiography, brain MRI, retinal fundus photography, dermoscopy, and histopathology. It combines controlled semantic probes with a suite of metrics designed to reveal failures hidden by aggregate benchmarks. Evaluating a diverse set of VLMs, from broad biomedical models to clinical-specialist models, we find that existing systems often fail on basic yet clinically meaningful semantic distinctions, despite strong reported performance on standard tasks. Our results suggest that current medical VLM evaluations may overestimate vision-grounded clinical understanding and highlight the need for targeted semantic benchmarks before such models are deployed in high-stakes medical workflows.

MedTEC-Bench is a diagnostic benchmark that probes semantic understanding in medical vision-language models across five levels of increasing specificity, from broad modality recognition through fine-grained findings to negation, spanning five imaging modalities. Evaluating eleven VLMs shows that models reporting strong retrieval and zero-shot scores often fail on basic yet clinically meaningful semantic distinctions.

Paper · coming soon

Federated Foundation Model for GI Endoscopy Images

A. Devkota, A. Amireskandari, J. Palko, S. Thakkar, D. Adjeroh, X. Jiang, B. Bhattarai, P. K. Gyawali

npj Digital Medicine 2026

Gastrointestinal (GI) endoscopy is essential in identifying GI tract abnormalities in order to detect diseases in their early stages and improve patient outcomes. Although deep learning has shown success in supporting GI diagnostics and decision-making, these models require curated datasets with labels that are expensive to acquire. Foundation models offer a promising solution by learning general-purpose representations, which can be finetuned for specific tasks, overcoming data scarcity. Developing foundation models for medical imaging holds significant potential, but the sensitive and protected nature of medical data presents unique challenges. Foundation model training typically requires extensive datasets, and while hospitals generate large volumes of data, privacy restrictions prevent direct data sharing, making foundation model training infeasible in most scenarios. In this work, we propose a FL framework for training foundation models for gastroendoscopy imaging, enabling data to remain within local hospital environments while contributing to a shared model. We explore several established FL algorithms, assessing their suitability for training foundation models without relying on task-specific labels, conducting experiments in both homogeneous and heterogeneous settings. We evaluate the trained foundation model on three critical downstream tasks—classification, detection, and segmentation—and demonstrate that it achieves improved performance across all tasks, highlighting the effectiveness of our approach in a federated, privacy-preserving setting.

We propose a federated learning framework to train foundation models for gastrointestinal endoscopy imaging, allowing hospitals to collaboratively develop general-purpose models while keeping data private. Our approach is evaluated on classification, detection, and segmentation tasks, demonstrating improved performance in a privacy-preserving, federated setting.

Paper Code Models

Towards Federated Learning Across Biobanks: Prototype Software from the 2026 Carnegie Mellon University–NVIDIA Hackathon

J. Mu et al.

BioHackrXiv 2026

The Carnegie Mellon University-NVIDIA Federated Learning Hackathon for Biomedical Applications (January 7-9, 2026) convened researchers from academia, government, and industry to implement federated frameworks for disease subtyping, genetic association studies, and multimodal clinical prediction using NVIDIA FLARE. This preprint presents ten projects spanning genome-wide association analyses, histopathology harmonization, pangenome construction, ancestry deconvolution, rare disease stratification, cancer subtyping, polygenic risk score aggregation, and multimodal fusion. These proofs of principle collectively demonstrate both the versatility of federated learning for biomedical applications and the technical considerations required for successful deployment.

This preprint presents prototype federated learning software developed during the 2026 Carnegie Mellon University–NVIDIA Federated Learning Hackathon for Biomedical Applications. The work demonstrates federated frameworks across biomedical tasks including disease subtyping, genetic association studies, histopathology harmonization, rare disease stratification, cancer subtyping, polygenic risk score aggregation, and multimodal clinical prediction.

Paper

FedVG: Gradient-Guided Aggregation for Enhanced Federated Learning

A. Devkota, J. Thrasher, D. Adjeroh, B. Bhattarai, P. K. Gyawali

CVPR Findings 2026 2026

Federated Learning (FL) enables collaborative model training across multiple clients without sharing their private data. However, data heterogeneity across clients leads to client drift, which degrades the overall generalization performance of the model. This effect is further compounded by overemphasis on poorly performing clients. To address this problem, we propose FedVG, a novel gradient-based federated aggregation framework that leverages a global validation set to guide the optimization process. Such a global validation set can be established using readily available public datasets, ensuring accessibility and consistency across clients without compromising privacy. In contrast to conventional approaches that prioritize client dataset volume, FedVG assesses the generalization ability of client models by measuring the magnitude of validation gradients across layers. Specifically, we compute layerwise gradient norms to derive a client-specific score that reflects how much each client needs to adjust for improved generalization on the global validation set, thereby enabling more informed and adaptive federated aggregation. Extensive experiments on both natural and medical image benchmarking datasets, across diverse model architectures, demonstrate that FedVG consistently improves performance, particularly in highly heterogeneous settings. Moreover, FedVG is modular and can be seamlessly integrated with various state-of-the-art FL algorithms, often further improving their results.

This study introduces FedVG, a gradient-guided federated aggregation framework designed to improve learning under client data heterogeneity. By using a global validation set and layerwise validation gradient norms, FedVG adaptively weights client models based on their generalization behavior, improving performance across natural and medical image benchmarks, especially in highly heterogeneous settings.

Paper Project Code

Local K-Similarity Constraint for Federated Learning with Label Noise

S. Amgain, P. Shrestha, B. Khanal, A. Devkota, Y. R. Shrestha, S. Baek, P. Gyawali, B. Bhattarai

IEEE International Symposium on Biomedical Imaging (ISBI) 2026 2026

Federated learning on clients with noisy labels is a challenging problem, as such clients can infiltrate the global model, impacting the overall generalizability of the system. Existing methods proposed to handle noisy clients assume that a sufficient number of clients with clean labels are available, which can be leveraged to learn a robust global model while dampening the impact of noisy clients. This assumption fails when a high number of heterogeneous clients contain noisy labels, making the existing approaches ineffective. In such scenarios, it is important to locally regularize the clients before communication with the global model, to ensure the global model isn’t corrupted by noisy clients. While pre-trained self-supervised models can be effective for local regularization, existing centralized approaches relying on pretrained initialization are impractical in a federated setting due to the potentially large size of these models, which increases communication costs. In that line, we propose a regularization objective for client models that decouples the pre-trained and classification models by enforcing similarity between close data points within the client. We leverage the representation space of a self-supervised pretrained model to evaluate the closeness among examples. This regularization, when applied with the standard objective function for the downstream task in standard noisy federated settings, significantly improves performance, outperforming existing state-of-the-art federated methods in multiple computer vision and medical image classification benchmarks. Unlike other techniques that rely on self-supervised pretrained initialization, our method does not require the pretrained model and classifier backbone to share the same architecture, making it architecture-agnostic.

This study introduces a local regularization objective for federated learning with noisy labels. The method uses the representation space of a self-supervised pretrained model to enforce similarity between nearby examples within each client, improving robustness in heterogeneous federated settings with many noisy clients.

Paper

2025

$AI analysis for ejection fraction estimation from 12-lead ECG$

AI analysis for ejection fraction estimation from 12-lead ECG

A. Devkota, R. Prajapati, A. El-Wakeel, D. Adjeroh, B. Patel, P. Gyawali

Scientific Reports 2025

Heart failure (HF) remains a leading global cause of cardiovascular deaths, with its prevalence expected to rise in the upcoming decade. Measuring the heart ejection fraction (EF) is crucial for diagnosing and monitoring HF. Although echocardiography is the gold standard for EF measurement, it is often inaccessible in remote areas due to its cost and complexity. In contrast, electrocardiography (ECG) is more readily available and affordable, and emerging research suggests a possible link between ECG signals and EF. In this work, we explore the potential of 12-lead ECG signals to estimate EF using various machine learning (ML) and deep learning (DL) models. While recent studies have considered the use of ML or DL for estimating EF, these algorithms are often trained and tested on urban-based populations. However, demographics like those in rural Appalachia, where disease prevalence is extremely high, have been overlooked, potentially due to the unavailability of large volumes of data. Moreover, there have been concerning reports regarding the fairness of AI predictions across different populations, making it crucial to understand the performance of AI models across diverse demographics before their widespread application. To address this, our study focuses on analyzing AI models for EF estimation in the rural Appalachian population. We utilized a 12-lead ECG dataset of 55,500 patients from WVU Medicine hospitals in West Virginia and employed a wide array of AI algorithms, ranging from Random Forest to modern deep learning-based methods like Transformers, to estimate EF. We also considered different thresholds for analyzing these AI algorithms and examined the impact of single and multi-lead ECG signals, and conducted model interpretability analysis. Overall, our comprehensive analysis demonstrated that deep learning-based algorithms achieved the highest performance, with an AUROC of around 0.86 for EF estimation from 12-lead ECG signals. Additionally, we found that while individual ECG leads were insufficient for accurate EF estimation, specific lead combinations significantly improved classification performance.

This study investigates the use of 12-lead ECG signals to estimate heart ejection fraction (EF) in a rural Appalachian population. Using a range of machine learning and deep learning models—including Transformers—our analysis shows deep learning models achieve the highest performance (AUROC ~0.86), with specific multi-lead combinations improving accuracy and model interpretability providing insights into predictive features.

Paper Code

Multimodal Federated Learning for Secure and Accurate Healthcare AI

J. Thrasher, A. Devkota, P. Siwakotai, R. Chivukula, P. Poudel, C. Hu, B. Bhattarai, P. Gyawali

Journal of Healthcare Informatics Research 2025

Recent advancements in multimodal machine learning have empowered the development of accurate and robust AI systems in the medical domain, especially within centralized database systems. Simultaneously, Federated Learning (FL) has progressed, providing a decentralized mechanism where data need not be consolidated, thereby enhancing the privacy and security of sensitive healthcare data. The integration of these two concepts supports the ongoing progress of multimodal learning in healthcare while ensuring the security and privacy of patient records within local data-holding agencies. This paper offers a concise overview of the significance of FL in healthcare and outlines the current state-of-the-art approaches to Multimodal Federated Learning (MMFL) within the healthcare domain. It comprehensively examines the existing challenges in the field, shedding light on the limitations of present models. Finally, the paper outlines potential directions for future advancements in the field, aiming to bridge the gap between cutting-edge AI technology and the imperative need for patient data privacy in healthcare applications.

This paper reviews the role of multimodal federated learning in healthcare, highlighting its potential to combine diverse medical data while preserving patient privacy. It surveys state-of-the-art approaches, identifies current challenges and limitations, and outlines future directions for advancing secure and effective healthcare AI.

Paper

2024

TE-SSL: Time and Event-aware Self Supervised Learning for Alzheimer's Disease Progression Analysis

J. Thrasher, A. Devkota, A. P. Tafti, B. Bhattarai, P. Gyawali, Alzheimer’s Disease Neuroimaging Initiative

MICCAI 2024 2024

Alzheimer’s Dementia (AD) represents one of the most pressing challenges in the field of neurodegenerative disorders, with its progression analysis being crucial for understanding disease dynamics and developing targeted interventions. Recent advancements in deep learning and various representation learning strategies, including self-supervised learning (SSL), have shown significant promise in enhancing medical image analysis, providing innovative ways to extract meaningful patterns from complex data. Notably, the computer vision literature has demonstrated that incorporating supervisory signals into SSL can further augment model performance by guiding the learning process with additional relevant information. However, the application of such supervisory signals in the context of disease progression analysis remains largely unexplored. This gap is particularly pronounced given the inherent challenges of incorporating both event and time-to-event information into the learning paradigm. Addressing this, we propose a novel framework, Time and Even-aware SSL (TE-SSL), which integrates time-to-event and event data as supervisory signals to refine the learning process. Our comparative analysis with existing SSL-based methods in the downstream task of survival analysis shows superior performance across standard metrics.

We introduce TE-SSL, a time and event-aware self-supervised learning framework for Alzheimer’s disease progression analysis. By incorporating time-to-event and event data as supervisory signals, TE-SSL improves representation learning and outperforms existing SSL methods in downstream survival analysis tasks.

Paper Code

2021

Near Real-Time Mobile Profiling and Modeling of Fine-Scale Environmental Proxies Along Major Road Lines of Nepal

N. B. Adhikari, S. Gautam, A. Devkota, S. Shikha, S. Pyakurel, M. P. Adhikari

ICMSI 2021

We present a methodology using GPS-enabled mobile sensors to collect and model fine-scale environmental proxies (e.g., temperature, CO₂, PM₂.₅) along major roadways in Nepal, demonstrating the effectiveness of ARIMA and RNNs for real-time and historical climate modeling.

Paper

2018

Sentence Ranking and Answer Pinpointing in Online Discussion Forums Utilizing User-generated Metrics and Highlights

S. Gautam, S. Shikha, A. Devkota, S. Pyakurel

NASCOIT 2018

This work presents a framework for extracting precise answers from online discussion forums by combining sentence ranking with user-generated metrics and highlights. The approach enables accurate answer pinpointing, improving search relevance and supporting better question-answering in forums.

Paper