Publications

Developed a method for diagnosis and severity assessment of PD using a model based on Gramian Angular Fields in combination with deep Convolutional Neural Networks (CNNs)

This work introduces Deep Geometric Moments (DGM) as a novel, training-free guidance mechanism for text-to-image diffusion models. Unlike existing guidance techniques (e.g., segmentation maps, depth maps, or CLIP features), which impose rigid spatial constraints or rely heavily on global semantics, DGM captures fine-grained, subject-specific visual features through robust geometric representations. The proposed method uses a pretrained DGM model during the diffusion process to steer image generation in a flexible yet identity-preserving manner. Experiments show that DGM achieves a better balance between control and diversity, enabling more nuanced and visually consistent image synthesis without retraining the diffusion model.

A privacy-preserving system that leverages Gramian Angular Field (GAF) transformations, Federated Learning, and wearable sensor data to detect Freezing of Gait (FoG) in individuals with Parkinson’s Disease

We introduce NutriGen, a framework based on large language models (LLM) designed to generate personalized meal plans that align with user-defined dietary preferences and constraints.

We introduce MetaBoost, a novel hybrid framework that integrates SMOTE, ADASYN, and CTGAN, optimizing synthetic data generation through weighted averaging and iterative weight tuning to enhance the model’s performance (achieving a 1.87% accuracy improvement over individual balancing techniques).

We introduces a novel Transformer-based framework, AttenGluco, designed to improve long-term blood glucose level forecasting using multimodal data—including CGM and activity signals. By integrating cross-attention and multi-scale attention mechanisms, the model effectively fuses time-series data with different sampling rates and captures long-term dependencies, outperforming a multimodal LSTM baseline by up to 12% in RMSE across multiple cohorts (healthy, prediabetes, and type 2 diabetes) using the AI-READI dataset.

MealMeter is a linear regression based technique applied on multi-modal data collected using a CGM sensor and a wristband and tracks meal macronutrients. MealMeter achieves as low as 0.37 average root mean squared relative errors (RMSRE) in carb tracking, which is at least 15.9% improvement compared to TabPFN foundational model and other baselines.

LEAD is an explainable AI technique that determines relative feature contributions by characterizing the decision boundary and perturbing critical samples along the decision boundary close to the test sample. LEAD achieves at least 6% improved fidelity and 7% improved consistency compared to LIME and SHAP.

This method reduces the labeling cost to train an efficient multi-task neural network by adding an additional clustering step in the multi-task active learning loop, within the sampling process

CUDLE is a self-supervised learning framework that improves cannabis use detection from wearable sensor data in real-world settings, outperforming supervised methods while requiring significantly fewer labeled samples.

Chronic cannabis users show disrupted diurnal cortisol rhythms, including a blunted cortisol awakening response and elevated afternoon cortisol levels, but no major differences in diurnal heart rate variability or electrodermal activity, except for increased evening heart rate. Acute cannabis use reduced cortisol, subjective stress, and electrodermal activity. These findings suggest dysregulated stress responses in cannabis users, potentially linked to later waking times and cannabis’s stress-relieving effects, warranting further research on the relationship between cannabis use, cortisol rhythms, and psychological disorders.

By integrating BERT-based word embeddings with domain-specific knowledge (i.e., MET values), FUSE-MET optimizes label merging, reducing label complexity and improving classification accuracy.

Glyman incorporates stakeholders' choices in producing counterfactual explanations to reduce the number of abnormal glycemic events T1D patients encounter.

We propose a self-supervised learning (SSL) approach for cognitive workload classification using wavelet-based augmentations of EEG signals.

We propose a personalized hydration monitoring system with wearable and machine learning.

We introduce a framework for the channel selection problem, mathematically formulate the minimum-cost channel selection (MCCS), and propose two novel algorithms to solve the MCCS problem.

Our patient-independent model achieved an overall accuracy of 78% in detecting FoG events using both medication ‘On’ and ‘Off’ state data.

In this work, we take the first stride in studying adversarial transferability in wearable sensor systems from four viewpoints: (1) transferability between machine learning models; (2) transferability across subjects of the embedded system; (3) transferability across sensor body locations; and (4) transferability across datasets used for model training.

We propose a resource-efficient, real-time human activity recognition framework, transforming the multi-class classification problem into a hierarchical model based on the Metabolic Equivalent of Task (MET), creating a personalized structure for each individual.

We present a data-driven approach for stress detection based on convolutional neural networks while addressing the problems of the best sensor channel and the lack of knowledge about stress episodes.

We use an MLP model to identify and detect moments of cannabis use. We later use transfer learning to further enhance model accuracy.

We propose an automated risk prediction system that can make recommendations to clinicians in real-time with machine learning classifiers that predict the risk of developing neurological impairment.

Accurately forecasts future blood glucose level and predicts dysglycemic events thereby.

GlySim forecasts blood glucose level from medication, consumption and exercise information and intervenes beforehand if an abnormal event is forthcoming.

Accurately estimates inter-beat intervals from noisy ECG signals by suppressing the noise and making the peaks more prominent.

Cross domain activity recognition using transfer learning

Briefly evaluates the factors behind medication adherence of patients at risk for atherosclerotic cardiovascular disease.

Optimizes change in consumed nutrients while driving users to their desired diet.

A multi-modal dataset for stress and alcohol relapse collected in real world settings.

We focus on devising algorithms that combine data about physical activity and engagement with the app to predict future physical activity performance.

CNN model developed, quantized, and deployed on Android platform to screen PD patients from hand-drawn spirals.

Warn users of potential hyperglycemic events in advance and provide interpretation for decisions made by models to modify lifestyle.

A cascading approach for optimizing energy and memory requirements of a machine learning model running on an embedded device.

A stress classification and personalization framework based on Convolutional Neural Networks.

Continual learning formulation in the context of sensor systems.

We propose an algorithm that can reconstruct the missing input data for unavailable sensors.

*Selected for oral presentation and received runner-up award in CVPR 2020 continual learning workshop

Diet and physical activity are known as important lifestyle factors in self-management and prevention of many chronic diseases. Mobile sensors such as accelerometers have been used to measure physical activity or detect eating time. In many intervention studies, however, stringent monitoring of overall dietary composition and energy intake is needed. Currently, such a monitoring relies on self-reported data by either entering text or taking an image that represents food intake. These approaches suffer from limitations such as low adherence in technology adoption and time sensitivity to the diet intake context. In order to address these limitations, we introduce development and validation of Speech2Health, a voice- based mobile nutrition monitoring system that devises speech processing, natural language processing (NLP), and text mining techniques in a unified platform to facilitate nutrition monitoring. After converting the spoken data to text, nutrition-specific data are identified within the text using an NLP-based approach that combines standard NLP with our introduced pattern mapping technique. We then develop a tiered matching algorithm to search the food name in our nutrition database and accurately compute calorie intake values. We evaluate Speech2Health using real data collected with 30 participants. Our experimental results show that Speech2Health achieves an accuracy of 92.2% in computing calorie intake. Furthermore, our user study demonstrates that Speech2Health achieves significantly higher scores on technology adoption metrics compared to text-based and image-based nu- trition monitoring. Our research demonstrates that new sensor modalities such as voice can be used either standalone or as a complementary source of information to existing modalities to improve accuracy and acceptability of mobile health technologies for dietary composition monitoring.