InfoFORCE: Research Posters

Author(s): Sai Shivani Parimi, Bishnu Sarker, Alexis Lamar, Hari Chandana Vadde

Recent advances in artificial intelligence have led to the emergence of multi-agent health systems capable of integrating diverse data sources and supporting complex decision-making in healthcare.These systems demonstrate the potential to analyze multimodal health data and generate personalized insights but their application has largely focused on individual-level care and often overlooks broader population health challenges,particularly those related to health equity.In conditions such as Type 2 Diabetes (T2D), traditional risk stratification models frequently rely on clinical indicators alone,failing to account for the significant impact of social determinants of health (SDOH) on patient outcomes.This study proposes a conceptual multi-agent AI framework designed to enhance population-level diabetes risk stratification by integrating SDOH into electronic health record (EHR)-based models.The framework extends existing multi-agent architectures by introducing specialized and coordinated agents with distinct capabilities.A population orchestrator agent manages task allocation and system coordination, a data science agent enables SDOH-integrated predictive modeling and subgroup analysis;a domain expert agent ensures clinically grounded,evidence-based interpretation,and a health coach agent translates insights into culturally sensitive, actionable interventions.In addition,the framework introduces supporting agents,including health equity agent for bias monitoring, a data quality agent for ensuring data completeness, and an SDOH extraction agent for structuring unstructured social data.The proposed framework emphasizes equitable identification of at-risk populations, particularly individuals with moderate clinical risk but high social vulnerability.This work is conceptual and has not yet been implemented or empirically validated.Future research is required to evaluate model performance, system integration feasibility, and real-world impact.Despite these limitations, the framework provides a foundation for advancing equitable,AI-driven population health interventions. 

Author(s):  Laxmi Shravani Mamidala, Ashwitha Reddy Venna, Gayathri Orsu, Lakshmi Triveni Muthyala, Tharun Swaminathan Ravi Kumar

Blood cancer, especially leukemia, is one of the most dangerous diseases, and early, accurate diagnosis is critical for improving patient outcomes. Conventional microscopic analysis of blood smear images is time-consuming, highly subjective, and prone to variability. This research proposes an automated Blood Cancer Detection System using a combination of machine learning, deep learning, and computer vision techniques to identify and classify abnormal blood cells.

The proposed system uses a hybrid artificial intelligence approach. Deep features are extracted from microscopic images using a pre-trained ResNet50 convolutional neural network, and these features are classified using machine learning algorithms such as Logistic Regression, Support Vector Machine (SVM), Random Forest, XGBoost, and K-Means clustering. To improve classification, YOLOv8 is used for accurate object detection and localization of blood cells. Model interpretability is enhanced using Grad-CAM explainability techniques, which highlight the most important regions of the image for the classification decision.

The dataset contains 5000 labeled microscopic images of blood cells, representing the five most prominent cell types essential for leukemia diagnosis. Experimental results show excellent performance of the supervised models, with SVM and Logistic Regression achieving about 98–99% accuracy, while YOLOv8 demonstrates strong accuracy in blood cell detection.

As an additional contribution, this research proposes an intelligent clinical decision-support system that allows users to upload images, display predictions, analyze explainable heatmaps, and compare predictions across models. Future work may include integration with real-time clinical databases, the use of federated learning for privacy-preserving collaboration, and multimodal patient data to further improve prediction accuracy. Overall, this system shows the potential of combining deep learning, machine learning, and explainable AI to accelerate biomedical image analysis and support faster, more accurate blood cancer diagnosis.

Author(s):  Unknown

As artificial intelligence increasingly mediates how individuals learn, communicate, and access information, questions of identity and authenticity have become central to human experiences and expression. This poster explores how AI-driven information environments influence an individual’s ability to develop, express, and sustain an authentic sense of self. Grounded in humanistic psychology, Self-Determination Theory, narrative identity theory, and information science, this work synthesizes current research to examine the complex relationship between AI systems, digital information ecosystems, and identity formation. A conceptual model is proposed in which AI systems shape the information environment, which in turn influence identity processes such as self-reflection, autonomy, and narrative construction, ultimately affecting authentic self-expression. Findings from the literature suggest that AI tools can support self-awareness and creativity, yet algorithmic curation may also narrow identity exploration and dilute personal voice. Implications are offered for educators, AI designers, and researchers seeking to promote agency, transparency, and critical information literacy in AI-mediated contexts. This work highlights the importance of intentional design and reflective practice to ensure that authenticity—and the human capacity for meaning-making—remains possible in an increasingly algorithmic world.

Author(s):  Kamrun Nahar, Sai Harshith Kolavasi, Krishna Mogaveera, Maimuna Patwary

This research focuses on the automatic detection of defects fabrics and quality prediction in the textile industry. Our study analyzed the finished textile product using a deep learning framework based, Convolution Neural Network or CNN in short. CNN is an artificial intelligence model that processes data and extracts meaningful insights. Just as human neurons receive input through our nervous system, CNNs receive input from cameras or sensors and process it through layered 2D and 3D structures. These networks can be trained to classify images, test patterns, and generate accurate outputs. Thus, we developed a supervised Convolutional Neural Network (CNN) Model that can detect and categorize and classify the defects in textile fabrics. High quality defective and non-defective images of various kinds were used to train and validate the CNN model architecture, to detect the good quality fabric automatically for inventory and final distribution.

We aim to create a system that can automatically detect, classify, and predict textile defects. Raw data was collected, cleaned, verified, and finally classified for its defect identification according to the industry standard. This project seeks not only to improve the quality assurance process in the textile industry but also potentially to revolutionize defect prediction and prevention. Integrating machine learning into textile manufacturing is crucial for enabling smart and automated quality control. By combining deep learning with domain expertise, we can create a system for textile industry’s employees or workers who mostly are work on finish products and test excellence that enhances fabric production quality, reduces material waste, minimizes costly errors in the production cycle, and support the evolution of intelligent textile manufacturing environment.

Author(s):  Ampana Jayaram, Lizal Adhikari, Dr. Haihua Chen

If you've ever tried to compare academic programs across multiple universities, you already know how frustrating it can be. Every school has a different website, a different catalogue format, and a different way of presenting the same information. There is no single place where a student, advisor, or researcher can search across thousands of institutions, compare them side by side, and get personalised guidance until now.
Edublox was built to solve exactly that problem. The project collects academic program data from 6,186 accredited U.S. universities and brings it together into one clean, searchable interface. The challenge was that university websites are notoriously inconsistent some use modern JavaScript frameworks that traditional scrapers cannot read, others publish their catalogues as PDFs, and many have broken or missing links altogether. To address this, the system uses Crawl4AI, an open-source browser automation tool that operates like a real human browsing the web, rendering pages fully before extracting content. The pipeline automatically detects program listings, follows promising sub-pages, parses PDF catalogues, and recovers gracefully from failures. Each university entry is enriched with real institutional data including tuition costs, graduation rates, median graduate earnings, and Carnegie Classification.

What makes Edublox go beyond a simple directory is its AI comparison engine. Users can describe their personal preferences field of study, budget, location, career goals and receive an intelligent analysis.
This project sits at the intersection of information science, learning technologies, and applied AI and it raises a question worth asking: why should the data students need to make one of the most important decisions of their lives be this scattered and hard to find in the first place?

Author(s):  Ananya Agarwal, Bishnu Sarker

Depression is one of the most prevalent and debilitating mental health disorders worldwide, affecting emotional well-being and cognitive functioning, and the World Health Organization estimates that over 280 million people globally (around 5–6% of adults) suffer from depression, making it a leading cause of disability worldwide. Despite this widespread impact, more than 75% of affected individuals in low- and middle-income countries receive no treatment. We propose a depression detection framework in this study using a large-scale social media dataset collected from X (Twitter), Reddit, and Facebook, comprising approximately 244,388 posts, evenly balanced between depressive (label = 1) and non-depressive (label = 0) classes. To ensure data quality and semantic consistency, an extensive preprocessing pipeline was applied, including punctuation removal, stop-word elimination, and URL-link removal. For traditional machine learning models, textual features were extracted using TF-IDF vectorization, while deep learning models leveraged tokenized and padded sequences to capture contextual dependencies. Through careful preprocessing and developing hybrid deep learning architectures, we propose a CNN + logistic regression model that achieved an accuracy and an F1 score of 0.95, while a bidirectional CNN-LSTM captured the nuanced flow of emotional expression at 93% accuracy—both outperforming traditional approaches like linear SVM at 93.24%. Despite these promising results, several challenges remain, including dataset bias arising from platform-specific language, limited linguistic and cultural diversity, the dynamic and evolving nature of social media language, and the difficulty of distinguishing transient emotional expression from clinically significant depression. 

Author(s):  Kostiantyn Savchenko, Viktoriia Savchenko

This research explores the application of learning technologies in transforming B2B sales training through a structured, consultative approach. Traditional six-step B2C sales presentations have proven insufficient in engaging business partners, fostering long-term motivation, and driving active retail sales. In response, a 12-step B2B sales presentation model was developed, integrating two distinct consultative sales cycles (B2C and B2B), enhanced question-handling, and an embedded micro-training module using role-play and immediate feedback. This structure leverages principles of adult learning, performance-based training, and attention-driven content delivery to improve knowledge retention, engagement, and sales outcomes. Field experiments with companies such as 3M and Omega Bittner demonstrated significantly higher buy-in rates and measurable sales growth, even during market crises. The model has since been successfully adopted across multiple industries in Ukraine, illustrating the potential of learning technologies to reshape sales enablement strategies and improve business performance in B2B environments.

Author(s):  Dhiresh Venkateshwarlu Katuri, Sri Ramya Karuturi, Srinivas Kalyan Rajarapu

PolicyPulse.AI is a smart insurance assistant and monitoring platform designed to help users manage health insurance and vehicle insurance in a single application. Existing insurance systems are often fragmented, difficult to understand, and mostly reactive, making it hard for users to track policies, understand coverage, and make timely decisions. This project addresses that gap by developing an AI-powered prototype that combines Retrieval-Augmented Generation (RAG), BERT-based document understanding, pre-trained AI agents, and Python full-stack development to deliver a more proactive and user-friendly experience.

The system allows users to upload insurance documents, which are processed using BERT transformers for document breakdown, contextual understanding, and extraction of key information such as coverage details, exclusions, premium information, and claim-related terms. This processed content is then integrated into a RAG pipeline, enabling the application to provide accurate, context-aware answers to user queries based on the actual policy content rather than generic generated responses. To further enhance intelligence, pre-trained AI agents are used to perform specialized tasks such as policy explanation, recommendation generation, renewal reminders, and risk-aware guidance.

The prototype is developed using a Python full-stack approach, with Streamlit as the frontend interface for rapid development, interactive dashboards, and user-friendly visualization. The initial version focuses on core features such as document upload, policy summarization, RAG-based question answering, insurance monitoring, and personalized recommendations. By combining document intelligence, retrieval-based reasoning, and agent-driven support, PolicyPulse.AI aims to transform insurance management from a static process into an interactive, intelligent, and proactive digital experience.

Author(s):  Laxmi Shravani Mamidala, Shree Harshini Mamidala,  Haoxuan Zhang, Professor Dr. Ting Xiao

The rapid growth of scientific publications has placed significant pressure on traditional peer-review systems, making it difficult to maintain consistent, timely, and high-quality evaluations. Human reviewers often face challenges such as limited time, reviewer fatigue, and subjective biases, which can lead to variability in review quality. To address these issues, this paper introduces ProComp, a multi-agent large language model (LLM) framework designed to automate structured scholarly paper evaluation by emulating the roles of expert peer reviewers.

ProComp decomposes the review process into multiple specialized evaluation agents, each responsible for assessing a distinct aspect of a research paper. The framework employs three expert reviewer agents that independently analyze problem formulation and motivation, methodological soundness and technical rigor, and novelty and research contribution. Each agent produces structured assessments and quantitative scores aligned with common academic review criteria. To enhance contextual understanding and support informed evaluation, the system incorporates a Retrieval-Augmented Generation (RAG) component that retrieves related work and citation context from external scholarly databases.

A fourth agent acts as a comparative meta-reviewer, aggregating the outputs of the expert agents to generate a consolidated evaluation and final recommendation. The proposed framework is evaluated on a dataset of 194 research papers, comparing multiple LLM backends and analyzing their agreement with human review signals. Results demonstrate that the multi-agent architecture produces structured, interpretable evaluations and maintains meaningful correlation with human judgments. Ablation analysis further investigates the influence of individual reviewer agents on final scoring decisions.

Overall, ProComp demonstrates the potential of collaborative LLM agents to support scalable, transparent, and consistent automated scholarly evaluation, providing a promising direction for augmenting future academic peer-review systems.

Author(s):  Tayiba Raheem, Indira Priyaharshini Sundarrajan, Gahangir Hossain

Phishing remains a critical cybersecurity threat that bypasses traditional keyword-based detection through increasingly sophisticated social engineering. This study investigates whether Retrieval-Augmented Generation can improve phishing email classification by grounding a language model in dynamically retrieved threat intelligence. We constructed a dual-retriever pipeline combining semantic vector search and TF-IDF keyword matching to supply contextual examples to a Flan-T5 classifier. The system was evaluated across three detection approaches and tested for robustness against data poisoning attacks at varying injection rates. An interactive demonstration with confidence scores was developed to support K-12 cybersecurity literacy education, bridging the gap between AI-driven security  tools and student comprehension.​

Author(s):  Jahnavi Kovelamudi, Anuhya Koyyana, Syam Sai Konakalla, Kavya Sai Kotha, Whitworth Clifford  

Route optimization in navigation systems focuses on speed and distance rather than safety. Although current maps provide certain safety features, there are no accident risk predictions, which are based on historical data and current circumstances. There are 7.7 million historical accident data points in the US, covering the period 2016-2023, which creates a unique opportunity for safer routes by combining this data with real-time data.
 
We are developing the SafeRoute that combines historical patterns for the severity of accidents with live weather data by a route-based risk classification. The initial examination of 7.7 million US accidents found in the Moosavi dataset indicates some crucial trends. 47% of the accidents are due to traffic signals and junctions. The averages on the severity at nighttime are 5% higher than in the daytime. Weather conditions such as freezing rain and heavy snow are associated with high severity accidents.
 
Our methodology incorporates classification models based with feature aggregation to represent infrastructure characteristics and temporal patterns of accidents. Models like ensemble and gradient boosting are used to classify the route segments based on their severity. In addition, the model offers an understanding of each prediction representing the factors that are included in the prediction of the route segment, such as infrastructure type, weather conditions, and time of day.
 
By providing clear, explainable route safety assessments, SafeRoute enables informed driver decisions and supports evidence-based urban traffic management policies. This work demonstrates the practical value of integrating historical accident patterns with real-time data to address a critical safety gap in modern transportation systems.

Author(s):  Preethi Dasari, Sai Sahasra Katuri, Jayakrishna Pamuru, Priyanka Mada, Yuvasri Kumar

Silent myocardial infarction is clinically significant yet underdiagnosed condition that occurs without noticeable symptoms, often remaining undetected until later identified through electrocardiographic or imaging findings. These unnoticed events can lead to severe complications, including recurrent myocardial infarction, revascularisation, rehospitalization, and increased risk of sudden cardiac death. This highlights a critical gap in early cardiovascular risk detection and preventive care.

The present study aims to address this hidden clinical burden by developing an artificial intelligence AI -driven model for early prediction of silent myocardial infarction. The motivation is to identify high-risk individuals before the onset of major cardiovascular events, thereby enabling timely intervention and improved clinical outcomes. 

Our project focuses on detecting silent myocardial infarction (MI) using AI/ML models by analyzing demographic data such as age, gender, and height, along with ECG data. We utilized the PTB-XL dataset, which consists of 12-lead ECG records, and classified cases into normal, silent myocardial infarction, and acute myocardial infarction.Our model is a two-branch deep learning (DL) architecture developed using TensorFlow and Keras. One branch processes ECG signals using Convolutional Neural Networks (CNNs), while the other branch incorporates demographic inputs. The outputs from both branches are combined and passed through a softmax layer to classify cases into the three target categories.The model was trained and evaluated using performance metrics including accuracy, precision, recall, and F1-score.

For real-world applicability, we developed a tool that converts paper based ECG images into digital signals using OpenCV, NumPy, waveform databases. The images are de-skewed and divided into 12 leads, after which the waveforms are extracted, filtered, and resampled. These processed digital signals are then passed into the CNN model for prediction.

Author(s): Abhinay Peeta, Venkata Haritha Sai Sri Animilli & Dr Sayed K. Shah

Biometric authentication systems based on conventional modalities such as fingerprints or facial recognition are increasingly vulnerable to spoofing, environmental variability, and privacy concerns. In response, this work introduces a novel tongue-based biometric identification framework designed for enhanced security and robustness within biomedical and AI-driven authentication contexts. Leveraging the intrinsic morphological stability of the tongue, its protected intra-oral environment, and its inherent difficulty to forge or replicate, we propose this modality as a next-generation biometric trait for personal identification.

At the heart of our system lies a deep learning architecture: a convolutional neural network (CNN) that automatically extracts high-dimensional, discriminative features capturing fissure patterns, surface texture, and geometric tongue shape. To further boost performance and anti-spoofing resilience, we integrate a feature-level fusion strategy that combines these learned embeddings with complementary hand-crafted descriptors thereby forming a unified, highly discriminative feature vector. Experimental validation demonstrates that this hybrid approach significantly improves identification accuracy while reinforcing the modality’s robustness against adversarial or spoofing attempts. The results affirm the viability of the tongue as a reliable biometric modality and highlight the value of combining biomedical imaging with advanced AI techniques for secure authentication.

Author(s): Ananya Agarwal, Ampana Jayaram, Sravan Kumar Bhashyam, Harshith Harikrishnan, Kewei Sha, Cliff Whitworth

The evolution of autonomous driving is shifting toward cognitive intelligence through Vision-Language-Action (VLA) models that move beyond simple question-answering toward direct trajectory generation. This research presents a multi-platform framework using the Quanser QCar2 for real-time local LLM reasoning and the RobiX platform for large-scale 3D SLAM. By integrating these systems, we are developing a high-quality data pipeline for ""Small Expert Reasoner"" models designed for efficient in-vehicle hardware execution.
Our methodology incorporates cutting-edge architectures, including OpenDriveVLA for spatial grounding, AutoVLA for adaptive ""dual-mode"" reasoning, and DriveMLM for aligning linguistic decisions with behavioral planning states. Utilizing the 64K-clip LMDrive dataset, synthetic environmental stressors (rain, snow, fog), simulated sensor data, and VLM-generated clips, we bridge the ""execution gap"" between conceptual intent and low-level actuator control.
Crucially, this research systematically addresses four ""unsolved dilemmas"" currently preventing Level 5 autonomy. First, we tackle numerical and spatial reasoning deficits, specifically the waypoint ""coarseness"" inherent in discrete tokenizers. Second, we confront the temporal reasoning and alignment gap in modeling dynamic scene evolution, such as sequential turn signals. Third, we mitigate causal confusion and the interpretability-trust paradox, where models may generate plausible but incorrect linguistic hallucinations for their actions. Finally, we address the scarcity of rare edge case data by utilizing generative world models to simulate scenarios that deterministic rules cannot capture. By fusing deterministic sensor logic with the sophisticated reasoning of local LLMs, we aim to deliver an autonomous experience that is collaborative, intuitive, and physically grounded.

Author(s): Bishnu Sarker, Ananya Agarwal

Malignant blood disorders, including leukemias, lymphomas, and myelodysplastic syndromes, often manifest through subtle morphological abnormalities in peripheral blood that precede overt clinical symptoms, leading to delayed diagnosis and reliance on invasive bone marrow biopsy. Although peripheral blood smear microscopy is routinely performed, manual interpretation is time-consuming, subjective, and limited in its ability to detect early, low-prevalence malignant features. In this work-in-progress, we present a web application framework for identifying malignant blood disorder based on digitized peripheral blood smear images using deep learning. We have used transformer-based and convolutional neural network models that learn multi-scale representations of cellular morphology, spatial organization, and population-level patterns associated with malignancy. The long-term goal is to establish a scalable, objective, and interpretable image-based AI system for early detection and risk stratification of malignant blood disorders from routine peripheral blood smears.

Author(s): Gowtham Vuppaladhadiam, Baby Jahnavi Kovelamudi, Samuel Sibbi Rayan Jerome, Ana D. Cleveland, K S M Tozammel Hossain

Clinical chatbot assistants are increasingly deployed to improve patient outcomes and support informed healthcare decision-making, yet existing solutions fail to leverage the full spectrum of electronic health record (EHR) data types. Most existing systems are limited in scope primarily leveraging structured data or isolated text inputs and often lack multimodal integration and contextual reasoning. 

We present CareLens, a large language model (LLM)-based clinical chatbot that integrates structured and unstructured EHR data including clinical notes and medical images through a multimodal retrieval-augmented generation (RAG) framework.  While LLMs are powerful for reasoning tasks, they often lack factual accuracy, especially for numerical queries (e.g., retrieving basic statistics such as counts, averages, etc.). CareLens overcomes this limitation by utilizing an intent detection layer to classify queries as either factual- or reasoning-based. The system routes factual queries to a structured query engine to retrieve accurate numerical results and explains these results in clinical terms via LLM. For reasoning with unstructured text and images, a fine-tuned NLP model and domain-specific vision model are used.

CareLens offers a scalable and interpretable solution that synthesizes and presents factual and contextual insights from diverse EHR data types spanning free-text narratives, diagnostic codes, and imaging while supporting natural human interaction via speech and text. This chatbot aims to enhance clinician decision-making and accelerate clinical research by enabling seamless navigation of complex patient information.

Author(s): Komala Subramanyam Cherukuri, Kewei Sha, Zhenhua Huang

Object detection is crucial for Connected Autonomous Vehicles (CAVs) to perceive their surroundings and make safe driving decisions. Centralized training of object detection models often achieves promising accuracy, fast convergence, and simplified training process, but it falls short in scalability, adaptability, and privacy-preservation. Federated learning (FL), by contrast, enables collaborative, privacy-preserving, and continuous training across naturally distributed CAV fleets. However, deploying FL in real-world CAVs remains challenging due to the substantial computational demands of training and inference, coupled with highly diverse operating conditions. Practical deployment must address three critical factors: (i) heterogeneity from non-IID data distributions, (ii) constrained onboard computing hardware, and (iii) environmental variability such as lighting and weather, alongside systematic evaluation to ensure reliable performance. This work introduces the first holistic deployment-oriented evaluation of FL-based object detection in CAVs, integrating model performance, system-level resource profiling, and environmental robustness. Using state-of-the-art detectors, YOLOv5, YOLOv8, YOLOv11, and Deformable DETR, evaluated on the KITTI, BDD100K, and nuScenes datasets, we analyze trade-offs between detection accuracy, computational cost, and resource usage under diverse resolutions, batch sizes, weather and lighting conditions, and dynamic client participation, paving the way for robust FL deployment in CAVs.

Author(s): Umme Israt Afroz

Accurate biomedical question answering depends on linking responses to trustworthy scientific evidence. Large language models (LLMs), while effective at generating fluent and informative responses, they often produce hallucinations that reduce reliability in sensitive areas like clinical decision support and biomedical research. To address this challenge, we propose a framework for evidence-grounded biomedical Q&A using the BioMCP server, an open-source Model Context Protocol (MCP) implementation available on GitHub. The BioMCP server provides a stable interface for retrieving relevant PubMed articles, which are dynamically linked to each sentence of an LLM-generated answer. Our approach involves querying BioMCP with biomedical questions, retrieving candidate literature passages, and integrating these as verifiable references that either support or contradict the generated statements. By prioritizing contradictory evidence, the system encourages critical evaluation and mitigates overconfidence in LLM outputs. This work demonstrates how MCP-based integration of PubMed with LLMs can improve transparency, reliability, and trustworthiness in biomedical question answering.

Author(s): S M Saiful Islam Badhon, Praneeth Rikka, Ana Cleveland, and K S M Tozammel Hossain

Diabetic Retinopathy (DR) is a leading cause of preventable blindness worldwide, yet early diagnosis remains difficult due to scarce annotated data and limited access to ophthalmic expertise. AI-based screening systems using Convolutional Neural Networks (CNNs) show promise for DR prediction through retinal image analysis, but they require thousands of labeled images—an expensive, time-intensive demand unsuitable for resource-constrained settings. Semi-supervised learning (SSL) offers a solution by leveraging limited labeled data alongside abundant unlabeled images; however, existing SSL methods for DR detection suffer from unreliable pseudo-labels.

We propose a novel SSL approach that adapts Area Under Margin Self-Training (AUM-ST) from natural language processing to medical imaging for DR detection. We replace BERT with BEIT (Bidirectional Encoder Representation from Image Transformers) to effectively process retinal fundus images while retaining AUM-ST’s reliability assessment mechanism. To preserve lesion features essential for diagnosis, we incorporate saliency-preserving strategies that avoid aggressive augmentations such as random rotations and crops.

Experimental validation on a real-world dataset of 757 retinal fundus images demonstrates superior performance over baselines including ResNet-18, ResNet-50, DenseNet-121, MobileNet-v2, and EfficientNet-B0. Our method achieves notable gains in both accuracy and F1 score when training with fewer than 50 labeled samples per class, substantially outperforming supervised approaches that require extensive annotations for clinical-grade performance.

These results highlight the practical viability of our framework for scalable DR screening in resource-limited environments, including rural healthcare facilities and developing regions. More broadly, this work advances SSL in healthcare AI by addressing annotation costs and limited clinical expertise, paving the way for reliable, accessible medical imaging solutions.

Author(s): Fengjiao Tu, Ji Hyun Yu, Haihua Chen, Junhua Ding, Liu Dong, Chi-Jia Hsieh, Hannah Kim, Sunnie Lee Watson

This study develops and validates an interpretable transformer-based model for detecting cognitive presence (CP) in MOOC discussions to serve as a diagnostic foundation for instructional design and facilitation. Specifically, we built this model by fine-tuning the RoBERTa architecture with theory-derived structural cues from the Community of Inquiry framework to capture causal, synthesizing, and action-oriented language. Using data from six Social Work MOOCs (4,260 learners; 227,863 sentences), the model classified discourse into four CP phases and a fifth non-CP category. The cue-augmented RoBERTa achieved strong performance (macro F1 = 0.714; κ = 0.666), and SHAP analyses confirmed the model’s internal logic aligns with theory-consistent linguistic markers, ensuring the pedagogical transparency of the results. Learner-level hierarchical regressions revealed that while discussion quantity predicted behavioral persistence, higher-order CP indicators (Integration and Resolution) functioned as unique diagnostic proxies for motivation, satisfaction, and perceived knowledge transfer. These findings shift the role of CP analytics from descriptive observation to prescriptive action through a proposed Pedagogical Diagnostic & Engineering (PDE) framework. By mapping automated structural signals to specific instructional failure points, this framework offers concrete tactics for real-time scaffolding, such as Socratic prompts and practitioner spotlights, to move learners beyond exploratory monologues toward deeper knowledge construction and professional application.  

Author(s): Naaz Fatima Begum, Aishwarya Salla

Google research has recently conducted a study (Heydari et al., 2025) that underscores the future of personalized healthcare with the help of multi-agent architectures. Multi-agent systems assign tasks to specialized agents, including data science, clinical, and health coaching, to more accurately domain-knowledge decision-making, compared to single AI agents, which are usually constrained with large, complex datasets and do not have coordinating capabilities with other agents in the system. This solution helps to integrate different sources of data, such as electronic health records and wearables, to provide more individualized and proactive care. Building on this concept, we developed a blueprint for a mental health solution that extends the architecture by introducing additional agents, including a health equity agent and a resource allocation agent, aimed at addressing disparities in care and optimizing the distribution of mental health resources. Overall, this multi-agent framework offers a more scalable, equitable, and patient-centered alternative to traditional single-agent models.

The personal health agent is headed by an Orchestrator that coordinates specialized agents: Data Science Agent predicts based on the datasets: CDC records, EHRs, Social Vulnerability Index (SVI). Domain Expert Agent uses DSM-5 knowledge and evidence-based practice to provide insight. Health Coach Agent delivers personalized interventions. The Health Equity and Resource Allocation Agents collaborate to define disparities and populations at risk with the help of socioeconomic and access-related data. The resource allocation agent represents the most efficient option for distributing healthcare resources based on the needs and socioeconomic status of the patient and local services. This system integrates the input of user-reported symptoms, wearable data, hospital data, and socioeconomic factors to generate personalized recommendations. 

The model aims to offer individual and patient-centered solutions that meet personal needs and contribute to effective mental health management.

Author(s): Megan Barnes, Clelia Perez, Woorin Hwang, Yessi Ambarwati, Anany Sharma, Talar Terzian, Maryam Babaee, Lauren Eutsler, Andrea Salgado, Pasha Antonenko, D. Donihue, S. Bhunia, B. Chatterjee

The increasing use of artificial intelligence (AI) across all use contexts has created a need for AI curricula. The few available AI literacy frameworks focus on using and understanding large language model AIs, such as audiovisual generation and chatbot-based AI, but these are not the only use cases of AI. EdgeAI is more commonplace than teachers realize, and is less emphasized in the AI literacy literature. However, it has unique affordances for learning. 

This multi-stage design-based research project has undertaken both hardware and curriculum design. Phase one developed a custom learning board which includes a microprocessor, discrete sensors, and LCD panel output. This board is unique; design choices include utilizing five individual sensors in order to highlight the affordances and limitations of each, highlighting inputs versus outputs, and for durability in classroom environments. 

Implementation and feedback from the initial implementation at a museum in Florida lead to streamlining background knowledge teaching in Module 1 and expanding hands-on opportunities. Initial feedback from high school implementations of a condensed curriculum include providing more student-centered materials for classes that provide more self-paced learning, the addition of screen-free and gamified activities to emphasize AI concepts to lead into the final AI for Social Good design activity. 

Phase two is iterative curriculum design, implementation, teacher feedback, and revision. Multiple stages of teacher feedback will be continually solicited during the design and implementation phases in Texas classrooms during the 2025-2026 and 2026-2027 school years. The modules are: Introduction to IoT and Embedded Systems, Exploring the microcontroller without AI, AI Integration to the microcontroller - Edge Impulse, Challenging AI Hardware skills, and AHA! Social contribution. 

Author(s): Esther Eze

Small business owners are facing more cyber attacks, but they often can't get affordable training that fits their needs. Current cybersecurity education usually has the same content for everyone, which doesn't work well because people learn differently, have different skill levels, and threats change fast. This poster looks at research on a new AI e-learning platform that personalizes cybersecurity training for entrepreneurs. This system uses machine learning to watch how students perform and find their weak spots. Then, it changes how hard the content is and how it's taught. It also uses natural language processing to give helpful feedback. Data tracks how each person is doing and suggests extra lessons when needed. By customizing the experience and constantly changing to fit the student, the goal is to give training that is relevant and gets results for each entrepreneur. Early tests show that using AI to adapt the training helps people remember what they learned and keeps them interested compared to regular online courses. The platform can also be easily expanded to reach more people, giving entrepreneurial communities without resources access to good cybersecurity education. This study pushes forward information technology and AI by showing how adaptive learning can help close the gap between cybersecurity experts and small business owners. The project also points out chances for AI to be used in job training and to help colleges, businesses, and small businesses work together to improve cybersecurity.

Author(s): Tayiba Raheem, Gahangir Hossain

This study applies the SEIPS 2.0 framework to evaluate the socio- technical fit of an AI Coordination Agent designed for refugee pediatric Severe Acute Malnutrition (SAM) care. ​Using structured component mapping and scenario-based heuristic evaluation. We decompose the refugee SAM care work system into five SEIPS 2.0 components : People, Tasks, Tools, Organization, Environment . Then we assess how the Coordination Agent aligns with each. ​The evaluation reveals that AI-assisted coordination improves fit across People, Tasks, and Tools components. ​But introduces tensions in Organization and Environment dimensions, particularly infrastructure constraints and institutional readiness​. These findings yield actionable design requirements for AI systems in resource-constrained humanitarian settings, where socio-technical misalignment can compromise patient safety.​