Spring 2026 Data Science Poster Fair - Poster Details

Advisor: Feng Xie, Division of Computational Health Sciences

Abstract: Emergency departments are responsible for quickly and accurately identifying patients who require hospitalization or are at risk of critical outcomes. Machine learning models trained on electronic health record (EHR) data can be leveraged to improve early prediction of these outcomes in emergency department (ED) patients. These models can support clinical decisions that are made during key timepoints, such as ED triage and disposition, enabling more efficient resource allocation and improved care for patients.

Electronic health record data includes multiple modalities, including structured clinical data, clinical notes, and diagnostic ECG data. Each data modality captures different aspects of patient health. However, many existing predictive models rely on a single data modality, failing to fully leverage complementary information available across EHR data sources. This study evaluates whether combining multiple EHR data modalities within a multimodal model improves prediction of clinical outcomes for ED patients.

The MIMIC-IV, MIMIC-IV-ED, MIMIC-IV-Note, and MIMIC-IV-ECG datasets were used to develop predictive models for four binary classification tasks: (1) hospitalization prediction at ED triage, (2) critical outcome prediction at ED triage, (3) hospitalization prediction at ED disposition, and (4) critical outcome prediction at ED disposition.

For each prediction task, three models were trained on different EHR data modalities: structured clinical data, clinical notes (diagnostic and radiology), and electrocardiogram (ECG) data. Structured data was modeled using a gradient boosting classifier, clinical notes using a transformer-based natural language processing (NLP) model, and ECG data using a fine-tuned HuBERT-ECG transformer model for binary classification. The performance of each single-modality model was evaluated independently, and the predicted probabilities from each individual model were combined to form a multimodal prediction model.

Among the single-modality models, the structured data model achieved the highest AUROC performance, followed by the clinical notes and ECG models. The multimodal model outperformed the individual modality models with AUROC values of 0.80 for hospitalization prediction at ED triage, 0.87 for critical outcome prediction at ED triage, 0.998 for hospitalization prediction at ED disposition, and 0.92 for critical outcome prediction at ED disposition. These results suggest that multimodal approaches capture complementary clinical information from varying EHR data sources and achieve higher performance than single-modality models. Furthermore, these findings indicate that adjusting the prediction timepoint from ED triage to ED disposition improves model performance due to increased availability of relevant data during the ED visit.

This study demonstrates the potential of multimodal machine learning models to improve clinical outcome prediction in emergency departments and highlights the value of integrating structured data, clinical notes, and ECG data within a multimodal model to support informed clinical decision making.

Advisor: Tianxi Li, School of Statistics

Abstract: This project examines how large language model (LLM) agents form and update social perceptions in a multi-agent interaction setting. Agents from multiple model families (Gemini, GPT, Claude, and DeepSeek) participate in repeated group discussions across diverse topics and evaluate one another on dimensions such as trust, knowledge, influence, agreement, confidence, and relevance. The study focuses on how these evaluations evolve over time, including the stability of initial impressions, the extent of changes across interactions, and relationships between evaluation metrics. Initial results show that agents frequently revise their evaluations, but initial impressions remain influential, indicating anchoring effects. While models exhibit similar patterns in how evaluation metrics relate, they differ in how strongly these dimensions are connected. Network analysis shows that influence and trust are unevenly distributed, with certain agents consistently emerging as central nodes, while others remain peripheral. Difference in the perception of Agreement versus Influence is evident across all models.

The project also produces a dataset of multi-agent interactions across selected diverse controversial topics, for future analysis of conversation dynamics.

Advisor: Jaideep Srivastava, Department of Computer Science & Engineering

Abstract: Machine learning models deployed in dynamic real-world environments are prone to performance degradation as data distributions, usage patterns, and operational conditions evolve over time. Existing monitoring approaches primarily focus on detecting individual drift signals or performance decline, but often provide limited support for diagnosing the underlying causes of model degradation across multiple tasks and modalities. In this work, we present a unified post-deployment monitoring framework that integrates complementary analyses including performance monitoring, data and prediction drift analysis, data quality assessment, and model explanations within a single model-agnostic pipeline. Rather than presenting isolated alerts, the framework takes heterogeneous monitoring signals into interpretable diagnostic summaries to support root-cause analysis and informed intervention decisions. Preliminary experiments on machine learning tasks across structured data, text, and imaging demonstrate that combining multiple monitoring signals improves the ability to localize degradation sources compared to individual analyses. This work represents an initial step toward systematic, multimodal diagnostic reasoning for reliable machine learning systems operating in evolving production environments.

Advisor: Erich Kummerfeld, Institute for Health Informatics

Abstract: Causal graphs are widely used to represent possible cause-and-effect relationships, but tools for generating and analyzing these graphs often produce outputs in different formats, making them difficult to compare, explore, and present consistently. This creates a practical barrier for researchers and students who want clear, manuscript-ready visualizations without extensive manual reformatting.

This project presents a lightweight, web-based interactive visualization tool that brings causal graphs from multiple ecosystems, including Tetrad, causal-learn, DoWhy, and Dagitty, into a single consistent interface. The system supports importing, viewing, and interacting with graphs while reducing the burden of handling tool-specific format differences.

The tool enables interactive exploration and standardized visualization of causal graphs, supporting clearer interpretation and more consistent communication of results. By improving usability and visual coherence across workflows, this work lowers the barrier to comparing and presenting causal discovery outputs in research and educational settings.

Advisor: Seongjin Choi, Department of Civil, Environmental, and Geo- Engineering

Abstract: SPARC is an end-to-end deep learning–based approach for 3D reconstruction of traffic intersections from a sparse multi-camera setup. Accurate 3D reconstruction of traffic intersections is critical for enabling safer autonomous navigation, improving traffic modeling, and supporting the development of intelligent transportation systems. Traditional multi-view reconstruction approaches are often limited to 2D bird’s-eye-view representations or semantic predictions, while existing methods for full 3D RGB reconstruction typically require a large number of diverse viewpoints. To address these limitations, the SPARC pipeline combines monocular depth estimation with pairwise correspondence and point cloud optimization to generate accurate 3D RGB representations from as few as four camera views. The method is evaluated on a custom dataset generated using the CARLA simulator and demonstrates improved reconstruction fidelity compared to baseline approaches.

Advisor: Yicheng Song, Information & Decision Sciences

Abstract: Sequential search models in economics describe how consumers make rational, step-by-step decisions when searching for products. These models typically assume that product attributes are readily observable. However, in many real-world settings, consumers infer product characteristics from unstructured data such as review text. This study integrates feature learning from unstructured data into the sequential search framework by developing a theory-driven deep learning model. The model leverages clickstream data from consumers’ search processes and associated review texts to jointly recover latent product features and consumer preferences. The analysis is conducted in a controlled environment where the data-generating process is known. Clickstream data are simulated to represent consumers searching for restaurants, where decisions are driven by preferences over latent attributes such as food, service, value, and convenience. The proposed model enforces consistency with the economic structure of sequential search, enabling recovery of the underlying decision process. The results show that the model accurately recovers latent features and preference structures, while providing interpretable mappings between textual content and economically meaningful attributes. This approach offers a unified framework for combining unstructured data with economic decision models.

Advisor: Jaideep Srivastava, Department of Computer Science & Engineering

Abstract: Smart wearable devices drain battery rapidly by running all onboard sensors continuously, but not every sensor is needed to recognise every human activity. This study develops a methodology for activity-aware sensor selection that identifies the minimal sensor subset per activity, quantified through SHAP, permutation importance, and ablation analysis across four benchmark HAR datasets covering 74 subjects and over 40 activities. The complete accuracy-energy trade-off is characterised to enable practical long-duration wearable monitoring without meaningful loss in recognition performance.

Advisor: Jaideep Srivastava, Department of Computer Science & Engineering

Abstract: Small and mid-sized banks often lack access to sophisticated forward-looking financial health assessments, leaving regulators and analysts reliant on lagging indicators that fail to capture emerging risks early enough. To address this, we developed an end-to-end machine learning pipeline using LightGBM with two-stage gain-driven feature selection to forecast six key performance metrics (AER, ROA, ROE, ROTCE, TCE, and T1LR) six quarters ahead, benchmarked against an LSTM model and supported by a rigorous data validation framework. The system will produce Fitch-style composite grades and GPA-based bank rankings, delivering interpretable, forward-looking performance assessments to support early financial health monitoring across the banking sector.

Advisor: Jane Sumner, Department of Political Science

Abstract: While self-refinement has become an increasingly researched approach to improving responses from LLMs for objectively-defined tasks such as solving math problems or generating code, its effectiveness in subjective applications has been less explored. Objectivity is a convenient goal for evaluating a refinement strategy, but many general-use LLMs are tasked to generate and revise qualitative deliverables such as cover letters, ad copy, or cold emails, where the exact degree of a response’s persuadability varies by audience.

This project investigates iterative direct preference optimization (DPO) as a strategy for fine-tuning LLMs for self-refinement. A framework is developed where the model generates multiple drafts of rhetorical text, which are then fed back in as preference data to fine-tune the model. To investigate the validity of this approach, a comparative study is pursued where the model’s refined drafts are qualitatively evaluated against human-edited benchmarks. 

These results aim to determine the boundaries of self-refinement in subjective tasks, and provide an additional strategy for aligning LLMs with nuanced human preferences when an objective ground truth is not available.

Advisor: Ju Sun, Department of Computer Science & Engineering