Poster Abstracts

My abstract is: Human mesenchymal stem cells (hMSCs), while useful in many therapeutic treatments, often struggle with producing consistent results in clinical trials. This is due to the inherent heterogeneity of hMSCs. To address this, many laboratories have used existing tools to sort hMSCs based on surface marker expression. Many of these techniques rely on fluorescent staining of biological markers, underscoring the need for label-free approaches that preserve the cells’ native state by leveraging intrinsic properties for sorting Dielectrophoresis (DEP) is a label-free technique that sorts cells based on their inherent electrical properties when exposed to non-uniform electric fields, with cell behavior modulated by voltage and frequency. In this study, we used an insulating DEP microfluidic device containing a large region of insulating posts to generate spatially non-uniform electric fields. Under specific voltage-frequency combinations, a subset of cells became trapped at these posts while others continue to flow through the device. By tuning these parameters, we enriched subpopulations of adipose tissue derived hMSCs that exhibited distinct DEP responses. Following sorting, the enriched cell populations underwent a 14-day adipogenic differentiation protocol. Cells that flowed through the device (remained untrapped) under low-voltage, low-frequency conditions demonstrated enhanced adipogenic potential compared to unsorted controls. These findings demonstrate that DEP can enrich for functionally distinct hMSC subpopulations, offering a promising tool to address cellular heterogeneity in regenerative therapies.

Stress is a pervasive factor influencing mental health and overall well-being, with prolonged exposure linked to adverse physical, psychological, and social outcomes. Recent advances in mobile health technologies and wearable sensing have enabled the continuous collection of physiological and behavioral data, creating new opportunities for early stress detection and intervention. In parallel, large language models (LLMs) have demonstrated strong reasoning capabilities across diverse domains, including health prediction from time-series data. However, most existing approaches rely on cloud-based models, raising concerns around privacy, latency, and real-world deployability. This work investigates the use of on-device language models (ODLMs) for stress prediction, emphasizing their potential to support health and wellbeing through privacy-preserving, low-latency inference directly on personal devices. 

We systematically evaluate how different data modalities, prompting strategies, and model scales affect stress prediction performance in ODLMs. Using the PMData life-logging dataset, which includes both objective physiological signals (e.g., steps, heart rate, sleep duration) and subjective self-reports (e.g., mood, fatigue, sleep quality), we design prompts that contextualize health data across multiple temporal intervals. Experiments are conducted using several state-of-the-art quantized ODLMs deployed on iOS devices, allowing us to jointly assess prediction accuracy and device-level performance metrics such as latency, throughput, and memory usage.

Our results show that compact ODLMs can achieve competitive stress prediction accuracy while operating entirely offline. Across multiple experimental settings, objective data prompts consistently yield a comparative prediction error to subjective or combined prompts, suggesting that passive sensing can effectively support stress assessment with minimal user burden. Among the evaluated models, a sub-billion-parameter ODLM demonstrates the best balance between accuracy and efficiency, outperforming larger models in both prediction error and inference speed. Additionally, we find that statistical summaries of longer time windows improve performance relative to raw natural language descriptions, highlighting the importance of prompt structure when working with time-series health data. Beyond model performance, we propose a proof-of-concept trust architecture that situates ODLMs within a closed social and clinical loop involving patients, caregivers, and clinicians. Rather than replacing human judgment, the system is designed to augment care by enabling timely insights into stress while maintaining safeguards against harmful or inappropriate outputs. This framing aligns with a well-being-centered approach to AI deployment, prioritizing user safety, autonomy, and clinical oversight.

Overall, this study demonstrates that on-device language models can serve as a practical and responsible component of mobile health systems for stress prediction. By combining wearable data, thoughtful prompt engineering, and efficient on-device inference, ODLMs offer a promising pathway toward scalable, privacy-preserving tools that support mental health monitoring and proactive wellbeing interventions in everyday life