CompHealth Corner: July
Computational and digital health ideas evolve quickly. Our monthly research roundup highlights the latest publications — from our Center faculty and other leaders in the field. Here’s what you should be reading this month!
Enhancing robustness in machine-learning-accelerated molecular dynamics: A multi-model nonparametric probabilistic approach
Quek, Ariana, Niuchang Ouyang, Hung-Min Lin, Olivier Delaire, and Johann Guilleminot. “Enhancing robustness in machine-learning-accelerated molecular dynamics: A multi-model nonparametric probabilistic approach.” Mechanics of Materials 202 (2025): 105237.
Summary
The authors, including Center member Johann Guilleminot, introduce a system-agnostic probabilistic framework to quantify model-form uncertainties in MD simulations driven by ML interatomic potentials.
Key highlights:
- A probabilistic framework for model-form uncertainty
- Representation of uncertainties from ML potential choice
- Use of a stochastic reduced-order model (SROM)
- Application to sodium thiophosphate for sodium-ion batteries
- Demonstrated capture of variability across ML potentials
Position: Retrieval-augmented systems can be dangerous medical communicators
Wong, Lionel, Ayman Ali, Raymond M. Xiong, Zejiang Shen, Yoon Kim, and Monica Agrawal. “Position: Retrieval-augmented systems can be dangerous medical communicators.” In Forty-second International Conference on Machine Learning Position Paper Track.
Summary
The authors, including Center member Monica Agrawal, analyze how current retrieval-augmented generation (RAG) systems answer medical queries, and argue that—even when every sentence is factually grounded—these systems can mislead patients by decontextualizing facts, omitting critical sources, and reinforcing misconceptions. They call for pragmatic reasoning about user intent and improved design to ensure safer
medical communication.
Key highlights:
- Large-scale evaluation of RAG outputs on disputed diagnoses and procedure safety
- Evidence of decontextualization: facts pulled from context lose critical nuance
- Identification of omitted sources that patients most need
- Analysis of how biased prompts reinforce patient misconceptions
- Recommendations to embed communication pragmatics and richer source comprehension
Building an open-source community to enhance autonomic nervous system signal analysis: DBDP-autonomic
Dunn, Jessilyn, Varun Mishra, Md Mobashir Hasan Shandhi, Hayoung Jeong, Natasha Yamane, Yuna Watanabe, Bill Chen, and Matthew S. Goodwin. “Building an open-source community to enhance autonomic nervous system signal analysis: DBDP-autonomic.” Frontiers in Digital Health 6 (2025): 1467424.
Summary
Wearables and smartphones can now capture rich autonomic nervous system (ANS) signals—but without robust, community-driven tools, much of that data remains untapped. Center member Jessilyn Dunn and others introduce DBDP-autonomic, which bridges the gap by offering:
- A unified, open-source framework for preprocessing and analyzing peripheral psychophysiological signals
- Transparency and reproducibility through community-guided development practices
- Scalability across devices and settings, so researchers can focus on discoveries instead of wrangling formats
Why read it?
You’ll learn how to turn raw ANS data into actionable insights, contribute to and benefit from a growing toolbox of analysis modules, and accelerate biobehavioral health research by making your work open, sharable, and future-proof.
Adaptive Physics Refinement for Anatomic Adhesive Dynamics Simulations
Martin, Aristotle, William Ladd, Runxin Wu, and Amanda Randles. “Adaptive Physics Refinement for Anatomic Adhesive Dynamics Simulations.” In International Conference on Computational Science, pp. 268-282. Cham: Springer Nature Switzerland, 2025.
Summary
Tracking how circulating tumor cells (CTCs) interact with vessel walls at the molecular level is essential for understanding cancer spread. But simulating these ligand-receptor dynamics across full vascular networks has been computationally out of reach—until now. Our team introduced a GPU-accelerated adaptive physics refinement (APR) method that dramatically reduces the cost of simulating CTC adhesion withoutcompromising the submicron precision needed to capture the physics of cancer cell behavior. Their approach features:
- A dynamic, high-resolution “window” that follows the cell and captures adhesive interactions in real time
- Hybrid CPU-GPU acceleration to scale across large microfluidic geometries
- Smart handling of data movement, memory, and algorithmic trade-offs to keep compute costs low
Why read it?
This work opens the door to high-fidelity simulations of cancer cell transport that were previously infeasible—letting you simulate complex adhesion events in anatomically realistic vessels. You’ll gain insight into GPU-accelerated strategies, learn how adaptive modeling can preserve critical biophysical details, and see how scalable methods like APR can push the boundaries of cancer mechanobiology research.
This paper was also awarded Best Paper at ICCS (https://www.iccs-meeting.org/iccs2025).
