Research

This initiative addresses vulnerabilities of learning systems to adversarial inputs and environmental distribution shifts. It advances concept-based learning that is robust both at training time and at run time, including under transformations like noise, rotations, or lighting changes. The research also explores adaptation in dynamic environments, and verification & monitoring of learning systems as they evolve, inspired in part by how children learn and generalize.

Project page:  https://aro-muri2020.seas.upenn.edu/index.html

AutoWean seeks safe, effective methods for autonomous weaning from mechanical ventilation — a process critical in ICUs that, if mismanaged, causes substantial morbidity, mortality, and healthcare costs. The project develops data-driven, context-aware, human-in-the-loop control strategies that coordinate patient monitors and ventilators. It addresses challenges in identifying patient state, adapting control policies, and ensuring safety and reliability in the presence of uncertainty.

This project investigates remote sensing and analytical methods to promote safe pain management with opioids. It aims to monitor both safety and adherence, using non-invasive sensing technologies and data analysis to detect misuse, safety risks, or deviations from prescribed therapy. The goal is to provide insight into patient behavior and support interventions while preserving patient privacy and trust.

Smart Alarms 2.0 addresses clinical alarm fatigue by suppressing non-informative alarms in both hospital and home settings, while keeping caregivers (nurses, physicians, parents) in the loop. The project develops novel design and analysis methods for robust alarm suppression and models caregiver interaction. Use cases include pediatric wards and home monitoring for patients with Bronchopulmonary Dysplasia (BPD). The objective is to optimize alarm settings so that caregivers receive the right prompts at the right times, improving responsiveness without overload.

This project addresses the global burden of glaucoma-related vision loss by developing an artificial intelligence–based clinical decision support tool for early diagnosis and prediction of disease progression.

A multi-modal deep learning framework integrates diverse longitudinal data, including retinal nerve fiber layer (RNFL) thickness from OCT, visual field assessments, and fundus photography, to enhance diagnostic precision and anticipate functional loss.

The research emphasizes trustworthy human-AI interfaces, providing interpretable outputs, transparent performance metrics, and robust uncertainty estimation to help clinicians make informed decisions. In addition, algorithms are designed to handle data heterogeneity, ensuring reliable performance across patient populations, imaging devices, and care settings, supporting equitable adoption in real-world clinics.

This research area explores how sensors, actuators, and networked devices can be combined into safety-critical, closed-loop systems for patient care. By integrating continuous-time physiological models with formal verification techniques (such as timed automata), the work ensures safe behavior even under network delays or failures.

Case studies include:

• A patient-controlled analgesia (PCA) infusion controller that prevents overdosing through real-time monitoring and automatic dose adjustments.
• Simulation and evaluation of hospital insulin infusion protocols using virtual patient populations to assess reliability and safety before deployment in clinical environments.

Modern healthcare increasingly depends on interoperable medical devices, where failures can have complex, distributed causes. This project develops causality analysis algorithms that process execution traces and component contracts to explain how a safety violation occurred.

The approach identifies which components or interactions contributed to an incident, even when direct reproduction in a clinical setting is impractical. Results provide actionable insights that support accountability, guide debugging, and inform system design improvements.

The MD PnP program is a multi-institutional effort to establish interoperability standards and tools for medical devices, enabling safer, more efficient clinical environments.

Contributions include:

• Participation in the development of the Integrated Clinical Environment (ICE) standard.
• Building interoperable prototypes, such as systems that synchronize ventilators with anesthesia machines.
• Creating safety interlocks for PCA pumps and other infusion devices.
• Conducting studies on regulatory pathways, usability, and workflow integration, demonstrating how interoperability can reduce errors and improve patient outcomes.

LogSafe is a secure, fault-tolerant data logging platform designed for large-scale Internet-of-Things deployments in untrusted cloud environments.

The system provides:

• Confidentiality and integrity for collected sensor data through lightweight cryptographic mechanisms.
• Tamper detection to ensure data authenticity and reliability.
• A scalable architecture that supports analytics and monitoring in environments where device or network compromise is a concern. LogSafe offers a trustworthy foundation for IoT systems that require strong data security guarantees.