Projects

Project 1 — AFRL: In-Sensor AI for Wearable Physiologic Monitoring

The A3I Hub participated in a SEMI and Air Force Research Laboratory (AFRL)-funded research program focused on developing ultra-low-power, in-sensor artificial intelligence for continuous physiologic monitoring. The central objective was to move machine-learning inference from conventional cloud-based platforms, e.g. Apple watch, directly into wearable and sensor-integrated hardware, enabling real-time detection of clinically relevant physiologic abnormalities from patient vital signs.

The research explored digital, analog, and bio-inspired neural computing architectures, including artificial neural networks (ANNs), reservoir computing, and echo-state networks implemented in CMOS. These approaches were designed to substantially reduce the energy and computational requirements of machine-learning inference, supporting continuous operation in resource-constrained wearable devices. Target applications included arrhythmia and sepsis detection from physiologic signals, ECG-based stress detection, cardiovascular disease prediction, and broader at-home health monitoring.

A major technical theme was demonstrating that clinically relevant machine-learning models could be implemented directly in low-power integrated circuits rather than requiring transmission of raw sensor data to an external processor. Representative work included real-time stress detection from wearable ECG using digital machine-learning circuitry; a bio-inspired reservoir computer for ECG analysis; a 7.5-nJ/inference CMOS echo-state network for coronary heart disease prediction; and integrated reservoir-computing hardware aimed at real-time, at-home patient monitoring. Parallel studies of analog neural-network classifiers established circuit architectures and custom activation functions for extremely energy-efficient embedded inference.

Collectively, the AFRL-supported work established an early foundation for the A3I Hub's wearable-sensor research: integrating sensing and AI computation at the edge to create low-power, autonomous systems capable of continuously interpreting physiologic data. This paradigm has potential advantages for wearable healthcare, including reduced communication bandwidth and power consumption, lower inference latency, greater independence from network connectivity, and improved privacy by minimizing transmission of raw physiologic signals.

Project 2 — NIH SCH: MUSE — Multimodal At-Home Monitoring for Acute Kidney Injury Survivors

The A3I Hub is currently leading in an NIH Smart and Connected Health (SCH)-funded project focused on developing MUSE (MUlti-modal SEnsor), an AI-enabled, non-invasive platform for monitoring patients at home following acute kidney injury (AKI). The project addresses an important gap in post-AKI care: survivors face substantial risks of recurrent AKI, progression to chronic kidney disease, cardiovascular complications, hospital readmission, and mortality, yet routine specialist follow-up is difficult to scale and can impose substantial burdens on patients and healthcare systems.

The central concept of MUSE is to replace reliance on frequent in-person assessment with patient-operated multimodal sensing coupled with AI-based risk prediction. The platform combines urinary biomarkers, wearable physiologic measurements, and patient clinical information to identify changes associated with AKI recurrence and alert patients when their estimated risk warrants medical attention. MUSE integrates four urinary biomarkers—urea, creatinine, lactate, and pH—with electrocardiogram (ECG) measurements and electronic medical record (EMR) information. Urinary measurements provide indicators of renal status, while ECG adds cardiovascular information relevant to post-AKI risk. Rather than treating each measurement independently, the system uses multimodal AI to integrate these heterogeneous data streams into a consolidated assessment of the patient's evolving health status.

A smartphone serves as the computational and user-interface hub for the system. Patients capture images of a urine dipstick-based sensor using the smartphone camera, with computer-vision algorithms quantifying the urinary biomarkers. ECG measurements are transferred to the phone using near-field communication (NFC). Patient medical information provides additional clinical context. Computer-vision and AI algorithms running on the smartphone then process and fuse the sensor and clinical data to estimate risk and generate alerts when potentially adverse changes are detected.

The proposed monitoring paradigm calls for patients to collect paired urine and ECG measurements at least twice daily, separated by approximately 6–8 hours, providing longitudinal measurements rather than isolated clinical snapshots. This repeated multimodal sampling creates an opportunity for AI models to detect evolving physiologic changes and potentially identify deterioration before it results in an emergency visit or rehospitalization.

From the perspective of the broader A3I wearable-sensor research program, MUSE represents an important progression from Project 1's in-sensor AI and low-power physiologic inference toward an integrated digital-health system combining wearable sensing, biochemical sensing, computer vision, clinical data, and multimodal AI. The project emphasizes translation into a patient-facing application: the technology is designed to be non-invasive, affordable, easy to operate, and usable without continuous involvement of a clinical expert.

Key innovation: MUSE establishes a framework for multimodal, AI-enabled remote patient monitoring, in which inexpensive biochemical sensing and wearable physiologic measurements are combined with existing clinical information to provide individualized longitudinal risk assessment. If successfully validated, this approach could help identify high-risk AKI survivors who require clinical intervention while reducing unnecessary visits for lower-risk patients, potentially improving post-AKI outcomes while making specialized follow-up more scalable.

Project 3 — Multimodal Wearable Biosensing for CAR-T Toxicity and Cardiotoxicity

The A3I Hub is leading a translational effort to develop and clinically validate a multimodal wearable biosensing platform for continuous monitoring of patients receiving CAR-T therapy, in partnership with Terasaki, Empaticaand Mayo Clinic clinical experts in oncology and cardiology. The project targets two major complications of CAR-T treatment: Cytokine Release Syndrome (CRS) and treatment-associated cardiotoxicity. 

CAR-T therapy has transformed the treatment of hematologic malignancies, but its potentially severe toxicities create a significant monitoring challenge. CRS can develop rapidly and unpredictably, progressing from early inflammatory changes to severe hemodynamic instability within hours. Current surveillance relies largely on intermittent vital signs, laboratory testing, and clinician assessment, making monitoring episodic and often reactive. In particular, inflammatory cytokines such as IL-6 and TNF-α are not continuously measured in routine clinical care, meaning that molecular evidence of immune activation may precede recognition of clinical deterioration. Consequently, interventions such as Tocilizumab may occur only after symptomatic CRS is established.

The proposed platform seeks to shift this paradigm from episodic clinical monitoring to continuous molecular and physiologic surveillance. Wearable and biosensor technologies will capture longitudinal inflammatory biomarker trajectories together with physiologic signals. These multimodal measurements will be integrated with relevant clinical information—including demographics, comorbidities, treatment characteristics, and medications—using AI-based risk models.

A principal research objective is to determine whether dynamic changes in inflammatory biomarkers and physiologic signals can provide an early-warning signature of CRS before overt clinical symptoms emerge. Rather than relying on individual threshold measurements, the AI framework will analyze temporal and multimodal patterns to generate individualized risk estimates. Demonstrating clinically meaningful lead time could enable earlier risk-adapted surveillance and more precisely timed therapeutic intervention.

A complementary objective is early characterization of CAR-T-associated cardiotoxicity. Cardiovascular complications may evolve before overt cardiac dysfunction becomes apparent through conventional clinical evaluation. Continuous longitudinal sensing provides an opportunity to identify subtle physiologic patterns associated with emerging cardiovascular injury and establish a foundation for predictive rather than retrospective cardiac risk assessment.

The translational program combines prospective clinical studies, wearable and biosensor deployment, real-time AI-driven risk modeling, and clinical validation. The Mayo Clinic A3I Hub brings together AI scientists with expert cardiology and oncology investigators, while Empatica contributes expertise in deployable wearable technologies. Beyond demonstrating algorithmic performance, the effort is designed to establish a scalable, production-ready monitoring system suitable for high-risk oncology care.

Key innovation: Project 3 advances the A3I wearable-sensor program from physiologic and biochemical monitoring toward continuous multimodal molecular + physiologic sensing coupled with real-time clinical AI. By integrating inflammatory biomarkers, wearable physiology, and clinical context, the project aims to provide actionable early warning of CRS while simultaneously establishing digital signatures of cardiotoxicity. The broader goal is to transform CAR-T toxicity management from reactive detection of established complications to proactive, individualized risk prediction and intervention.

Project 4 — PerfusionEdge: On-Chip AI for Transplant Organ 

In collaboration with transplant surgeon at Mayo Clinic, ASU and Terasaki institute, PerfusionEdge is an edge-AI platform for continuous, real-time assessment of organ viability during machine perfusion. Building on the team's dual-compartment biochemical monitoring (DCBM) system, which continuously measures glucose, lactate, and pH in perfusate and bile, and SCOPE, a self-supervised AI model for detecting abnormal graft physiology, the project moves AI inference from offline workstations directly into sensing hardware.

PerfusionEdge will deploy SCOPE on ultra-low-power analog in-memory computing chips, enabling nanojoule-scale inference without cloud connectivity. The system will be integrated with DCBM electronics and prospectively evaluated alongside clinical liver perfusion at Mayo Clinic Arizona. The collaboration combines Mayo Clinic's clinical perfusion and AI expertise with ASU's analog AI and system-on-chip design.

Key innovation: integrating continuous biochemical sensing with embedded AI to provide autonomous, real-time organ viability assessment at the point of perfusion, creating a reusable platform for transplantation and other applications requiring continuous physiologic intelligence.

Despite the expert-level performance of artificial intelligence (AI) models for various medical imaging tasks, real-world performance failures with disparate outputs for various minority subgroups limit the usefulness of AI in improving patients’ lives. AI has been shown to have a remarkable ability to detect protected attributes of age, sex, and race, while the same models demonstrate bias against historically underserved subgroups of age, sex, and race in disease diagnosis. Therefore, an AI model may take shortcut predictions from these correlations and subsequently generate an outcome that is biased toward certain subgroups even when protected attributes are not explicitly used as inputs into the model. We explore various types of bias from shortcut learning that may occur at different phases of AI model development and develop mitigation technique from preprocessing (data-centric solutions) and during model development (computational solutions) and postprocessing (recalibration of learning).

Domain adaptation – While the kidney segmentation is a required step for various clinical analyses, starting from disease detection, surgical planning, radiation dose optimization, to longitudinal characterization, the model development has largely focused on using a single-phase CT, resulting in significant performance degradation caused by drift in contrast phase appearance. We target three domain shifts to propose a robust kidney segmentation model: (i) contrast to non-contrast, (ii) arterial to venous phase, and (iii) normal functioning kidney to abnormal. Leveraging a traditional segmentation residual network (SegResNet) backbone, we design the domain adaptation for phase resiliency using an auxiliary contrast-phase classifier with confusion loss. 

Causal debiasing – Our project introduces MOSCARD, a novel predictive modeling framework that integrates chest X-rays (CXR) and 12-lead electrocardiograms (ECG) to enhance cardiovascular risk assessment. While CXR provides insights into chronic conditions contributing to MACE, ECG captures real-time cardiac electrical activity and structural abnormalities. By combining these two complementary modalities, MOSCARD delivers a more holistic and accurate risk profile than traditional models based on clinical scores, CT scans, or biomarkers. Unlike much of the existing literature, which often overlooks the role of confounders and underlying causal structures in multimodal data, our approach explicitly integrates causal reasoning by incorporating comorbidities and mitigates bias.

Unstructured medical data analysis and integration of multimodal data (image + EHR) can unlock the large amount of electronic healthcare records (EHR) for clinical event prediction (e.g. ER visits, hospitalization, short-term mortality). Opportunistic screening occurs when healthcare providers use a patient’s existing clinical interactions (e.g., routine visits, imaging for unrelated issues) to screen for conditions like cancer, rather than relying solely on scheduled screening programs. Our research interest is multimodal clinical data integration and predictive modeling to perform opportunistic screening for ASCVD and pancreatic cancer. For example, A 65-year-old patient undergoes a non-gated chest CT for evaluation of chronic cough. The scan is not intended for cardiac assessment, but it incidentally a deep learning model captures the calcium in coronary arteries —a marker of subclinical atherosclerosis, and generates a patient-specific report with CAC images and guideline-based recommendations using prior clinical history.

Interested in developing computational methods that can extract quantitative information from images, integrate diverse clinical and imaging data, enable discovery of image biomarkers, and improve clinical treatment decisions. I am leading several innovative medical image analysis research projects related to cancer diagnosis, e.g. prostate cancer aggressiveness detection, histopathologic subtype classification of brain tumor, prediction of semantic features of bone tumor. I am developing a novel computational framework that can automatically interpret implicit semantic content from multimodal and/or multiparametric radiology images to enable biomedical discovery and to guide physicians in personalized care. I am responsible for the overall design of the framework, and development, execution, verification, and validation of the systems.