Biomedical AI Research Focus

Research Area Leaders

Dr. Bruce Gao

Clemson University

Dr. Feng Luo

Clemson University

Dr. Qi Wang

University of South Carolina

AI is increasingly transforming the designing, manufacturing, and utilization of medical devices. Biomedical AI is an emerging interdisciplinary field where innovations exist not only in the development of new theories, models, and algorithms in AI but also in the synergistic integration of AI with targeted biomedical devices and applications in complex biomedical settings. Despite the remarkable progress made in imaging and language processing, cognition, and causal analysis, other challenges remain in biomedical AI, especially in its integration with biomedical devices. These challenges include:

  1. Data acquisition and AI-ready data preprocessing. Data related to biomedical devices are often generated from multiple sources with multiple scales, modalities, and sizes, which can be noisy, biased, incomplete, and unannotated. One of the challenges in the biomedical device domain is processing the data to make them AI-ready and trustworthy. New AI theories and more efficient algorithms for processing the source data and interrogating their trustworthiness are needed. 
  2. AI-device integration, dynamic learning, and predictive modeling. The accumulated static and longitudinal data that are generated and consumed by biomedical devices provide a wealth of information for tracking, monitoring, analyzing, and prognosticating for targeted tasks. The need to use longitudinal data to integrate AI seamlessly with physical as well as virtual devices to enhance their functionality is an overarching theme linking all ADAPT thrusts in this project. This requires innovation at the nexus of data-driven, physics-informed machine learning; time-dependent dynamics learning; physical models; and physical devices. 
  3. XAI for biomedical devices. To ensure the trustworthiness of the biomedical devices used in healthcare, the AI technologies implemented must be explainable to clinical stakeholders and users. One emerging field to accomplish this goal is XAI, which offers explainability of the AI models in trustworthy AI. 
  4. AI ethics and acceptance related to AI-enabled biomedical devices. Successful implementation of AI into biomedical devices requires that AI be trusted by clinical stakeholders and users. 
  5. Security and robustness of AI-enabled biomedical devices. These necessitate technical security, stability, proper training, and trustworthiness of AI. Developing AI technologies for biomedical devices requires innovation in making biomedical data AI-ready, niche deep learning models, seamless AI–device integration, trustworthy assessment, and availability of safe delivery systems.

The ADAPT in SC Project will conduct research to address the above challenges in the next five years.

First, we will develop new DL models and algorithms in data acquisition and preprocessing to create AI-ready datasets for the targeted biomedical devices. We will develop new representation learning algorithms to address small, large unannotated, noisy, high-dimensional, and multiscale/multimodal data issues relevant to our research. We will explore new data augmentation techniques such as generative adversarial network (GAN)-based methods to generate synthetic data, or leverage similar patient data via similarity analyses to address data deficiency. We will investigate novel approaches in meta-learning and few-shot learning with similarity-based networks to make them work more efficiently and reliably with small and sparse datasets. In addition to latent variable space approaches, we will advance models for multimodal data fusion and deal with unannotated data issues by integrating multi-modal representation learning, transfer learning (TL), or reinforcement learning with few-shot learning ideas. 

Second, we will develop new models and methods in physics-informed ML and predictive modeling. Many biomedical devices require seamless integration of AI-enabled units with physics models such as mechanical balance equations, energy transport equations, or biomarker- and morphology-evolution models (e.g., of solid tumors). We will innovate new models to couple AI units with physics models to achieve physics-informed ML. Predictive modeling is highly desirable in biomedical applications in diagnosis, treatment, and therapeutic rehabilitation. We will establish advanced predictive modeling frameworks by incorporating causal analysis, analysis on transfer-entropy, and representation learning with longitudinal data in neural dynamical systems, stochastic neural dynamical systems, and controlled dynamical systems.

Third, we will develop a new digital twin framework to integrate AI technologies with ADAPT-targeted biomedical devices. In some biomedical devices, more intimate coupling between AI and the device can be done to revolutionize the device’s performance. We propose to develop the DT framework to integrate AI technologies and patients’ multiscale and multimodal data to produce DT-enabled biomedical devices. The DT is an in-silico counterpart of the target that dynamically recapitulates the patient’s physical body or disease state, monitors the past, and predicts the future of the patient’s well-being and possible treatment outcomes. This framework will be innovated to guide the development of several devices.

Fourth, we will also explore AI models for cyber-attack detection, behavioral analytics and device security. The targeted biomedical devices are cyber-physical systems (CPS) consisting of edge or cyber devices, which are vulnerable to cyberattacks (jamming, spoofing, denial-of-service, malware injection, black hole, eavesdropping, Sybil attacks, sensor or communication failures). Detecting false readings in the devices is imperative yet challenging due to the systems’ dynamic behavior. To address the threat, statistical models, especially the change-point models, and abnormality-detecting AI models will be used for real-time anomaly detection. We will develop and evaluate multiple detection models, focusing on cyberattacks on the CPS. Using simulated data, we will develop robust attack detection methods based on classical hypothesis testing-based change point detection algorithms and Bayesian reliability-based models (e.g., a cumulative summation-based algorithm, a Bayesian online change point algorithm, and a hidden Markov models-expectation maximization algorithm).

Fifth, we will research trustworthy AI issues associated with the development of the targeted devices. We will be mindful of AI ethics, accessibility, disparity, and societal impact when designing AI-enabled biomedical devices. We will develop AI standards for biomedical device designs and protocols for training project participants and AI professionals engaging in biomedical AI research. We will produce AI-enabled biomedical devices to improve healthcare and accessibility, reduce disparities in access to healthcare technology in our communities, and bring about positive societal impact.

Clinicians, researchers, innovators, industry colleagues and others interested in learning more about ADAPT in SC and how they can participate, please contact Nadim Aziz at