Engineering-Driven Learning Approaches for Bio-Manufacturing and Personalized Medicine

By Chen, Jialei

Georgia Institute of Technology

Download (PDF)

Licensed according to this deed.

Published on


Advisors: Chuck Zhang, Roshan Joseph, Arman Sabbaghi, Jianjun Shi, Ben Wang, Jeff Wu

Healthcare problems have tremendous impact on human life. The past two decades have witnessed various biomedical research advances and clinical therapeutic effectiveness, including minimally invasive surgery, regenerative medicine, and immune therapy. However, the development of new treatment methods relies heavily on heuristic approaches and the experience of well-trained healthcare professionals. Therefore, it is often hindered by patient-specific genotypes and phenotypes, operator-dependent post-surgical outcomes, and exorbitant cost. Towards clinically effective and in-expensive treatments, this thesis develops analytics-based methodologies that integrate statistics, machine learning, and advanced manufacturing. Chapter 1 of my thesis introduces a novel function-on-function surrogate model with application to tissue-mimicking of 3D-printed medical prototypes. Using synthetic metamaterials to mimic biological tissue, 3D-printed medical prototypes are becoming increasingly important in improving surgery success rates. Here, the objective is to model mechanical response curves via functional metamaterial structures, and then conduct a tissue-mimicking optimization to find the best metamaterial structure. The proposed function-on-function surrogate model utilizes a Gaussian process for efficient emulation and optimization. For functional inputs, we propose a spectral-distance correlation function, which captures important spectral differences between two functional inputs. Dependencies for functional outputs are then modeled via a co-kriging framework. We further adopt shrinkage priors to learn and incorporate important physics. Finally, we demonstrate the effectiveness of the proposed emulator in a real-world study on heart surgery. Chapter 2 proposes an adaptive design method for experimentation under response censoring, often encountered in biomedical experiments. Censoring would result in a significant loss of information, and thereby a poor predictive model over an input domain. For such problems, experimental design is paramount for maximizing predictive power with a limited budget for expensive experimental runs. We propose an integrated censored mean-squared error (ICMSE) design method, which first estimates the posterior probability of a new observation being censored and then adaptively chooses design points that minimize predictive uncertainty under censoring. Adopting a Gaussian process model with product correlation functions, our ICMSE criterion has an easy-to-evaluate expression for efficient design optimization. We demonstrate the effectiveness of the ICMSE method in an application of medical device testing. Chapter 3 develops an active image synthesis method for efficient labeling (AISEL) to improve the learning performance in healthcare and medicine tasks. This is because the limited availability of data and the high costs of data collection are the key challenges when applying deep neural networks to healthcare applications. Our AISEL can generate a complementary dataset, with labels actively acquired to incorporate underlying physical knowledge at hand. AISEL framework first leverages a bidirectional generative invertible network (GIN) to extract interpretable features from training images and generate physically meaningful virtual ones. It then efficiently samples virtual images to exploit uncertain regions and explore the entire image space. We demonstrate the effectiveness of AISEL on a heart surgery study, where it lowers the labeling cost by 90% while achieving a 15% improvement in prediction accuracy. Chapter 4 presents a calibration-free statistical framework for the promising chimeric antigen receptor T cell therapy in fighting cancers. The objective is to effectively recover critical quality attributes under the intrinsic patient-to-patient variability, and therefore lower the cost of cell therapy. Our calibration-free approach models the patient-to-patient variability via a patient-specific calibration parameter. We adopt multiple biosensors to construct a patient-invariance statistic and alleviate the effect of the calibration parameter. Using the patient-invariance statistic, we can then recover the critical quality attribute during cell culture, free from the calibration parameter. In a T cell therapy study, our method effectively recovers viable cell concentration for cell culture monitoring and scale-up.

Cite this work

Researchers should cite this work as follows:

  • Chen, Jialei (2021), "Engineering-Driven Learning Approaches for Bio-Manufacturing and Personalized Medicine,"

    BibTex | EndNote


MATIN Development Team