Radiotherapy is a well-established technique for treating durable malignant cells. In breast radiotherapy, regions of the breast containing cancerous cells are exposed to x-rays to shrink and kill tumors. However, this treatment method remains unsatisfactory due to crude setups and poor localization techniques that prevent effective normal organ sparing. Overlapping and nearby healthy cells may be unintentionally damaged by radiotherapy in addition to the targeted cancer cells, which results in life-threatening acute and chronic toxicities in breast cancer patients after treatment. To control healthy organ sparing and provide a reproducible setup, this work experimentally develops a pneumatically actuated soft robot to safely isolate the breast from other organs for imaging and treatment using Ecoflex, a silicone elastomer with a low Young’s Modulus. We pneumatically actuate the soft robot by pumping air into a network of air channels embedded within the robot’s body, causing specific sections of walls to expand and press against the breast. This expansion fixes the breast in a treatable position as far away from the rest of the patient’s body as possible. Upon actuation, the thickness of the inner wall pressing against the breast is less than 250µm, which minimizes interference with imaging and unwanted radiation exposure. Each device costs less than 5 USD to manufacture, so it is practical to custom-fit the robot to each patient and dispose of it after treatment. This work demonstrates a promising future for soft robots in medical applications due to their lightweight, adaptable, reproducible, and inexpensive features.

Protein-Protein Interaction (PPI) network methods are an increasingly popular way to predict drug downstream effects. For example, PathFX is a novel algorithm that uses PPI network methods to identify drug pathway associations and drug-related phenotypes. However, these algorithms often predict more drug effects than evidence supports. These predictions can be tested by conducting observational studies in the Electronic Health Record (EHR). However, instead of testing each individual drug-disease prediction in the EHR, it is more practical to test groups of drugs based on shared gene pathways. This study will focus on the specific disease areas of diabetes and lung cancer to illustrate how PathFX can be used to analyze drug-gene and drug-disease associations to identify hypotheses for shared drug-gene pathways. We analyzed PathFX networks for drugs used to treat diabetes and lung cancer. We analyzed the frequency of shared genes and shared phenotypes, and used downstream proteins to cluster treatment drugs. We identified 44 and 34 drugs for diabetes and lung cancer respectively, and found drug network clusters are distinct from ATC groups. We used GO enrichment to discover functions associated with network clusters and found that diabetes and lung cancer pathways had distinct functional categories. We hypothesize that we will be able to distinguish clinical and non-clinical drugs by their downstream pathways and provide a means to reduce PathFX over-prediction. We will later use observational studies in the EHR to test the utility of network-identified clusters and expand this analysis to other disease areas.

A communication system, which consists of a transmitter and a receiver, models the process by which information is sent and received. The transmitted symbols that are generated by a transmitter go through a noisy channel and reach the receiver end. The receiver needs to estimate the transmitted symbols by their noisy version. Claude Shannon developed a theory that determines the maximum rate at which the receiver can reliably estimate the transmitted symbols based on the noise’s statistics. To achieve the maximum rate, the transmitted signals need to approximately follow an optimal probability distribution, which can be done through probabilistic shaping. One method for probabilistic shaping is using a distribution matcher that takes a sequence of bits equally likely to be ones and zeros and maps it bijectively to a new sequence of symbols with the desired probability distribution.There are two types of distribution matchers denoted as constant and multi-composition distribution matchers or CCDMs and MCDMs. We coded a CCDM and a MCDM, which is a union of CCDMs. Two different versions of the MCDM based on a high probability and typical set rule were constructed. We found that MCDMs outperformed CCDMs in both normalized Kullback–Leibler (KL) divergence, a measure of how well the desired distribution is met, and matching rate, meaning we can send more information using less bits. By applying MCDMs to channels, we can achieve higher transmission rates and better noise correction to increase the efficiency and speed of the internet and communication systems around the world.

Transportation infrastructures such as roadways in urban areas contribute to pollution via contaminated stormwater runoff. Implementing soil-based stormwater infrastructure such as biofilters could capture and treat the contaminated runoff. However, required compaction for roadside soil limits biofilter’s infiltration and treatment capacity. The addition of bulking agents such as sand or large aggregates such as expanded shale, clay, and slate (ESCS) can increase the infiltration capacity. However, the quantity of the bulking agent required to achieve the desired infiltration rate is unknown. To estimate the optimal amount of bulking agent, we mixed the soil with two bulking agents, sand (0.6 - 0.85 mm), and expanded shale, clay, and slate (ESCS, < 2.8 mm) with different mixing ratios. Further, we amended the soil-bulking agent mixture with biochar to enhance the contaminant removal performance. The result shows that the biofilter media mixture with 50% (v/v) bulking agents, 25% soil, and 25% biochar, meet the required infiltration rate of 1-5 inch.h-1. Under compaction, ESCS-based media exhibited a 3 times higher infiltration rate than sand. While both compacted biofilter media effectively remove E. coli, biofilter amended with sand showed relatively higher removal than ESCS-based media owing to higher straining in sand amended biofilters. The results would help develop design guidelines for roadside stormwater treatment systems that require the compaction of filter media.

An essential tool to furthering dynamic spectrum sharing, which allocates spectrum based on user demand, is knowing the signal strength on particular frequencies and at locations, allowing optimized base station placement for efficient use of the spectrum between many devices in an area. Given the location of existing fixed transmitters and the locations and signal strength of their respective receivers, our goal is to find the signal strength at any point in an area of interest due to a transmitter at any location, but without any active transmission from the transmitter. The path loss model, employing a least squares linear regression, is a traditional method for this problem; however, finding the signal strength in an urban area, because of building obstacles, has a complex pattern of loss, so we attempt to predict that strength through a deep learning model. Specifically, we use a 2D format to feed in data, ideally giving the model spatial context of multiple receivers’ signal strengths at once. We use the U-Net architecture, which is a type of convolutional neural network with image-to-image translation: the input is two matrices, one representing the transmitter location and the other representing all the receivers’ locations, and the output is an image of predicted signal strengths at the receiver locations used in the input. Using simulation-based evaluation, we find that, on using a wide range of available transmitters and receivers to train the model, this method does achieve a more accurate prediction of signal strength than the path loss model.