In recent years, Electronic Health Records (EHRs) have increasingly been incorporated in medical practices and hospitals. EHRs contain a multitude of data, including patient medical histories, medications, treatments, and, in inpatient settings such as the intensive care unit (ICU), physiological time-series and lab test results. These many different types of data tend to be too numerous to interpret easily. Visualizations can enable clinicians and researchers to gain a better understanding of the different types of available data and how they fit together.
Our visualization integrates patient timelines in the ICU (including vital signs, lab results, and intervention administration) with rolling risk estimates from a predictive model for adverse outcomes. We demonstrate our visualization tool on the Medical Information Mart for Intensive Care (MIMIC) III, an openly available database composed of ICU EHR data from the Beth Israel Deaconess hospital in Boston, Massachusetts. Our tool enables researchers to search for patient cohorts fitting specific criteria and to explore the rich data associated with their hospital stays. Our patient-centered visualization enables researchers to investigate how their predictive model risk estimates correlate with physiological changes and interventions while in the ICU.
In the field of artificial intelligence, most learning mechanisms are either inherently spe- cialized or have a flat knowledge base that makes them only useful when confined to a single domain, reducing the capability of both handling complex problems and specializing in multiple domains. This project is to design a contextual knowledge system and that easily fits a knowledge abstraction for use by various learning mechanisms. The resulting system provides automatic compositionality for learning mechanisms which rely on knowledge artifacts, enabling rapid learning of rich models.
Large volumes of data are routinely collected from patients admitted to hospitals, particularly in intensive care units where critical illness requires close monitoring. This data can be mined for a variety of medical and academic purposes, but it generally contains highly sensitive information about the hospital's patients, limiting the extent to which it can be shared with researchers.
The Health Insurance Portability and Accountability Act has laid out guidelines for patient data de-identification, including the removal of eighteen specific identifying data elements. Removing these eighteen fields, which include patient names, telephone numbers, addresses, etc.,is not straightforward. Furthermore, additional steps may be required to remove distinctive details that indicate a patient's identity when cross-referenced with public records often isn't sufficient because a patient's identity could potentially be pieced together using clues from the data that are then supplemented with information from public records. This makes it difficult to make the data available to researchers. Thus, it is important to develop robust methods for de-identifying patient data while still leaving enough information in the set to maintain its value for academic and industrial research, healthcare quality improvement initiatives, and higher education coursework.
The Lab for Computational Physiology has received several highly detailed clinical datasets from various hospitals, providing us the unique opportunity to develop novel de-identification methods. In response to demand for a de-identification package, the lab has developed a de-identification package that is more generalizable, easier to use, and able to significantly outperform existing methods when benchmarked against a gold standard corpus.
About 400 million people in the world have diabetes, with about half of them not even knowing they have it. According to the International Diabetes Federation, in 2015, 1 person every 6 seconds died from diabetes-related issues. One way to decrease the death toll is to find the individuals who have diabetes but are undiagnosed, and provide them with the care and medication they need before the diabetes worsens. The most common techniques for diagnosing diabetes can be expensive and time-consuming. Moreover, many of these techniques require invasive procedures.
Since spectroscopy can be used to measure the amount of glucose in the bloodstream, our solution is to create a simple spectrometer that will analyze the skin and provide an image of the resulting spectra. This image will then be processed, and return the individual's diagnosis. The goal of this device is to make it noninvasive, low-cost and accessible.
For my project this year, I have worked on creating a low-cost device that will be able to attach to a smartphone and analyze the quantity of glucose in the bloodstream with a quick examination of the skin. Last semester, I created and tested the device, then over IAP, I went to India to conduct trials at a partner hospital on both diabetic and non-diabetic patients. The resultant data will be analyzed this semester to see if the device is able to distinguish between the diabetic and non-diabetic patients. Afterwards, the mobile application can be made to work with the device.
The electrocardiogram (ECG) records the electric potential that triggers the mechanical activities of the heart, and it plays a vital role in the diagnosis of cardiovascular diseases. Accurate classification of the ECG is important in medical engineering as it could enable early alerting of physiologic abnormality. This project aims to develop a collaborative annotation tool for classifying different types of clinical data, initially focused on labelling ECG recordings that exhibit atrial fibrillation (AF).
While performing initial investigations of ECG waveform data using Julia, the author realized that AF classification is a challenging problem because of the complexity of ECG signals and the limited availability of annotated data. Existing AF detectors are limited in applicability because they are typically developed using only classification of normal and AF rhythms on sets of carefully-selected data. This project classifies ECG waveform into four categories: normal sinus rhythm, AF, an alternative rhythm, or too noisy to classify.
The annotation tool this project presents enables collaborative-crowdsourcing of annotations for ECG waveforms, with the quality of the annotations being evaluated using a dataset of rhythms labelled by expert cardiologists. The tool will be hosted on PhysioNet, a repository of recorded physiologic signals, as a general tool to facilitate the classification of signals. Future work will focus on extending the tool beyond AF detection, including other ECG rhythms and other physiologic signals.
Recently, there has been increasing interest in the use of perfusion cultures in the biomanufacturing industry, due to its many advantages, which include higher attainable cell concentrations, higher productivity, and reduced space and cost. The defining characteristics of a perfusion culture are the continuous reintroduction of fresh media into the bioreactor and the continuous removal of spent media. The latter is accomplished using a variety of cell-retention devices, the majority of which are membrane-based, using hollow fiber membranes. The design of cell-retention devices is a challenge due to the necessity of maintaining a sterile environment in the device, as well as the propensity for such devices to become clogged or fouled during the duration of the culture.
In this project, we investigate the sorting capability of a novel membrane-less spiral microfluidic device in a perfusion culture setting, that is not only resistant to fouling and clogging, requiring little to no upkeep or maintenance, but is also easily scalable for industrial applications. We demonstrate successful retention capability by continuously removing cellular debris and nonviable cells from the bioreactor, while retaining the majority of the viable cells. We also investigate the effects of debris and dead cell removal on the quality and quantity of IgG antibody production, as well as explore other possible benefits of the device. This project is part of many ongoing efforts to discover more efficient and robust methods of cell-retention, and will also contribute to our current understanding of the mechanisms behind microfluidic flow.
Heavy metal contamination from industrial activity poses an ecological and human health hazard. Current methods of removing heavy metals from wastewater have distinct disadvantages: chemical precipitation generates residual waste sludge, while ion exchange technology is easily soiled and requires expensive infrastructure. Bioremediation using yeast has been proposed as an alternative to conventional treatment technologies, but current yeast biosorbents are nonspecific and less efficient than ion exchange technology.
Yeast that specifically accumulate metals could lead to more efficient removal of heavy metals that are present in minute quantities, as well as increase the feasibility of recycling heavy metals. We propose to use protein engineering and molecular cloning techniques to increase the specificity of yeast biosorbents, as well as to harness and enhance the natural ability of yeast to sequester metals. My Super UROP project focuses on two strategies: 1) improving the specificity and uptake magnitude of the yeast transporters and internal sequestration system and 2) mineralizing metals on the yeast surface using hydrogen sulfide precipitation and engineered peptides. Bioengineered yeast has the potential to be an inexpensive, environmentally benign, and specific bioremediation agent.
Internet of Things devices such as building sensors, health monitors, and industrial equipment often communicate using the Bluetooth Low Energy protocol, which is vulnerable to selective jamming attacks. With IoT applications in mind, we present a novel ultra-low-power fast-hopping wireless transmitter architecture co-designed with an on-chip security engine and control protocol. Our transmitter operates in the 2.4 GHz band and offers greater security against selective jamming by hopping between carrier frequencies every microsecond. The ultra-fast hopping rate prevents jammers from accurately detecting a target transmitter's carrier frequency and initiating interference before the target's next hop.
We also introduce the Rapid Adaptive Broad-Band frequency-hopping protocol for the Internet of Things (RABBIT). RABBIT is asymmetric, designed for sets of transmitter nodes communicating with a base station node, and adaptive, dynamically adjusting its configuration based on real-time interference conditions. The transmitter architecture's rapid hopping performance enables RABBIT to send each message bit on a different frequency, offering a theoretical transmit speed of one megabit per second for one transmitter. Together, the low-energy RF architecture and RABBIT offer new security properties for low-power wireless devices.
Our visualization aims to bridge the gap between observing the trackers and presenting that information to the user. Our web extension stores and updates our visualizations over time as needed. The visualization makes use of the taxonomy of trackers and presents the information as bar charts over time of different tracker types over the browsing history of the user.
Crack propagation prevention by implementation of spiral shapes in composite materials
Cracks are prevalent in all kinds of infrastructure used today. The propagation of cracks in buildings and bridges can cause severe damage to the integrity of these structures over extended periods of time and potentially endanger the lives of many people.
Nature has already found an effective way to deal with high stress in materials. Spirals, which are found in nature, demonstrate a good way of distributing stress to prevent breakage. They can be found in structures that experience high tensile and compressive stresses, such as shells, spider webs, and bone osteons. The goal of this project is to slow and trap the propagation of cracks by using spiral designs inspired by nature. Composite materials with a spiral-like pattern are designed with CAD software, manufactured with 3D printing, and tested with a tensile testing machine.
We demonstrate that, by using a combination of stiff and soft materials, cracks under tensile stress can be diverted. A soft spiral in a stiff matrix and vice-versa can make a crack follow a spiral path to prevent the crack from rapidly propagating through the entire structure in a straight path. This allows the material to remain connected even after reaching its failure point, slowly releasing the fracture energy and making this material concept more safe for engineering applications.
Often nature is used as a source of inspiration for scientific research. Over time nature has slowly eliminated less efficient solutions in favor of better answers to problems such as resistance to fracture. Many kinds of organic tissues show an amplification of mechanical properties with respect to their constituents far larger than man-made materials, despite being composed of weaker materials. For example, bone has a higher strength-to-weight ratio than mild steel, despite being made of collagen and hydroxyapatite, two materials with properties inferior to mild steel. The key to bone's strength lies in the way that these materials are organized at different levels. With current technology not only have we gained a better understanding of how these structures result in stronger materials, but recent developments in 3D-printing now allow for these structures to be easily replicated. If these structures were to be replicated using modern engineered materials the resulting composite material could be just as many times stronger. This research consists of studying the structure of bone, replicating it using 3D printers, and then examining the resulting properties. Through repeated testing and variation of several variables the design will maximize the desired properties. Results could be applied in potentially any manufacturing field with further improvements in 3D-printing, leading to stronger consumer plastics, implants, and perhaps construction material.
Quantum Key Distribution (QKD) is one of the most established applications of the quantum information field. It is a way to encrypt data to guarantee secure communication due to the "no-cloning theorem" in quantum mechanics. On the other hand, the mostly used class of digital cryptography techniques, such as the RSA, relies on the fact that it is computationally hard to decompose a large number that is a product of two prime numbers . The most efficient algorithm in solving the RSA problem takes an exponential number of steps. In addition, a successful implementation of a quantum computer will compromise the security of these algorithms in the future, since quantum computers can solve the RSA problem in polynomial number of steps. To further elaborate, it would take a current computer (with a 1 GHz CPU) about 1 million years to crack a 1000-bit RSA encryption, but a quantum computer (with a 1 MHz CPU) would take only about 1 day to crack the same encryption. Figure-1 shows the computation time to solve an n-bit RSA encryption by different quantum computer architectures and digital processor architecture. QKD is currently the only established cryptography method that is resistant against attacks using quantum computers.
Since the time QKD, also known as Quantum Cryptography, was proposed in the early 80s, many QKD protocols were developed. The most prominent protocol is the Bennett and Brassard 1984 (BB84) protocol, where Alice (the sender) randomly sends the bits 0 or 1 in two non-commuting bases to Bob (the receiver). Current QKD implementations are, however, limited in its rate at approximately 1 Mbits/s because the bits must be encoded at single-photon level. One solution to this problem is to exploit multiple wavelength channels in its transmission, such that a single channel, an optical fiber or a free-space link, can carry multiple QKD channels at the same time: a scheme called wavelength-multiplexing.
This poster describes the design, implementation, and testing of a 4-channel wavelength-multiplexed QKD system that consists of an FPGA-based driver and a quantum photonic chip. The FPGA-based driver modulates four different channels on the quantum photonic chip to send random bit signals synchronously at 1 GHz rate with 16-bit precision. The quantum photonic chip consists of four ring resonators that encode quantum information in time by producing a weak light pulse either in either the early (bit 0) or the late (bit 1) time bin, as well as their superposition. The design ensures that the QKD system can generate quantum-secure keys at a high-rate, while being compact.
A recently discovered class of defects in gallium nitride (GaN) wafers could provide exciting opportunities for implementing solid-state qubits. These defects can act as single photon emitters (SPEs) and are appealing due to the flexibility that GaN can provide in device fabrication. Photoluminescence measurements of these defects have shown high intensity over a wide range of wavelengths at room temperature. Despite the broad spectrum observed, we believe many of these emitters are from the same type of point defect and that differences in luminescence occur due to strain in the wafer.
To further understand GaN SPEs, we use a combined experimental and computational approach. We use cathodoluminescence microscopy to determine variations in the structure and chemical composition of GaN defects. We also plan on measuring the strain dependence of the emission wavelength to determine the role of lattice strain on photoluminescence. Experimental data from these studies will allow us to use density functional theory (DFT) with the local density approximation-1/2 method to obtain transition energies and other electronic properties. We will then compare transition energies with experimental data to validate the method's accuracy and use the DFT results to determine the feasibility of implementing qubits with these defects.
Year to year, the amount of metagenomic data increases exponentially, at a rate that computing power will soon no longer accommodate. Two problems arise with data at this scale: storage and read mapping. Compressive algorithms are key in order to solve both efficiently. We intend to develop compressive meta-genomic algorithms to address these issues in a manner that maintains accuracy and increases efficiency. By accomplishing this, we can unlock the potential of metagenomics to target cures to diseases at the microbial level.
Metagenomic data breaks down into roughly two parts: sequences and their corresponding quality scores, a measure of sequencer read accuracy. Current methods focus compression efforts on quality scores, since they make up a majority of metagenomic storage. The project's results prove that quality score compression degrades downstream metagenomic analysis. Additionally, we explore compression as applied to read mapping. Here, we have demonstrated a 6x speedup compared to current approaches by taking advantage of redundant sequences across both the reference genome and the reads. Compression with regards to biological data is crucial in deriving tangible results from increasingly large datasets and the proposed methods dramatically accelerate the metagenomic pipeline.