Blood pressure (BP) serves as the primary indicator of a patient’s cardiovascular health. Today, cuff-based BP monitors are the gold standard for routine blood pressure monitoring. However, they are prone to inaccuracies and cannot provide continuous readings. Continuously monitoring BP would allow patients to observe their BP fluctuations from eating, medicine intake, and exercise, thus empowering individuals diagnosed with hypertension to make better-informed health decisions. This motivates the need for a non-invasive, continuous blood pressure monitor. Prior studies have already shown the potential for pulse transit time (PTT), which is the time for a pulse wave to travel between two arterial sites, to be used for non-invasively measuring BP. In this work, I hope to improve upon this technique. To do this, I focus on testing and improving the pulse detection accuracy of a system incorporating an optical sensor array in a surgical eye protection face mask. By getting a better resolution of the pulse waves, I believe the estimate of BP will be more accurate and, in turn, provide a valuable dataset to further investigate the relationship between PTT and BP.

With cardiovascular disease as the leading cause of death worldwide, there is a need for improved wearable monitoring tools for assessing the health of the cardiovascular system. Photoplethysmography (PPG) is a continuous, non-invasive measurement that encodes a multitude of informative vital signs, including heart rate, heart rate variability, respiratory rate, cardiac output, and arterial stiffness. Although existing PPG sensing technologies record from the finger or wrist, the face presents a promising and underutilized location for wearable pulse sensing. This work presents a novel wearable PPG sensing system that records at multiple wavelengths and facial locations. As a proof-of-concept, we seek to evaluate a potential application of our system incorporated in a surgical face mask for use in intra-operative hemodynamic monitoring. By collecting data with our system alongside ground truth cardiovascular vital signs, we can build and test non-invasive inference algorithms. After validating our system’s heart rate detection accuracy with a standard error of 2.84 beats per minute, we now proceed to test our device’s ability to infer additional cardiovascular parameters. In addition to showing promise for novel non-invasive, continuous surgical monitoring, this work has broader implications for wearable health applications based on face-worn form factors such as glasses, helmets, and headsets.

Cellular regulation of transcription is a complex phenomenon, with a wide range of biological determinants influencing the binding of transcription factors (TFs). Being able to accurately predict TF binding is important to pinpoint noncoding DNA regions where mutations are likely to cause disease. Existing approaches are able to accurately predict TF binding across the genome, but are not focused on interpretability. Our approach, EXPERT: Explainable Prediction of Transcription Factor Binding based on Histone Modification Data, achieves state-of-the-art predictive performance while simultaneously offering local interpretations that reveal biologically meaningful relationships by using a stage-wise boosting technique and local additive feature attributions. We utilize a performant non-linear model (XGBoost) and an efficient local feature attribution method (TreeSHAP) to demonstrate how to increase the signal learned from a set of biologically meaningful features (histone binding in DNA) using a stage-wise gradient boosting scheme while maintaining high performance. EXPERT enables understanding predictive models through biological explanations and can further our foundational understanding of the epigenome by highlighting novel biological relationships.

There is currently extensive demand for optical media like CDROM, DVD and Blue ray disks for data storage with computer technologies. Here we combine mathematical modelling studies and photonic laser diffraction experiments to study the optimization of data storage in different types of optical media. Using calculus-based studies, we estimated the data storage capacities in these systems and calculated the CD, DVD and blue ray disk arc length and data storage linear densities. These are in good agreement with reported values. Using red, blue and green laser sources at our photonics lab, we conducted laser diffraction studies and estimated the line spacing of CDROM, DVD and Blue ray disks. The advancement from CDs to DVDs yields higher data storage densities. In the high capacity blue rays disks, because the physical structures called pits that store data on the disks become smaller, there are other challenges in realizing these smaller devices, which make it more expensive. The CD/DVD players' lasers operate at the diffraction limit resolution of light and provide maximum data capacity for their geometry. Magnetic media like floppy disks, hard disk and magnetic tapes are also used for computer data storage. We have estimated the maximum data storage capacity from magnetic floppy discs. We used curve fitting methods to analytically represent the magnetic read-back pulse as Lorentzian functions for data modeling. Our studies provide an integrated STEM learning of data storage in optical and magnetic media.

As information processing machines approach the nanoscale level, DNA has emerged as a powerful tool in molecular engineering systems. The specificity and programmability of its hybridization interactions offer flexible and fine-tuned control over reacting species. Among the DNA computing techniques used today, strand displacement circuits are highly popular, with potential applications ranging from disease diagnostics to DNA-based artificial neural networks. The fundamental mechanism of these circuits is the hybridization of a single-stranded DNA input strand to a double-stranded complex which triggers the release of a prehybridized output strand. When released, this output can be detected and used to characterize circuit behavior. The output strands of strand displacement circuits are typically read out using fluorescence spectroscopy. However, due to spectral overlap of traditional reporters (e.g. FAM, TAMRA, Cy5), the number of outputs that can be detected in parallel is severely limited. To address this, we present the use of nanopore sensing technology as an alternative readout device that enables highly scalable, real-time detection and quantification of DNA strand displacement circuits. We demonstrate dynamic sensing of an operating circuit within the flow cell of a commercially-available high-throughput nanopore sensor array (Oxford Nanopore Technologies’ MinION device) and show that strand capture frequency can be correlated to concentration, allowing for direct quantification of desired circuit elements. To investigate this reporter strategy’s multiplexing potential, we present a collection of ten orthogonal circuit output sequences (barcodes) that can be classified at the single-molecule level from raw nanopore signal data using machine learning, with the potential to scale to larger barcode sets. We conclude that nanopore-based detection of strand displacement circuits holds key advantages over fluorescence-based methods for real-time, multiplexed circuit readout on an inexpensive, portable sensor device.

Automatic gender recognition (AGR) is a subfield of facial recognition that has recently been scrutinized for bias in the form of misgendering and erasure against various identity groups in our society. Recent studies have found that several commercial AGR classifiers (from Microsoft, IMB, Face++) are biased against women and darker-skinned people as well as gender non-binary people. In this work, we investigate and quantify AGR classifier bias against transgender people by developing and evaluating three different convolutional neural networks (CNN): using images of cisgender individuals, using images of transgender individuals, and using images of both cisgender and transgender individuals. We find that the cisgender trained classifier is 91.7% accurate when evaluated on cisgender people, but only 68.9% accurate when evaluated on transgender people, with the worst performance on trans men with 38.6% precision. We also find that the classifier trained on the combined dataset performs nearly as well as and occasionally outperforms both other classifiers when evaluated on their own datasets, highlighting potential methods for avoiding overfitting. Additionally, we visualize how the classifiers differ by obscuring different parts of the face. Overall, the disparities of accuracy between each classifier demonstrate the degree to which they are impacted by the composition on their dataset and highlight the possibility for commercial AGR classifiers potential to misgender trans people, in particular, transgender men, at a high rate.

 The internet has changed the norms for how we write and communicate. Many major technology companies communicate with their users in much more casual and conversational language than the formal written English taught in schools. For instance, contractions and sentence fragments are common, the word “like” often used in place of “such as”, and “info” often used instead of “information”. This casual writing style may help users view a company as more approachable, but they may also perceive the casualness as lacking professionality and trustworthiness. Trust and taking the right action are especially important in security-related interfaces, as a user’s security and privacy may be compromised if an interface fails to educate users on secure behaviors. Through an online quantitative study we will explore the effects that formality of language has in security-related prompts. These effects include: how a user understands a prompt, their perception of the prompt’s formality, and how likely they are to take the action the prompt suggests. We will also investigate how users perceive the formality of various major technology companies and whether these perceptions match how the companies actually communicate with users. As average-level formality is different for everyone, we will analyze our results across different demographics, such as education-level, age, and the country the person grew up in. Our goal is to measure the likelihood of a person to take an action based on the wording of a security prompt, the person’s sentiments towards the prompt, and whether these depend on the person’s demographics and the company with which they are interacting. Our work serves a practical purpose, that is, helping technology companies decide what tone they want to address users with to accomplish their goals. It also serves a more theoretical purpose, in furthering understanding in the intersection of human-computer interaction and security.

Software engineering is a known difficult task that spans a wide variety of problems; as such, an important skill for seasoned software developers is problem solving. Current software engineering research focuses on building tools that support software development processes, but very little research has been done to assist developers with learning programming strategies. Yet, previous research has shown that developers using explicit programming strategies (i.e. procedures in problem solving that were verbally described) were objectively more successful at code design and debugging tasks. In this research project, we extend the previous work in understanding how programming strategies may be used at scale, and whether it is a potentially effective way of improving developer productivity. We propose a novel platform composed of a repository of explicit programming strategies across various programming activities, such as debugging, design, and testing. Developers will be able to search, use, create, and provide feedback on programming strategies on this platform, which will require innovations in defining how explicit programming strategies are searched and indexed. With this explicit strategy sharing platform, we want to understand the experiences of developers who are strategy seekers or givers as well as their motivations using the platform. For strategy seekers, we would like to understand their experience in using strategies, as well as how the feedback process may evolve strategies. For strategy givers, we would like to understand their experience in writing strategies and why they do it. After building the platform, we will deploy the platform in a classroom setting that relates to software engineering, and allow students to use the platform organically. Then, we will perform user interviews and data analysis to evaluate our research question.

In the general population, adherence to a daily routine is linked with well-being.This appears to be the case to an equal if not greater extent among individuals with schizophrenia. Individuals with schizophrenia-spectrum disorders who consistently engage in activities that typically occur in a routine – e.g. employment, education, healthy sleep, social connections, and physical activity – enjoy an array of physical and mental health benefits. The study of adherence to routine has been limited by the use of retrospective scales, which are common in clinical research. These measures require respondents to provide estimates of the amount and frequency of behaviors over weeks or months. Such estimates are insufficiently granular to assess adherence to routine. The present study aims to examine the relationship between behavioral stability and symptoms in schizophrenia. Our team previously deployed a multi-modal mobile assessment system in a sample of individuals with schizophrenia for twelve months. In this study we revisit those data to develop models that quantify within-day adherence to routine among individuals with schizophrenia. We operationalize adherence to routine as defined in an individual’s behavioral stability, or the extent to which their behaviors detected by sensors stay stable from one day to the next during the study period. The present study has four main aims: whether (1) passively sensed behavioral stability can be quantified, (2) whether it is associated with symptoms and dysfunction in schizophrenia, (3) whether the addition of behavioral stability improves predictions of symptoms and dysfunction, and (4) whether this behavioral stability metric might be associated with future risk of poor outcomes.

Although DNA is best known as the molecule that encodes the genetic information of all living things, they can also be utilized as chemical building blocks. Using DNA as the building material of choice, I am working on constructing a complex DNA circuit, in specific, a binary neuron network that utilizes DNA strand displacement reactions to compute a winner-take-all or majority voting operation. Winner-take-all computation is just one type of competitive neural network model, mirroring the lateral inhibition and competition seen in biological neurons of the brain. DNA circuits are efficient at collecting and responding to information within a biochemical environment; processing information locally and producing specific outputs in response to changing environmental conditions. DNA strand displacement is the process by which two DNA strands that are partially or fully complementary hybridize to one another, thereby displacing one or more pre-hybridized strands. Strand displacement reactions are facilitated by a "toehold" domain, a region of exposed DNA on a double-stranded gate complex that is complementary to an input strand.The winner-take-all function can be broken down into sub-functions that use four distinct seesaw DNA gate motifs: weights, thresholding, annihilator, and catalytic amplifier. Additionally, we aim to combine spatial separation in microfluidic droplets with a stricter choice of network architecture to address previously seen issues of scalability. By isolating each computational primitive in droplets, DNA species can be re-used for all primitives of the same network layer. Recently, we have experimentally tested a 5-input neuron and used manual pipetting to simulate droplet operations. While the leak caused problems for patterns close to the decision boundary, we could successfully compute well-separated patterns. Next, we plan on optimizing the sequence design to reduce leak and scale up the network size by automating the network execution on a microfluidic droplet device.

Understanding cause-and-effect is key for informed decision-making. The gold standard in causal inference is performing controlled experiments, which may not always be feasible due to ethical, legal, or cost constraints. As an alternative, inferring causality from observational data has been extensively used in statistics and social sciences. However, the existing methods critically rely on a restrictive assumption that the population of study consists of homogeneous units that can be represented as a single flat table. In contrast, in many real-world settings, the study domain consists of heterogeneous units that are best represented using relational databases. We propose and demonstrate CaRL: an end-to-end system for drawing causal inference from relational data. In addition, we built a visual interface to wrap around CaRL. In the demonstration, I will use CaRL, which I have implemented, to show a live investigation of causal inference from real academic and medical relational databases.

Although transcriptomes are highly tissue and cell type specific, curated gene set collections are not constructed or analyzed with recognition of this bias despite much of transcriptomic analysis depending on curated gene set collections. Prior work has recognized the potential for gene bias due to the variable nature of manually curating gene set collections, but coverage has yet to be characterized across all tissues and in all commonly used gene set collections. The goal of this study was to perform a comprehensive analysis of curated gene set collections from the data repository Molecular Signatures Database (MSigDB) based upon tissue specific expression. We analyzed KEGG, REACTOME, BIOCARTA, and Gene Ontology (GO) including Biological Processes, Cellular Components and Molecular Function gene set collections available on MSigDB. We curated lists of enriched and elevated genes as defined by Human Protein Atlas for 36 tissues. Analyses, visualization, and statistical analyses were performed using the R statistical programming language. We revealed that the MSigDB gene set collections differ among themselves in the fraction of tissue genes covered. GO Biological Processes has the highest gene coverage. BIOCARTA has the lowest gene coverage. Additionally, each collection differs among tissues in the fraction of genes covered. We also showed differential gene coverage among tissues even when collections are combined. Within elevated tissues, the liver has the highest and the fallopian tube has the lowest gene coverage. Within enriched tissues, the lymphoid has the highest and the testis has the lowest gene coverage. We created a database describing the presence or absence of tissue specific genes for each tissue with which researchers can elect the most appropriate gene set collection to use for analysis of a specific tissue. This increases the utility of our findings and creates a direct resource for researchers in the field.

During surgeries, constant blood pressure sensing is important to counteract the possibility of hypotension, which is a dangerously low drop in blood pressure. Although monitoring blood pressure with invasive arterial catheters can provide continuous information to the anesthesiologist, discomfort and health risks related to using an invasive method limit their use to only a few high-risk surgeries. While blood pressure cuffs to non-invasively measure blood pressure do exist, they are usually uncomfortable and can only periodically record blood pressure. This motivates the need for a tool to perform continuous, non-invasive blood pressure sensing. Here, we validate the use of facial photoplethysmography (PPG) signals to accurately infer blood pressure. Using our wearable eye face mask mounted with optical sensors, we collect PPG signals while the subject is undergoing surgery. Then, we can calculate blood pressure from the PPG signals and subsequently determine the accuracy of the blood pressure measurements. To infer blood pressure from non-invasive facial PPG signals, we apply temporal deep learning techniques that can model dynamic changes in the cardiovascular system. First, we test potential filtering methods by performing peak detection on noisy PPG data to determine which filtering method cleans the signals the best. Then, we incorporate several machine learning models, including autoencoders, to compress parts of the PPG signals into more featurized components. In the final step, we test the face mask sensor data to find the root mean square error (RMSE) of the predictive model compared to that of the ground truth. We expect that it is possible to infer blood pressure from noisy sensor data, as an alternative to invasive arterial catheters.

Deep convolutional neural networks (CNNs) have seen widespread application across problems in the life sciences where probabilistic models built on simple assumptions are insufficient. One area where Deep Learning has seen considerable success is protein structure modeling, where a protein’s tertiary structure is predicted using physio-chemical information. State-of-the-art structural prediction methods often yield high fidelity structures, but some regions (e.g. loop regions) still pose a significant challenge. To augment low fidelity structures, I propose a novel framework based on a conditional deep generative model for improving residue-residue contact predictions in unreliable local regions (ULRs), implemented as a residual convolutional neural network with high attention to contextual protein information. The work is extended from Nao Hiranuma's DeepAccNet developed in the Baker Lab. My network will supplement existing structural refinement protocols in regions where contacts are poorly predicted. If successful, this will greatly improve the ability of modern protein refinement protocols to recognize more difficult structural motifs.