Drink Guardian Pitch for JHU Design Day
About 20 million students will attend college and universities in the fall of 2020.1 For many, college should be some of the most exciting years of young adult lives, lurking beneath the excitement hides a dangerous national issue—drug-facilitated sexual assault in college communities. According to a study conducted by the American Addiction Centers, out of people who have had their drink spiked, 52% had their drink spiked for the first time in college2, while more than half of college sexual assault cases occured in the first semester.2
To combat this issue, and hopefully decrease the occurrence of drug-facilitated sexual assault in college communities, we propose Drink Guardian: a robust drug-detecting device that continuously monitors a user’s drink, and has the ability to signal both the user and a sober friend or partner upon detection of a drug. In particular, we plan to accurately determine the presence of Rohypnol, a common date-rape drug, as well as limit false negative results using both fluorescence and electrochemical methods. The fluorescence test will detect two peaks at 250 and 315 nm for unprotonated Rohypnol, or a single peak at 280 nm for protonated Rohypnol.3 The electrochemical test, on the other hand, will look for a current peak at around -0.3 V which lies between Rohypnol’s first oxidation and second reduction reaction equilibrium.4 When both tests are positive, a notification and vibration will be sent to the user and a sober friend of choice, via bluetooth from Drink Guardian to their smartphones.
This report will further detail Drink Guardian’s product specifications and development, manufacturing process, and financial analysis. With our proposed manufacturing process, we will produce 1,132 devices a day to be able to launch 293,500 devices in our first year on the market. We expect universities to cover the $100 cost of Drink Guardian and provide our devices to students. Moreover, the cost of fixing broken equipment will be provided through a warranty, as it is built into our business costs. Based on our financial analysis we predict to break even after roughly five years.
Project pitch of Better Life Asphalt, a novel, multibillion market, bio-based asphalt binder.
A video describing background, motivation, methods, and results of our sulfur-infusion technique for improving Hole Transport Layer charge transport in Lead Sulfide Colloidal Quantum Dot solar cells.
Demonstration of our car moving and stopping. As well as a short powerpoint presentation on our mechanisms.
A description of each of our mechanisms with our data.
This poster covers the simulation and synthesis of aluminum nanoparticles for photocatalytic applications.
This video covers the background and prevalence of widefires followed by a detailed debrief on the scope of our solution and its economics.
The poster provides a concise, one-stop shop for all important information of our project.
A description of Plerion Medical’s POET Catheter CBID project.
A description of Pidgin’s Hearing Rehabilitation CBID project.
GORD21 Project Overview Video. Outlines reason why pump is necessary and the objectives of our testing.
Ventilator-induced organ damage has been associated with the delivery of high tidal volumes and, more recently, with the delivery of high mechanical power (MP) by the ventilator, which describes the energy transferred from the ventilator to the lung tissue per unit time. This study aims to build machine learning models that leverage knowledge of ventilator settings and other patient features to predict physiological deterioration and mortality in mechanically ventilated patients. To this end, patient data from the Phillips eICU Database were used to build predictive models. Inclusion criteria include age ≥18 years, ICU stay duration ≥48 hours, volume- or pressure-controlled mechanical ventilation, and ventilation duration ≥48 hours. Physiological deterioration was defined as any increase in the daily Sequential Organ Failure Assessment (SOFA) score between days 3–7 from initiation of mechanical ventilation. Classification models including logistic regression, random forest, and support-vector machines were evaluated for predicting mortality and physiological deterioration using area under the receiver operating characteristic curve (AUC). Using these models, we show predictive power for mortality (AUC: 0.78) and for overall physiological deterioration using total SOFA scores (AUC: 0.74). We have also shown predictive power for organ-specific deterioration, specifically for renal (AUC: 0.74) and pulmonary systems (0.79), using their respective SOFA subscores as outcome measurements. Inclusion of time-series physiological data may further improve the performance of these models.
Model Design and Test Results
During the acute stroke period, there is a disruption of the blood-brain barrier and cerebral blood flow autoregulation, which results in a direct relationship between hemodynamics and brain perfusion. Improved understanding of the relationship between hemodynamic trends and clinical outcomes could provide predictive insight. The goal of this work was to predict hospital length of stay using the patient’s demographic information and hemodynamic profile. We conducted a retrospective cohort study for stroke patients with data that included hemodynamic parameters such as pulse, temperature, and blood pressure as well as demographics and comorbidities. Our primary outcome was length of stay greater than or equal to 7 days. The ROC and precision recall curves for the models are presented with the XGBoost showing the best performance. Models that used the first 24, 48, or 72 hours of patient data were built. The application of these models could augment the care coordination process and provide a categorization of patient recovery trajectory.
Here, we summarize our project and the impact we plan to have with it.
A powerpoint presentation of what we’ve accomplished this past school year.
Video describing the inspiration, application, and results of our product thus far with possible next steps.
5:30 min video
We presented our solution to a problem that was brought forward by Oceaneering International. In our video we introduce the problem and demonstrated our working prototype in action.
This video describes the background of the project from a real user, as well as the various ways the we have attempted to make the socket. Our final prototype is shown along with some results from initial testing. A future user also describes what excites him about our socket.
PRIME21: Oxygen Concentrator for Low-Resource Settings
This video is a brief overview of the design goals and prototypes fabricated over the past academic year.
Here you can see some of results and what we are working with, from the robot arm to the thermal camera results. We look forward to design day!
Quick summary of the project and work the FMIC21 team has completed from fall 2020 to spring 2021.
Music: https://www.purple-planet.com
Music: www.bensound.com
Stock Videos (microscope): www.videvo.net
A brief voiceover of our project poster.
This video describes the need for the mirror device we created that allows for students and professors to share written work on zoom in a low-tech and hands-free way. We shed light on parts of our design process and also demonstrate how the device can be used.
Project overview starting with problem statement and objectives, followed by design and prototyping, and closing with final design and next steps
This video introduces Xingyu (Jasmine) Hu’s senior design project at the Department of Material Science and Engineering, Johns Hopkins University. The project focuses on screening the response of the cellular transcriptome of mouse mesenchymal stem cells (mMSCs) to different combinations of biophysical cues by using RNA sequencing and data analysis tools.
In this video, we show how we created a system accessible to associates with varying abilities, through a design process driven by understanding associates’ perspectives.
Our video highlights the difficulties that students and educators face when trying to engage with virtual lab experiences. We demonstrate our solution, a Choose Your Own Adventure virtual lab experience, and descibe the feedback we recieved from testing with college students.
CBID: LaparoscopiX Training – Expanding Minimally Invasive Surgery in Kenya
CBID: LaparoscopiX Training – Expanding Minimally Invasive Surgery in Kenya
Two-minute video which provides an overview of the project.
Short video providing project background, clinical need, solution, and future steps.
Poster on the project background, clinical need, design figures, and future steps.
Project handout
CBID: LymphaSeal – Helping Children with Postoperative Lymph Leaks Return Home Faster and Healthier
A powerpoint presentation of what we’ve accomplished this past school year.
Deliverable for Design Day Presentation
The results show that gelatin methacrylate serves as a promising alternative to polyacrylamide in forming a tough interpenetrating network that can serve as a medical adhesive. The addition of tannic acid to the material system dramatically improved important mechanical properties of the material, but further work will be required to increase the fracture strain (stretchability) of the hydrogel.
Overview of Quest2Learn: our current progress and how we got here
Our video follows a hypothetical student through the pandemic lockdown as they lose motivation to exercise and regain it with the help of HealthTime.
Watch a short introduction to our team, why we’re competing in the Collegiate Wind Competition, and what we’ve built this year!
Poster for Quest2Learn project and learning
Project poster
This video shows how our Dynamic Brace reimagines the clubfoot bracing experience, making the healing journey more comfortable for kids.
Animation explaining the vector surveillance process and the role the VectorCam system plays in enabling expansion of this work
This poster outlines the necessary background, methods, results, and conclusions of this project. There are several figures that show observed trends on how lipid composition influences transfection efficiency that ultimately lead to our conclusions.
This is a pitch video which briefly describes the Sprout platform, the problem it solves, and how exactly we will build it out.
This project poster explains the Sprout investment platform in detail.
This project poster outlines our Ackoji Outlook add-in, describes the problem it solves, the software architecture of the add-in, as well as next steps for the future.
We display the general method in conducting the baseball scheduling optimization and give a list of cooperated leagues and projects we’ve worked on.
Faceoff wins and losses matter but not as much as the starting situation of a given faceoff. In the case of offensive zone or defensive zone faceoffs, the team with the more favorable offensive state at the location of the faceoff will still be expected to achieve more expected goals than the other team regardless of which team wins the faceoff. The logical end point at which non-faceoff, in-game events take precedence over game play rather than the preceding faceoff is worthy of debate. However, using time of a zone change as a logical starting point and noting that on average 22.5 seconds elapse before the first zone change after a faceoff, we can estimate that a faceoff win in the offensive zone by the offensively positioned team is worth about 0.052 expected goals and a faceoff win in the defensive zone by the defensively positioned team is worth about 0.049 expected goals. On a more general note spanning all situations, a faceoff win is worth an average of 0.015 expected goals. This may seem tiny but becomes quite notable when considered in the context of the average NHL game featuring 59.3 faceoffs. This suggests there are nearly 0.89 expected goals per game up for grabs at the faceoff dot. Accounting for both the gain of winning a faceoff for your team and forfeited gain of stealing a faceoff win from the other team, winning just six more faceoffs a game would be the equivalent of adding 0.18 additional expected goals in offense each game. That translates to 15 additional expected goals over the course of a full season or roughly the equivalent of adding an additional middle-six forward that could easily cost $4 million annually against the salary cap for the likely lower cost of personnel that can win six more faceoffs. That surplus value represents nearly five percent of the salary cap, which is invaluable to any team with Stanley Cup aspirations. Our research suggests that faceoffs represent a market inefficiency and ripe opportunity for NHL teams to cost-effectively win more games. Faceoffs matter.
This poster includes a summary of the bioretention basin and project site, including visuals of the basin design.
This work focuses on optimizing and enhancing the absorption of a MoTe2 photodetector using plasmonic aluminum nanoparticles. Simulations were conducted using the finite-difference time-domain (FDTD) method to solve Maxwell’s Equations. Optimization considerations involved both the thickness of the SiO2 layer below the structure and the morphology and size of the Al structures. The resulting device proposed involves prolate Al hemispheres, which remain understudied in literature. The size-tuning allows for enhancement control in either the visible or near-infrared (NIR) region.
The motivation behind this project is to setup a thief joule as a DC-DC converter relying on Bipolar Junction Transistors and generate a functioning PCB component to be used on low voltage power sources.
Abstract, Introduction, Film Characterization, Prototypes/Schematics, Conclusion, and Acknowledgements
This poster shows the method and the results derived that led to improvements in the upper bound.
The poster discusses the baiting and ambushing problem and the methods used to solve the problem. Meaningful results from the simulation environment are used as a discussion point and these results allow us to form conclusions.
The goal of this project is to determine which NFL metrics are predictive of future NFL performance, by comparing players who have signed second contracts with teams. By determining these key metrics, we can predict which second contracts were good deals (“booms”) and which second contracts were bad deals (“busts”).
We seek to determine the metrics that NFL teams can use in determining whether to give a second contract to players after their rookie contracts are over. We can also answer whether advanced metrics improve accuracy in determining future production (and, subsequently, contract valuation) and figure out which years in a player’s rookie contract are most indicative of second contract performance.
Poster Voiceover
Poster of algorithm summary and preliminary experiment results on synergistic lifelong learning
Poster of comparing decision forests and deep networks in different data domains with preliminary experiment results
A poster with information about the current Malaria diagnostic market, market gaps, and our proposed Malaria breathalyzer device. Several images with reaction membrane, device, and manufacturing schematics, as well as a financial forecast and development timeline.
BreatheBeauty Project Poster
Poster includes details on our shampoo concentrate refill pods project with charts and graphs to support our work
Poster includes details of macroporous alginate-nanofiber composite device, including synthesis overview, outcomes of device deployment in a swine model, and mechanical properties.
Our presentation gives an overview of our design solution including a description of our deliverables, a maintenance plan, and a cost estimate.
Almost one million Americans are admitted to the intensive care unit with sepsis. The standard treatment for treating patients with sepsis is to administer broad-spectrum antibodies because of the lack of microbiological data available upon arrival to the ICU. This is especially problematic given that at least 50% of original “presumed sepsis” diagnoses are incorrect and in fact not of bacterial origin, resulting in the substantial overuse of antibiotics. Consequently, clinicians need a faster and more accurate method of diagnosing the microbiological origin of sepsis for children in the pediatric intensive care unit in order to deliver antibiotic-specific treatments in a timely manner, while preventing the overuse of broad-spectrum antibiotics. We aim to use machine learning to accurately predict the microbiological origin of infection in children faster than the time it takes hospital lab tests to return. We hypothesize that a predictive model analyzing a patient’s Physiological Time Series Data and electronic medical records can identify the origin of infection in the first 48 hours of PICU admission. Patient encounters were labeled as Bacterial, Nonbacterial, or Not Infected by clinical experts based on observed temperature instability and blood culture results in the PICU. Using the Random Forest Classifier in a One Vs. Rest Multi-Class Classification, we were able to predict the encounter’s microbial origin of infection using physiological time series data, all with AUC over 0.6. We were able to identify windows of physiological signals that are relevant to each classification.
Johns Hopkins Design Day 2022 — Early Prediction of Length of Stay in Hospitalized Patients with Stroke and Traumatic Brain Injury
Stroke is one of the leading causes of morbidity and mortality worldwide, and traumatic brain injury (TBI) is one of the major causes of disability in children and young adults. ICU length of stay (LoS) is considered a primary driver of inpatient costs. The prediction of length of stay in the early phase of hospitalization can inform resource allocation and improve clinical decision-making to ultimately reduce medical spending. The team used patient data available in the first 24 hours of stay to predict length of stay for patients with traumatic brain injury and stroke in the NCCU. The predictive features driving length of stay were also identified and ranked. Accurate predictions about length of stay for NCCU can be made using patient data available in the first 24 hours. Considering an average cost of 4000$ per day in the NCCU, cost of stay can be estimated. Support vector machine (SVM) was the best performing model given our data, and the Glasgow Coma Scale (GCS) is the most important feature in predicting patient length of stay.
JHU BME Design Day 2022 presentation of Team Beaver’s project “Predicting Neurologic Injury in Pediatric ECMO,” for Precision Care Medicine.
Pediatric extracorporeal membrane oxygenation (ECMO) is used in emergent cases to provide life support for critically ill patients under the age of 17. However, there is a high mortality rate of pediatric ECMO patients (40%) which has mainly been attributed to the risk of acute neurologic injury due to the delicate balance of clotting and bleeding in ECMO. The purpose of this study is to analyze and predict neurologic injury through the development of a supervised machine learning models using data from the Pediatric ECMO Outcomes Registry (PEDECOR) database. The first model uses only pre-ECMO demographic and patient history features, and then a dynamic model is created by incorporating daily lab measurements and blood products. The static models resulted in low AUC, 0.77. The dynamic model produces daily predictions while the static model only has one overall prediction. Feature importance evaluation was performed on both models and resulting features can be studied further to help guide the administration of blood products and anticoagulation factors during ECMO.
COVID-19 is a highly transmissible infection caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), with 506 million cases worldwide and has resulted in 6.2 million deaths. Little is known about the human genetic and immunological basis of resistance to SARS-CoV-2. It has been observed that mean secondary attack rates for SARS-CoV-2 infections can reach up to 70% in some households, and several families reported that all their members except one of the spouses were infected. This suggests that some highly exposed individuals may be resistant to infection. In addition, little is known about whether the occurrence of COVID-19 resistance differs between people by health characteristics as noted in the electronic health record. In this study, we developed a machine learning model to predict COVID-19 resistance individuals with prior COVID-19 exposure using EHR data from the JH-CROWN dataset. Exploration of the dataset through clustering with Maximal-frequent All-confident pattern Selection and Pattern-based Clustering (MASPC) presented a discrepancy between patterns of diagnostic codes in resistant and non-resistant patient cohorts and XGBoost was found to have the highest performance in our modeling. We hope to validate the features found to be associated with resistance/non-resistance through more advanced association studies.
In-hospital cardiac arrest (IHCA) occurs in approximately 15,200 pediatric patients annually. However, only one out of four patients survives until discharge. We have created machine learning models to predict IHCA up to three hours in advance. Our data consists of 240 Hz ECG waveform data, 0.5 Hz vital signs time series data, and medications data from 1,145 patients in the pediatric intensive care unit. We generated five minute windows on our data to perform feature engineering. We created 23 heart rate variability (HRV) metrics from ECG R wave peaks (NN intervals) after filtering outliers and removing ectopic beats. A total of 96 summary statistics were calculated for 12 vital signs, such as respiratory rate and blood pressure. For the medications data, drugs were classified into 46 therapeutic classes, and boolean features were created. We trained six machine learning models: logistic regression, support vector machine, random forest, XGBoost, LightGBM, and a soft voting ensemble. After evaluation on held out testing data, the XGBoost model performed the best, with 0.971 auROC, 0.798 auPRC, 99.5% sensitivity, and 69.6% specificity. Upon investigating feature importances, we determined that a combination of heart rate variability metrics, vitals signs data (e.g. a low respiratory rate), and several therapeutic classes of medications are key indicators of pediatric IHCA. We have created high-performing models to predict pediatric IHCA up to three hours in advance by combining ECG waveforms, vital signs time series data, and medications data. This increased warning time will allow clinicians to intervene earlier, improving patient outcomes.
Our project poster summarized our team’s design ideas for the sustainable renovations to the Northern End of Levering Hall and The Glass Pavillion
Purpose: When to start weaning a patient from mechanical ventilation can be a challenging dilemma for a clinician. This study therefore aims to develop a clinical decision support tool that leverages machine learning models to predict the risk of weaning failure upon admission to the intensive care unit.
Methods: Data from the MIMIC-III Clinical Database were utilized to obtain the patient cohort. Inclusion criteria included: 1) Age 18 years and 2) ICU admission undergoing ventilation. Patients were classified into one of four groups according to the WIND classification and excluded if they did not experience a weaning process or died within 7 days after extubation. To develop a binary classification model, patients in groups 1 (short weaning) and group 2 (difficult weaning) were labelled as weaning success and patients in group 3 (prolonged weaning) were labelled as weaning failure. Classification models included logistic regression (LR), random forest (RF), and Extreme Gradient Boosting (XGBoost) models. A clinical decision curve analysis (DCA) was constructed plotting clinical risk threshold vs. net benefit to evaluate the clinical utility of the models.
Results: Of 46520 records, 12813 met the inclusion criteria. 1569 patients did not undergo any separation attempt and were excluded from the analyzed cohort. Of the remaining 11295, patients were classified into 3 groups: short weaning (n = 7082), difficult weaning (n = 2548), and prolonged weaning (n = 1614). 1234 underwent a tracheostomy during their admission, of which 172, 256, and 799 were within groups 1, 2, and 3 respectively. The RF and XGBoost models had superior performances at AUCs of 0.91 and 0.90 respectively compared with an AUC of 0.652 for LR. Although RF and XGBoost had similar performances on the ROC curves, the DCA showed XGBoost to have a 20% additional net benefit compared to the RF model for patients at high-risk thresholds. For the XGBoost model, the top 3 predictive features for weaning failure included congestive heart failure, atrial fibrillation, and diabetes mellitus.
Conclusions: We were able to successfully classify patients into one of four groups according to the WIND classification. XGBoost and RF demonstrated high precision in predicting patients at risk of weaning failure.
Clinical implications: Although the WIND classification has demonstrated promise in classifying ventilated patients into meaningful groups, there are a lack of reliable tools that aid clinicians in predicting which patients are prone to weaning failure upon admission. We have designed a clinical decision support tool that will identify patients at risk of weaning failure upon ICU admission to allow for implementation of appropriate early interventions to reduce patient morbidities associated with weaning failure. Future prospective multi-institutional studies are needed to validate the utility of this tool within the clinical domain.
We developed algorithms to automatically assess ocular torsion patterns based on both fundus images and oculography videos, using deep-learning based image classification, image segmentation, and video classification models. The feasibility of our project is supported by established deep-learning based models that can perform image segmentation on fundus images and video motion detection on oculography videos. With data synthesis, augmentation and model transferring, ocular torsion can be well detected from fundus images. The model performance can be further improved when more image data becomes available for training and fine-tuning. Label generation method for video classification is reliable, but more machine learning training and validation is needed in the future.
Epilepsy affects 50 million people globally. 30-40% of patients with epilepsy have medically refractory epilepsy (MRE), which cannot be completely controlled by currently existing drugs and whose treatment accounts for 80% of the $16 billion spent on treating epilepsy annually. For these patients, first-line treatment is the surgical removal of the epileptogenic zone (EZ), but this requires accurate identification of the EZ. This remains a challenge, with surgery success rates ranging from 30-70%. The current gold standard for identifying EZ is to passively capture and analyze a plethora of neurophysiology data, a largely manual process that can take over two weeks.
We are engineering an innovative machine-learning tool trained on cortico-cortical evoked potential (CCEP) data actively measured from single pulse electrical stimulation. This project aims to identify and validate a computationally-derived CCEP-based marker to accurately identify areas of the brain likely to contain the EZ, suggest possible EZ regions that were not identified by conventional analysis, and predict successful or unsuccessful outcome after resection surgery.
To develop our CCEP computational marker, we have trained a random forest model on a combination of CCEP stimulation, response, and network-extracted features from 19 MRE patients that had successful surgical outcomes. Our model performs at a sensitivity of 92% and specificity of 75%. It can identify the EZ agnostically across brain regions as well. We envision this model integrating into the clinical workflow for MRE surgery. Clinicians can leverage our CCEP marker as a decision-support tool to increase EZ localization accuracy and improve surgical outcomes. Future work involves model evaluation within the clinical workflow, with some preliminary results on patient cases shown today.
This is the poster that showcases the abstract and evidence
Project Poster for Presentation
This poster demonstrates my contributions to the Galaxy for Anvil research project during my time in Schatz Lab in Summer 2021 and beyond.
Poster includes computational and experimental XRF spectra and design results
The Pluto Brace is an innovative design to treat Clubfoot disorder.
Univariate Independence Test based feature selection methods have been the standard for transforming datasets prior to classification. Wrapper based transformers that utilize classifier estimators are also highly effective at transforming datasets prior to classification with the same estimator. However, univariate based methods fail to capture multivariate dependencies, and wrapper methods are highly effective mainly when paired with the appropriate classifier; I propose a novel transformer that uses a user-specified Multivariate Independence Test as a wrapper object and provide specific examples of when this transformer can be optimally applied.