FastTrackFDA

Partners: Natasha Maniar, Prabhakar Kafle

​

FastTrackFDA  addresses the critical challenge faced by medical device startups: the time-consuming and expensive FDA approval process, which has remained unchanged despite the rapid development of software-enabled devices. FastTrackFDA simplifies this journey from device conception to FDA clearance, offering an intuitive interface that reduces dependence on costly consultants. It features advanced technologies including a custom PDF extractor for 510k summaries, matrix similarity algorithms for identifying predicate devices, and AI-powered tools for generating clinical trial designs. Our team tackled challenges like diverse PDF formats and deployment issues, resulting in a versatile tool that streamlines the regulatory pathway, reduces approval times, and cuts costs. Our solution, built with a Next.js frontend, Flask backend, and supported by OpenAI, Pinecone, and other APIs, is a testament to the potential for technology to transform medical device market entry, making lifesaving innovations accessible sooner. FastTrackFDA's future includes expanding its scope to cover more aspects of the FDA process and continuously updating its dataset to reflect the latest device approvals, demonstrating an ongoing commitment to improving the regulatory experience for medical device innovators.

Scrabble Playing Robot

Partner: Rumaisa A. 

A single-player Scrabble-playing robot that leverages suction technology to manipulate game pieces with precision. Our research, focused on the complex manipulation and perception challenges presented by board games, culminated in a system capable of solving Scrabble puzzles, handling the pieces to play the game, and refilling the player's tile rack. The project involved modeling the internal state of the game, the hardware involved, and employing a finite state machine for the game's simulation loop. All the game pieces were designed in Fusion 360. The end effector is a suction gripper that mimics the dynamics of a real-world suction mechanism. To perform the game logic, the game utilizes the LeafSystem in Drake. Through our work, we not only demonstrated the robot's ability to engage in the game of Scrabble but also laid the groundwork for future exploration in automating other board games like chess and checkers, highlighting potential improvements and the incorporation of a multi-player mode.

Augmenting Expert Domain Image Inputs for Enhancing Visual Language Models Performance

This blog post explores enhancing visual language models, particularly for expert domains like scientific literature, where standard models struggle. By integrating domain-specific knowledge and advanced image embeddings, the research aims to refine the performance of visual language models such as OpenFlamingo. Leveraging graphical structured embeddings and graph neural networks, the study tests different methods of representing images to improve the models' interpretive capabilities.

A Novel Approach for Detecting High-Impact Words Using Sentiment Analysis Models

This blog post explores enhancing visual language models, particularly for expert domains like scientific literature, where standard models struggle. By integrating domain-specific knowledge and advanced image embeddings, the research aims to refine the performance of visual language models such as OpenFlamingo. Leveraging graphical structured embeddings and graph neural networks, the study tests different methods of representing images to improve the models' interpretive capabilities.

Enhancing Image Deblurring with ControlNet: A Novel Approach to Privacy Protection

Partner: Raz Gaon

​

Deep generative models, particularly diffusion models, have emerged as a powerful tool in image synthesis, demonstrating impressive performance across various applications. Building upon this, ControlNet was developed to refine output constraints and enhance precision in tasks, eliminating the need to retrain large diffusion models. In this study, we propose a novel ap- plication of ControlNet, fine-tuning it to generate clear, sharp facial images from blurred facial image inputs. This approach is particularly promising for privacy protection, as it enables the replacement of individuals in public settings with synthetic faces, preserving the origi- nal image’s characteristics while safeguarding individual identities. Our findings indicate that this approach offers substantial potential in the realm of privacy protection, while maintaining the integrity of image characteristics.

Development of a Machine Learning Algorithm to Predict the Path of Joints for Gait Rehabilitation

Advisor: Dr. Anurag Purwar​

​

Using a public gait data set from 42 individuals, we trained a supervised learning algorithm to trace the paths of the ankle, knee and hip joints. The ankle, hip and knee flexion and extension angles with the demographic information of the participant were used as the training and testing data. The goal was to obtain a prediction model that could attain the desired trajectories for a person's gait cycle based on their demographics. With a neural network of one hidden layer, a relationship between the input data and the trajectories was sought and found. The reconstruction loss was the quantitative analysis to gear the algorithm towards optimization with the Adaptive Moment Estimation (Adam) Optimizer. With several trials and configurations of the data, the desired paths were created after 1.2 million iterations; there was a training loss of 9.2 and testing loss of 8.9. The algorithm was validated using a random participant from the public dataset, an individual with Parkinson’s disease and the first author’s data. In the end, a mechanism was developed to demonstrate the practicality of the algorithm and the projections it had predicted for the student researcher

Exploration into the Housing Crisis of the Greater Boston Area

Boston has a severe housing crisis. Recently, many articles tell the story of buildings being converted into condos, and residents being displaced to find new places. I was interested in learning more about the history of these condo conversions and understanding how they impact Boston's population. An initial glance at the dataset reveals potential gaps in data collection for certain post years. Additionally, there may be default or null values that impact the cleanliness of the data. Lastly, sanity checks were completed to ensure that prior and post property conversions made sense. Through the deeper analysis of condo conversions in Boston, we notice a significant shift in the city's housing landscape. A large portion of residential buildings containing two to three families were particular targets for conversions. Older buildings, properties built in the early 1900s, were part of a significant number of conversions. A geographic analysis revealed that conversions were more frequent in certain zip codes (02127, 02130, 02131), suggesting a pattern of targeting areas with potentially lower-income households and more affordable housing. To examine the financial incentives that might be available for property conversions, I found that there was mostly an increase in building, land, and total valuation when a conversion happened. A shift towards condominiums has altered Boston's residential areas, potentially affecting the city's socioeconomic diversity.

Condo Conversions in Greater Boston

Partners: Abigail Klein, Stephen Wilson, and Caroline Cunningham

​

This project explores the impactful trend of condominium conversions in one of the U.S.'s most expensive housing markets, Greater Boston. Through detailed, interactive maps and visual data, users can explore how condo conversions from 2015 onwards have affected various zip codes in Boston and Cambridge. The visualization segments into different perspectives: the age of buildings most likely to be converted, the types of buildings undergoing conversions, and trends in condominium-related complaints. This project aims to shed light on the pressures that drive housing prices up and the displacement of residents, providing a platform for users to visualize and understand the dynamics at play in their own communities.