John Pace, PhD

Project Goal: To use BERT, a state-of-the-art natural language processing (NLP) algorithm, to allow users to ask free text questions about medical records and obtain correct answers.

Background and Project Description: In healthcare, specific patient data is often difficult to find. Chart notes written by doctors or nurses are unstructured and can be voluminous. With patient records easily being 50 pages or more, it is time consuming to manually look through the records to find the… Show more

Project Goal: To use BERT, a state-of-the-art natural language processing (NLP) algorithm, to allow users to ask free text questions about medical records and obtain correct answers.

Background and Project Description: In healthcare, specific patient data is often difficult to find. Chart notes written by doctors or nurses are unstructured and can be voluminous. With patient records easily being 50 pages or more, it is time consuming to manually look through the records to find the answer to a question. In this project, I am utilizing BERT, a state-of-the-art NLP algorithm, to allow users to query the free text of chart notes and obtain answers quickly. Due to the unstructured nature of the data, significant pre-processing has to be done, such as dividing the notes into small chunks, or passages, that can be queried quickly. Additionally, BERT, by design, can only query small passages, so the data must be iteratively queried then the output data post-processed to display the correct answer.

Languages and Frameworks: Python, Tensorflow

AI and Machine Learning Techniques: Natural Language Processing

Code on GitHub: https://github.com/pacejohn/BERT-Long-Passages

Medium: https://medium.com/@johnpace_32927/question-and-answer-for-long-passages-using-bert-dfc4fe08f17f Show less

Project Goal: To develop an AI application that will transcribe American Sign Language (ASL) onto a screen so a non-ASL speaker can understand what is being signed.

Background and Project Description: Many deaf people communicate using American Sign Language. They are able to understand what a non-deaf person is saying by reading lips. However, a person who does not speak ASL cannot understand what is being signed. Thus, communication only occurs one way. The goal of this project is to… Show more

Project Goal: To develop an AI application that will transcribe American Sign Language (ASL) onto a screen so a non-ASL speaker can understand what is being signed.

Background and Project Description: Many deaf people communicate using American Sign Language. They are able to understand what a non-deaf person is saying by reading lips. However, a person who does not speak ASL cannot understand what is being signed. Thus, communication only occurs one way. The goal of this project is to allow for two-way communication. I am developing an application in which a person can use their phone camera to see a person who is signing. As the signs are formed, it will transcribe the signs onto the screen, allowing the person to understand what is being signed. This will allow for two-way communication. This project was showcased at the AI Summit Conference in San Francisco in September 2019 and will be showcased at the Super Computing 2019 conference in November 2019 and the AI Summit Conference in New York City in December 2019.

Languages and Frameworks: Python, Tensorflow, OpenCV, IBM Power AI Vision

AI Techniques: Computer Vision, Deep Learning (Convolutional Neural Networks, Recurrent Neural Networks, Natural Language Processing)

YouTube video: https://youtu.be/gBQTo1hpM3Q Show less

Project Goal: Create a computer vision workshop to introduce basics of computer vision to diverse audiences.

Background and Project Description: I have developed a workshop that introduces the basics of computer vision and allows the attendees to perform computer vision tasks such as image classification, object detection, and segmentation. There is also an overview of convolutional neural networks and how they are used in computer vision. I have taught the workshop around the country… Show more

Project Goal: Create a computer vision workshop to introduce basics of computer vision to diverse audiences.

Background and Project Description: I have developed a workshop that introduces the basics of computer vision and allows the attendees to perform computer vision tasks such as image classification, object detection, and segmentation. There is also an overview of convolutional neural networks and how they are used in computer vision. I have taught the workshop around the country to data scientists, C-level customers, sales representatives, IT administrator, programmers, and even high school students.

Languages and Frameworks: IBM Power AI Vision

AI and Machine Learning Techniques: Computer Vision, Image Classification, Object Recognition, Image Segmentation, Convolutional Neural Networks Show less

Project Goal: To train convolutional neural networks to be able to read very “messy” text from images or video, transcribe the text, and output to usable formats.

Background and Project Description: “Messy” text is defined as text within images that is typically poorly read by other image recognition software, due to varying background, shadows, dirt, obstructions, numerous challenging fonts, and weathering. Tombstones are a massive dataset of objects with "messy" text. Websites like… Show more

Project Goal: To train convolutional neural networks to be able to read very “messy” text from images or video, transcribe the text, and output to usable formats.

Background and Project Description: “Messy” text is defined as text within images that is typically poorly read by other image recognition software, due to varying background, shadows, dirt, obstructions, numerous challenging fonts, and weathering. Tombstones are a massive dataset of objects with "messy" text. Websites like genealogy.com, ancestry.com, and findagrave.com allow people to upload pictures of tombstones, but they must be manually transcribed. In this project, I have developed several convolutional neural network models that will read images or video of tombstones and transcribe the letters into text. This speeds up transcription time and helps to reduce errors caused by human data entry of the information. This project was showcased at the IBM Think conference in San Francisco in February 2019. To see examples of the text recognition, see my YouTube videos at https://www.youtube.com/watch?v=RwOCFxAszfc&list=PLz8DxxKmiM0BLK9YWfxuGeX6-8yrsO7_B

Languages and Frameworks: Python, Tensorflow, IBM Power AI Vision, OpenCV

AI and Machine Learning Techniques: Convolutional Neural Networks
Show less

Project Goal: Predict blood glucose levels using historical data for training.

Background and Project Description: Predicting blood glucose levels for individuals with diabetes is important when trying to determine the amount of insulin that must be given at various time points. Insulin pump manufacturers have to train models to predict how much insulin a pump should give to patients. In this project I used historical data to train both deep learning and classical machine learning… Show more

Project Goal: Predict blood glucose levels using historical data for training.

Background and Project Description: Predicting blood glucose levels for individuals with diabetes is important when trying to determine the amount of insulin that must be given at various time points. Insulin pump manufacturers have to train models to predict how much insulin a pump should give to patients. In this project I used historical data to train both deep learning and classical machine learning time-series models. For the deep learning part, I used a type of recurrent neural network called an LSTM (Long Short-Term Memory) for training and inference. For the machine learning part, I used SARIMA. For more details, see my blog post at https://www.ironmanjohn.com/home/time-series-benchmarks-between-the-ibm-power-9-ac922-server-and-the-nvidia-dgx-1


Languages and Frameworks: Python, Tensorflow, Keras

AI and Machine Learning Techniques: Recurrent Neural Networks, LSTM, SARIMA, Time-Series Analysis
Show less

Project Goal: To use historical data to create a dashboard that allows data driven prediction of which tires to use during an Indycar race.

Background and Project Description: In Indycar racing, all of the cars are almost identical, so no one team has an advantage due to the physical components of the car. In a race, a car must use 2 different types of tires, but it is up to the team to decide which tires to use at which times. Proper tire selection can be the difference between… Show more

Project Goal: To use historical data to create a dashboard that allows data driven prediction of which tires to use during an Indycar race.

Background and Project Description: In Indycar racing, all of the cars are almost identical, so no one team has an advantage due to the physical components of the car. In a race, a car must use 2 different types of tires, but it is up to the team to decide which tires to use at which times. Proper tire selection can be the difference between winning and not winning. In this project, I used historical data to create a dashboard of tire usage over previous years. The team was able to find patterns of tire usage by winning teams that they had not seen before. This allowed them to make data driven decisions about which tires to use during the race and at what time. The data was in the form of dozens of log files that had to be significantly manipulated to extract useful information. The cleaned data was imported into Microsoft SQL Server which was the data source for the Microsoft Power BI dashboard.

Languages and Frameworks: Python, SQL, SQL Server, Microsoft Power BI Show less

Project Goal: To determine the likelihood of a patient being readmitted within a given time frame.

Background and Project Description: Patient readmission is a significant challenge in the healthcare industry. Not only is it potentially harmful to patients, it is costly to healthcare institutions. In this project, I used classical machine learning techniques, such as XGBoost and logistic regression, to determine the likelihood of a patient being readmitted for the same condition… Show more

Project Goal: To determine the likelihood of a patient being readmitted within a given time frame.

Background and Project Description: Patient readmission is a significant challenge in the healthcare industry. Not only is it potentially harmful to patients, it is costly to healthcare institutions. In this project, I used classical machine learning techniques, such as XGBoost and logistic regression, to determine the likelihood of a patient being readmitted for the same condition within a 30-day time frame. Patient data came from the de-identified MIMIC-III database. This project was particularly challenging in that the data for patients was often incomplete, miscoded, or missing. As a result, a large amount of data manipulation and cleaning was required before analysis could begin.

Languages and Frameworks: Python, R, MySQL

AI and Machine Learning Techniques: Logistic Regression, Decision Trees, Random Forests, Clustering, XGBoost
Show less