Co-Founder

Millions of medical images are captured annually for clinical trials. Many trials collect images from multiple sites across the world. These all must be standardized, de-identified to protect patient privacy, put through quality control, annotated, and evaluated. Most of these processes are performed manually. This contributes to longer, more expensive clinical trials and delays effective treatments from reaching the patients who need them. It's also error-prone, potentially threatening patient safety and privacy. Some of the image processing tasks that should be done cannot feasibly be done by humans at the scale needed for Phase III trials (big clinical trials with many patients). I co-founded Saliency to address these challenges. We automated the medical image assessment tasks previously done manually for clinical trials and have introduced patient privacy protection tools that were previously non-existent in the industry. We participated in the StartX and Plug And Play accelerators early on, each with an acceptance rate under 5%. Saliency was acquired by Clario in 2020. Clario is a clinical research company whose technology platform is used in approximately 70% of all FDA approvals. I then served as their Director of AI.

Close Project

Co-Founder

Gosset enables optimal capital allocation in the biotech industry. We leverage AI and novel data sources to aid in investment research, catalyst event preparation, M&A valuations, deal sourcing, and competitive intelligence. Our products inform portfolio management decisions in pharmaceutical companies and investment funds.

Close Project

NBA Basketball Health Analytics

Created a new metric to assess basketball playing style and assessed how playing style affects NBA players' injury risk and their ability to return to play.

ACL tears are season-ending knee injuries that have had a major impact on the competitive landscape of the NBA. I co-led a study in collaboration with a team of orthopedic surgeons at Stanford on how athletes' style of play affects their ACL injury risk. We found that players who drive the ball to the basket more frequently have a higher risk of ACL tears, but those who make it back to the NBA after rehab perform at a similar level to matched controls in the remainder of their careers.

To achieve these insights, I created a new metric of playing style - Driving Tendency - that leverages player tracking data to identify how frequently a player drives to the basket per minute on the court after controlling for their playing time and utilization. This allowed us to determine how driving affects ACL injury risk irrespective of how much playing time a player gets and how often they get the ball. I also developed a new method for measuring the similarity between players in order to match ACL-injured players with controls in a nuanced, objective fashion.

This work was covered by several media outlets, including Sports Illustrated.

Close Project

Automatic Assessment of Knee Osteoarthritis in X-rays and MRIs using Deep Learning

Knee osteoarthritis is a debilitating disease that involves inflammation and degradation of the knee. It affects over 250 million people and has no cure. Research to improve our understanding of this condition relies on X-rays and MRIs. However, evaluation of these images by physicians and researchers can be subjective and requires significant time and expertise. This creates a bottleneck that hinders research progress. With colleagues at Stanford University's Neuromuscular Biomechanics Lab, I developed, validated, and deployed neural networks to automate the assessment of osteoarthritis in X-rays and MRIs.

We developed a model to automate the assessment of osteoarthritis severity from X-rays and compared its performance to that of musculoskeletal radiologists. Saliency maps were generated to reveal features that the model used to arrive at its disease severity assessments (as in the left image above). The model agreed with the consensus of a musculoskeletal radiologist committee as closely as individual musculoskeletal radiologists agreed with the committee. It predicts osteoarthritis severity scores with state-of-the-art accuracy. Saliency maps suggested the model’s predictions were based on clinically relevant information. We published a manuscript on this work in Radiology: Artificial Intelligence. The model code is available as a docker on Github and we've deployed it as a user-friendly web app. By open-sourcing the model, we've allowed others to explore its utility in research and healthcare. For example, researchers at the Veterans Affairs Medical Center have explored its potential for predicting outcomes after knee replacement surgery.

We also developed a model to automatically find the cartilage in knee MRIs (as in the right image above) and to evaluate the health of the cartilage with high spatial resolution. Assessments of cartilage health using the model agreed with those of an expert as closely as experts agreed with one another. We published a manuscript on this work in the journal CARTILAGE. The code for this model is also available as a docker on Github and we also deployed a user-friendly web app version of this model.

I have a 30-minute webinar on YouTube in which I summarize this work.

Close Project

Real-time, automated gait retraining for rehabilitation

Knee osteoarthritis affects 20-40% of individuals over 65 years of age. Knee mechanics plays a central role in the disease. In other words, the style in which an individual walks can affect the rate at which their osteoarthritis progresses. By teaching individuals a new walking pattern, we may be able to reduce their pain and slow the progression of their disease. Unfortunately, current gait retraining protocols require continuous measurement of patients' knee biomechanics in order to give them feedback. This requires the use of a laboratory full of expensive equipment which is incompatible with routine care in typical clinic. Therefore there is an urgent need for simpler methods to quantify the "load" that individuals place on their knee joint as they walk.

In collaboration with colleagues at the Stanford BioMotion lab, I developed a method for predicting knee loads using only measurements of plantar pressure (the pressure between one's feet and the ground). This method is highly attractive because knee loads could be estimated just by having the patients walk on a pressure mat in their regular clothes, without supervision or preparation time. I presented our work at the International Society of Biomechanics Congress in 2019.

An even easier alternative to plantar pressure measurements would be to simply record a patient's walking patterns using a smart phone camera and estimate their knee loads from the video using a neural network. I worked with colleagues in the Stanford Neuromuscular Biomechanics lab to explore whether this is possible. We simulated a situation in which markerless motion capture algorithms were used to track individuals' joint locations in standard videos and then demonstrated that this joint location data can be used to analyze movement patterns and estimate knee loads. We published a manuscript on this work in the journal Osteoarthritis and Cartilage in 2021.

Close Project

Philadelphia Phillies Major League Baseball Team

Interned as a Quantitative Analyst in the team's R&D Department

During my internship as a Quantitative Analyst, I used a data-driven approach to identify strategy adjustments that pitchers could make based on their personal strengths to improve their outcomes. I then developed a tool that allowed the team to easily measure the statistical similarity between any two players of the same position using the full performance history of each player. I also conducted diligence on biomechanics products and services being considered by the team.

Close Project

Fulbright Scholar

Conducted sports engineering research in Canada in collaboration with a world-renowned research group and Adidas

The Fulbright Program funds U.S. Scholars to conduct research abroad as a means of cultural exchange. I chose to go to Canada because the University of Calgary Human Performance Lab (HPL) in Alberta, Cananda is world-renowned for its sports engineering research. I spent one year at the HPL developing methods to monitor athletes' technique for sports-specific tasks using wearable sensors. I utilized inertial measurement units, electromyography, and machine learning algorithms to characterize elite athletes' technique for basketball free throws and sprint starts. I presented this work at the International Society of Biomechanics Congress. The project was funded in part by an NSERC grant that promoted collaboration between academia and industry. I worked with Professor Benno Nigg at the University of Calgary and executive R&D leadership at Adidas.

Close Project

Vertical Jump Biomechanics @ EXOS

Lorem ipsum dolor sit amet consectetur.

Use this area to describe your project. Lorem ipsum dolor sit amet, consectetur adipisicing elit. Est blanditiis dolorem culpa incidunt minus dignissimos deserunt repellat aperiam quasi sunt officia expedita beatae cupiditate, maiores repudiandae, nostrum, reiciendis facere nemo!

• Date: October 2019
• Client: Window
• Category: Photography

Close Project