REVEAL

This project aims to develop a non-surgical diagnostic method for use by practitioners in the diagnosis of Eosinophilic Esophagitis (EoE) in patients. Current industry standards require surgically collected biopsy for diagnosis, so this project explores a machine learning capability to improve practitioner understanding and confidence in a potential EoE diagnosis based on patient symptoms.

Given that EoE typically has a "median diagnostic delay time of 6 years" due to practitioner concern to perform the surgical diagnostic procedure, patients are left suffering and have lower prognosis for recovery. REVEAL (Rapid EoE Variant Evaluation AI Learning) is a machine learning model that determines the probability of EoE based on symptoms/patient history input. 

While the prevalence of EoE is rare, studies into effective and safe non-invasive diagnostic options would not only improve the quality of life in EoE patients, but also open the door to similar exploration of diagnostic options within related disorders.

Eosinophilic Esophagitis (EoE) is a chronic, rare disease of the esophagus which currently has limited diagnostic and treatment options to those who suffer despite spanning all age groups and genders. The only current diagnostic option following clinical signs of EoE is an endoscopy with corresponding biopsy. Due to a combination of hesitations as a consequence of invasiveness, cost, and other factors, it has been found that “typically EoE is not diagnosed immediately after symptom onset with a median diagnostic delay time of 6 years” despite the fact that delays in diagnosis have been shown to lead to “increasing rates of stricture” and other significant symptoms, causing an overall declining prognosis for patients. 

One of the struggles with EoE diagnosis is that the clinical symptoms patients present are wide ranging and similar to many other gastrointestinal disorders. As early diagnosis is key to improving EoE prognosis, improvement in diagnostic options could lead to a lower timeline between initial symptom onset and diagnosis for patients. Studies into effective and safe non-surgical diagnostic options would not only improve the quality of life in EoE patients, but also open the door to similar exploration of diagnostic options within related digestive disorders.

To augment the diagnosis process, the goal is to develop a machine learning model (REVEAL) that can determine probability of EoE based on patient provided symptom input and patients clinical history data input. Machine learning can help improve practitioner confidence in EoE diagnosis or progression to biopsy collection to support diagnosis. Since machine learning models require a large repository of data, this project will need to work with patients and doctors to collect sufficient data for both EoE and other GI diseases to train and test a model. Additionally, various models need to be tested to determine which unsupervised or hybrid model works best for this type of data. As a newer field, machine learning often requires trial and error to improve the model and understand which patient symptoms are most helpful in determining likelihood of EoE. In previous medical applications of machine learning for medical diagnosis, the use of “casual learning” was shown to improve diagnostic accuracy, so developing the code to support this based on previous research will be explored. Overall success for the purposes of this project would be at least 50% classification accuracy to improve practitioner confidence. 

According to Apfed, "Eosinophilic esophagitis is a rare disease, but increasing in prevalence with an estimated 1 out of 2,000 people affected." It has been found that EoE affects those of all ages and genders, although it occurs at a higher frequency in males than females. As of 2019, there are “no Food and Drug Administration approved drug[s] for the treatment of EoE,” leaving patients subject to experimental treatment options with unknown safety. Given the rare nature of the disease, research funding is typically limited and gathered by special organizations like Apfed. Due to this, most of the research has focused on treatment options rather than improvement on diagnostic options that are less invasive, costly, and painful to the patient. Research has shown that patients are interested in non-surgical options for EoE diagnosis, even if they are less accurate. This solution offers a revolutionary ability to predict the likelihood of an EoE diagnosis completely using machine learning capabilities which would have significantly lower cost, be completely non-invasive and non-surgical, and would lend practitioners the ability to wait less to pursue a potential EoE diagnosis. During the development process, patients and practitioners would be engaged to test out the REVEAL tool and provide feedback to improve it, ensuring buy in from the patient and medical community.

This project directly aligns with the first dimension of the challenge to utilize big data to improve diagnosis of rare diseases. REVEAL is a machine learning tool that aids practitioners in deciding whether to pursue any additional diagnostic testing to determine whether a patient has EoE. It also is connected to the goals of improving the lives of patients who are typically underserved within funding and research by the FDA and connecting patients and practitioners to the process of developing new tools.

As a current graduate student at Johns Hopkins University, this concept was developed concurrent with my studies. It was initially developed for submission as an grant proposal for funding and is also being submitted for this contest to secure funding. 

The REVEAL tool takes a new approach at improving the diagnostic experience for patients who may or may not have EoE. The current accepted approach requires a surgically collected biopsy for diagnosis, and most other researched approaches still require some invasive collection of biopsy to diagnose an individual with EoE. The ability to diagnose, or at least accurately predict, patient EoE without any invasive procedure will be catalytic. This would fundamentally change the cost and time involved in such a diagnosis, allowing this to be a simple test performed in a regular doctor's visit rather than a special outpatient procedure. If successful, it could also be expanded to other diseases, improving the lives of patients and lowering the cost of diagnosis for countless people.

The core technology of REVEAL is machine learning. Given the novel nature of the approach, testing is still necessary to determine what machine learning model is best for the given data within the unsupervised and hybrid models that exist currently. Based on research of previous machine learning applications within the medical field, use of "casual reasoning" will be incorporated to improve outcomes within the chosen unsupervised or hybrid model. 

This technology has not been used in EoE to date, but has been explored in other medical applications. As discussed in Richens et al., most machine learning gives a diagnosis that is equivalent to an average doctor, but their use of “casual reasoning” improves the diagnosis to the top 25% of doctors or an expert diagnosis. Further exploration is needed to determine if this is good enough for use within the EoE community, but it shows promise to be able to remove surgical requirements for EoE diagnosis. This technology has not been explored to date in EoE diagnosis, but it shows large potential to improve diagnosis delay in patients, especially by practitioners not as familiar with the disease.

Richens, J., Lee, C., & Johri, S. (2020, August 11). Improving the accuracy of medical diagnosis with causal machine learning. Retrieved March 30, 2021, from https://www.nature.com/articles/s41467-020-17419-7?6598

Firstly, if the proposed methods are not as sensitive or specific as current surgical diagnostic methods, this would hamper incorporation into clinical practices. Secondly, the data repository needed to train and test a machine learning model will likely be difficult to gather. This is both because EoE is a rare disease and the fact that collection of medical data in general requires patient sign off. Additionally, since other GI disease data is needed, the repository will likely be quite large and may require supercomputing or other large GPU access to run. Finally, patient and practitioner buy-in is critical to the success of the project both in clinical studies and through implementation in industry. If there is mixed success from one or both parties, this will slow down the ability for the technology to roll out to patients.

Even if the proposed method is not as sensitive or specific, this may be offset by the level of patient and/or practitioner buy in. By networking and advertising the study, the amount of data collected to train and test the machine learning model can be improved. This can be done through pamphlets in doctor’s offices or advertisement on social media. Supercomputing access is potentially available for this project using MARCC as a Johns Hopkins University resource. Finally, feedback from patient surveys can be incorporated and the proposed methods can be iterated on for improvement either during this or future projects. 

REVEAL does not currently serve any patients because it is still in the early stages of development. Within one year, with funding assistance, the model will hopefully be built and tested to achieve the initial goal of 50% classification accuracy. At this point, it would be ready for initial clinical testing and could positively impact one or two dozen people. However, at five years, it could be helping thousands of people since the technology will be tested and matured and would be ready for clinical use. Also, if the technology is able to be diversified towards other diseases for diagnosis, the impact could be closer to millions at the five year mark.

Within one year, the primary goal is to develop and test a prototype of REVEAL that achieves the initial goal of 50% classification accuracy towards the probability of EoE within a prospective patient. Looking beyond that, the goal is to develop a software that can be sold to GI practitioners to improve diagnosis of EoE. This software will be iterated upon to improve the accuracy to one day be fully competitive with other surgical diagnostic options in terms of sensitivity and specificity. 

There are many near terms goals for the REVEAL project. The first is the identification of a viable machine learning technique that processes this medical data well and can be scaled for use in a large dataset to identify potential EoE patients. The second goal is to successfully collect the necessary data to train and test the machine learning model by networking with doctors and collecting (and securing) patient data of a wide variety of GI diseases. The third near term goal is the successful development of the REVEAL tool to achieve a 50% classification accuracy based on patient data from previously diagnosed patients. The final near term goal is to perform a study utilizing the REVEAL tool to gather and adjudicate patient and practitioner information to improve the first generation version of the tool for initial release to the medical community.

The solution team is currently comprised of two people. Amanda Gelbart is the solution team lead and is a current graduate student at Johns Hopkins University. Wendy Gelbart is a entrepreneur who has been a successful small business owner for over 20 years.

Amanda is the technical expert in the team. As a graduate student at Johns Hopkins studying Biomechanical Engineering, she brings the engineering and technical experience needed to bridge the medical and engineering community together to deliver this solution. Her partner, and mother, Wendy brings the business knowledge to take this idea to the next level. As a small business owner for over 20 years, Wendy has the experience in running a business, managing external funding, and marketing to improve outcomes for the business. She also has experience within Academia, as a published author in the field of Educational Technology, researching and bringing a project from idea to reality.

Beyond the technical qualifications this team has, we understand what it is like for patients to struggle to receive a proper diagnosis. Due to being under-insured for most of Amanda's childhood, Wendy knows the struggle to be the parent of a child in pain, looking for the answers. Doctors visit were expensive. Testing was even more financially straining. This is why it took Amanda over a decade to receive her EoE diagnosis from initial symptoms that were deemed significant by her childhood doctor. As a mother/daughter team, we bring compassion to the project, understanding just how difficult it is to wait to find the answers and help your child feel better and just get back to being a kid.

As a female led team, we are focused on making sure equity is first in our leadership model. In addition to serving the under-served community of patients with rare disease in this project, we aim to lift up under-represented technical experts as the project grows. 

Yes the team lead, Amanda Gelbart, has Eosinophilic Esophagitis. She was diagnosed at 17 after battling GI issues for her entire life. Lack of access to proper medical care due to being under-insured prevented diagnosis earlier in life.

As a rare disease patient who suffered for many years to get a diagnosis and answers to why I was constantly in pain and "different," I understand just how terrible it can be to not have the answers. I have suffered, so I want others to not have to. This project is driven by my passion to help others improve their lives by shortening the time to diagnosis of EoE. The Horizon Prize can help take my idea to the next level not only through funding but also the potential mentorship beyond what my solution team currently has the knowledge and experience with in this field. 

While our solution team has experience in running a business or developing new technologies, there are some key aspects we are not experts in and could use the help of the Horizon Prize team. First off, we are not legal experts, nor do we have any legal experts currently on our team. Given the sensitive nature of medical data, this is key to ensuring regulatory hurdles are properly navigated. Next, the development and distribution of the REVEAL tool could be assisted via the Horizon Project team given the vast network that exists within the research and industry communities you have. 

We would like to partner with the American Partnership for Eosinophilic Disorders (Apfed) and the National Organization for Rare Disorders (NORD) given their extensive history in funding research and technology to improve the lives of those with EoE and rare disorders in general.