Optimizing research and development of A.I. models

The scarcity of available data for rare diseases (RDs) complicates health conditions that are difficult to manage at several levels. This leads to having an extensive diagnosis timeline of 5-8 years which ultimately worsens outcomes for patients. This situation calls for innovative solutions via quantitative and automated tools which focus on improving the datasets available in predictive models. More training data on identifying biomarkers accelerates time to diagnosis.

Gradient Health, Inc has created a platform that provides instant access to a vast database of ethically sourced medical imagery. The goal of sharing this database to interested academic institutions is to enable researchers to quickly search for the data they need, access pre-labeled images, and develop AI algorithms that are based on truly representative data. In turn, rare diseases will be more quickly recognize outliers leading to a faster diagnosis. Gradient is currently partnered with world leading research institutions such as Stanford University and Duke University who have chosen to work with Gradient due to the speed they are able to deliver large datasets and the diversity of the data.

The following is a brief description of our other accomplishments so far that will support the research of RDs:

1. Indexing and dashboard software. Most academic or imaging centers, whether in the US or abroad, are not even aware of the data sitting in their servers.  We designed a dashboard based on radiolong “findings,” a collection of terms and descriptions used by radiologists to describe the pathologies in the images (see attached). The dashboard also reflects patient age, demographics, and geographical location, among other data. Our software can be integrated into any vendor’s PACS and reporting system. The software will update in real-time and will allow for immediate creation of patient cohorts based on their radiology findings. 

2. De-identification software. We continue to refine our de-identification software. Although we anticipate that the majority of DICOM images will already be de-identified, we want to add an additional layer of privacy by detecting and redacting burns out protected health information. 

3. Labeling software. We have developed web-based image labeling software that is not only user-friendly, but encompasses the majority of tasks that machine learning engineers could want, including bounding box and segmentation. A major benefit of our software is that multiple users can label images remotely, without the need to download additional software. In addition, the newly labeled data can be transmitted seamlessly and securely to the development team. In addition, our labeling software can also function as a PACS (Picture Archiving and Communication System). Although it lacks some of the post-processing functionality seen with established companies, our software will still represent a significant step-up for certain imaging facilities in Africa.

This solution solves A.I. companies and academic researchers’ needs for large, diverse annotated medical images. This project also serves hospital systems by permitting them to have an active role in the development of A.I., improves their understanding of how AI models are deployed, provides them a PACS system and educational training for radiologists, with a special focus on rare diseases. Our database also visualizes worrisome trends that could manifest in radiology images, such as in abnormal obstetrical ultrasounds, congenital anomalies, organ dysfunction, or cancer spread that would warrant additional action by health agencies. Ultimately, at our core, is a strong desire to remove barriers to the efficient and safe delivery of healthcare. We believe that this is not only a sound business strategy, but also one that will hopefully save lives. 

The data sourced from these communities will be anonymized and labeled for use in training and validation of AI algorithms that could then be deployed for the very people whose data was accessed. This will hopefully improve the speed and quantity of healthcare delivery while providing a revenue share model that is fair and respectful. The revenue share model is necessary to incentivize hospitals to engage with the data network. Because images are obtained in a DICOM standard, these images can also be deployed in developed countries, but at a higher price point, which would incentivize well-funded industry partners to participate as well. 

We have a diverse, accomplished group of people working to solve this problem. Our founders, Josh Miller and Ouwen Huang have previously worked together on a computer vision agricultural technology company FarmShots, a satellite imagery platform that was deployed across Africa and Brazil. Dr. Sophie Chheang is Interventional Radiologist and Assistant Director of Informatics at Yale. We are currently partnered with Telelaudo, a teleradiology company that provides remote radiology reports to imaging centers and hospitals in Brazil and other Portuguese-speaking countries. Nico Addai is a Research Consultant for MIT STEP lab where she is trained in A.I. ethics development.

Gradient is applying to Solve because we recognize the scale of the problem and our proposed solution. Artificial intelligence and developments in healthcare technology will affect the whole world, and we believe that partnering with MIT Solve will allow us to have greater impact.

Private companies measure their success solely by how profitable their measures are, often times regardless of the ethical costs to get there. While Gradient also means to be profitable, our mission is based around creating a strong foundation for all of medical A.I. which is why we have opened access to academic institutions to utilize our ethically sources datasets. 

One of the major problems facing medical algorithm design is lack of access to large, diverse datasets. By focusing on acquiring medical data that represents both geographical diversity as well as ethnicity and age, Gradient Health, Inc is improving medical algorithms robustness. We use advanced natural language processing techniques to organize the images according to radiology findings, and make those images available to developers  who create algorithms that are fairly and accurately deployed for the very people that they were trained on.Gradient has a global view and continues to build data partnerships with hospital systems and clinics around the world. 

Our goal for the next 5 years is to integrate with at least 10 data sources in Africa and South America in at least 5 different countries. We are in the process of  developing partnerships with private radiology annotation companies to train radiologists and radiologist technicians remotely. We also have begin to deploy our platform and build algorithms for deployment in low resource settings. By the end of five years, our goal is to have integrated our open-source dashboard to allow anyone in the world to look up open source de-identified DICOM images from at least 30 countries. These open source documents will also include metrics on the number of images that are indexed and labeled, algorithms that are in the pipeline and-qualitative feedback on the experience for all stakeholders: researchers, industry, hospitals and imaging centers, radiology labelers. 

Our goal is to create an international research community that will have the ability to easily share and compare results using the same datasets. We plan on measuring that success based on a) how many academic institutions we are partnered with b) how many datasets are utilized in the study of pathologies and c) the success rate of prediction models created using our datasets. We are aiming to partner with 20 academic research organizations in the coming five years to greater expose our datasets and tracking their use of our datasets. 

Radiology data that has been siloed will be surfaced, contextualized, and labeled for specific use cases. Significantly, this index will be made available to any researcher or company with the desire to build useful tools that will also positively impact healthcare delivery. 

This will dramatically scale up the development of computer vision algorithms that can be deployed in environments where radiologists are scarce, which includes much of Africa and South America. The ability to detect severe, potentially life-threatening, imaging findings will not only help the community directly, but will also provide a 30,000 view to global health needs. Our database could also visualize worrisome trends that could manifest in radiology images, such as in abnormal obstetrical ultrasounds, congenital anomalies, organ dysfunction, or cancer spread that would warrant additional action by health agencies. Ultimately, at our core, is a strong desire to remove barriers to the efficient and safe delivery of healthcare. We believe that this is not only a sound business strategy, but also one that will hopefully save lives.

Beyond developing deep, mutually beneficial relationships with hospitals, we prioritize  Our core solution makes use of many aspects of artificial intelligence: automation, natural language processing (NLP), classification and segmentation.

Gradient Health, Inc is dedicated to enriching its company culture in hiring candidates from diverse backgrounds. Our team comes from a variety of backgrounds, races, religions, and other metrics of diversity. As for our technology, we focus on acquiring medical data that represents both geographical diversity as well as ethnicity and age. Gradient Health, Inc is improving medical algorithms robustness by curating large, diverse datasets. Gradient has a global view and continues to build data partnerships with hospital systems and clinics around the world. 

We are a for-profit private start-up that aims to bring together A.I. companies needs for medical data with data partners' desires to be fairly reimbursed for their contributions. We provide an annotation service for DICOM images, in addition to a DICOM viewer and access to annotated, off-the-shelf datasets.

We have two major stakeholders in our business: our data partners and our A.I. company partners. Our A.I. companies act as our customers and purchase access to annotated DICOM images or utilize our DICOM viewer to annotate and analyze their images. For our data partners, we create a revenue share model with per utilized image in addition to offering a free PACS system for hospitals in developing countries to store their hospital information.

Gradient Health is currently generating revenue of over $100,000 per year, with a goal of reaching $500,000 in 2022. We have already raised $2.5M in seed round funding to assist in hiring talent and deepening partnerships to reach this goal.