From Rare Cases to Big Data

Patients often find their doctors and other providers have little experience with rare diseases. On diagnosis, patients may ask friends, providers may ask colleagues, or both may go to the internet and read off figures, which may be quite scary. Much like a novel or emerging disease, rare diseases often appear where clinicians have limited knowledge of or experience with the disease. 

By their very name, rare diseases do not have many cases. This may mean the burden of disease may be low but on the flip side it is harder to learn about these diseases. This makes it harder to build clinical acumen among clinicians. It is less likely that someone in a rural or low resourced health setting will have a provider who has seen many cases.

There is also a growing digital divide in the development of new tools for rare diseases. Rare diseases frequently do not have the large clinical datasets needed for big data and AI. As machine learning brings improvements to healthcare, rare diseases are expected to lose out on these new tools. 

However, many rare diseases are either a) not rare in less resourced areas around the world but are often left undiagnosed or undocumented in available data sets as less research is published or b) data, though rare in any individual setting, if collected from many locations would be sufficient to power artificial intelligence tools. 

These diseases are often infectious diseases which have been globally neglected in research. There is a need for further tools to better identify and predict outcomes and treatment needs for those with rare infectious diseases. Without this work, cases may simply spread, undetected, making rare diseases a bit more common, unfortunately.

This issue highlights the need for clinicians and researchers in low resource settings to be able to share their clinical insights for great global learning and also for the development of AI tools.

Clinical decision tools can help decrease delays in diagnosis and also help predict outcomes of patients and the need for more advanced treatment or transfer to more specialized or higher levels of care. 

Those rare diseases that are infectious diseases may be rare now, but can multiply and spread further. Per the NIH, there are about 40 infectious diseases considered rare diseases; many of these are zoonotic, vector-borne or associated with waterborne transmission or associated with crowding, all of which are expected to become more common with climate change. We will see more outbreaks, especially with the mixing of vectors and animal reservoirs and pathogens as environments are fragmented and climates change. Such amplification of rare diseases is expected to affect vulnerable populations with the least resources and historically marginalized the most. Tools we develop now will better help us respond to the coming challenges brought by climate change By neglecting rare diseases we impact the unfortunately many who have these rare diseases and the many others in the future.

We wish to build upon our existing work. We have established sustainable telemedicine care first in 2010 in Haiti and also through Radiokam, whose CEO Djinaud Prophete, has built up capacity for teleradiology have built up a means to identify and support the care of those in need of care. Much of work has focused on tuberculosis, which we will continue to work on and which itself is considered a rare disease, but we also hope to expand our approach to include diseases often referred for TB evaluation but which are instead rarer diseases, which may be overlooked to the detriment of many patients.

We have been building a network of providers and researchers in low resource clinical settings to share de-identified data from patient care. We have focused on the and have developed a prototype model for the classification of different pulmonary findings and diagnoses.

We are working with researchers in Haiti, Ghana, Nigeria, Sierra Leone, and New York City to  build a clinical database and are working to expand to create more partners. Our work has focused on tuberculosis (TB), which is considered a rare disease by the NIH in the US, but is in fact, one of the world’s deadliest diseases, second only COVID-19. Other diseases which are labeled as rare, which we wish to study further include: Leptospirosis, Chikungunya, Brucella, and Legionella, which have been concerns for ourselves and others in these settings.

We have been using computer vision to analyze radiological images, but have also worked on predictive analytics to improve predictive outcomes for patients. All of the above diseases can have cardiopulmonary involvement seen on chest imaging. The specific questions to analyze depend on each disease. For example, both Tuberculosis and Chikungunya can develop into persistent and long-term symptoms even after the main symptoms have subsided. Imaging and clinical parameters can identify those more likely to have longer symptoms (pulmonary scarring in tuberculosis; demographic data, laboratory abnormalities or myocarditis signs in chikungunya) Legionella and Leptospirosis often present themselves in mild symptoms. Timely diagnosis and treatment of these diseases are crucial to cure and prevention of long-term damage. 

We could extend our approach to other diseases which are rare, relying on existing health record data through local institutions, existing databases, and working with networks of other clinicians to collect de-identified data, both in the US and abroad (here here here here here). We can also scale our methodology (here) to larger datasets for different rare diseases, while coupling this with computer vision (X-Rays or other imaging data). Data can be used from first visit to predict risk of subsequent outcomes.

We have an existing prototype for X-Rays (here here) which can be expanded for other classifiers. We can start from our existing work for other clinical data (here), and expand to other geographic locations and diseases. We would build on our existing data collection to build a more accessible database with a data dictionary we have used for clinical settings with less health resources to include data on clinical presentations and outcomes for these specific diseases (on first visit: vital signs, labs, imaging, review of systems, demographics (avoiding Protected Health Information (PHI)), and specified outcomes for each disease).

Our approach can be used to predict different outcomes depending on the disease and ensure outcome choices are not biased by location, if different treatment options are available or if delays in accessing care differ. These outcomes can include disease recurrence, longterm pain (Long syndromes) hospitalization, higher level of care, death. We also will analyze data available for Leptospirosis and Dengue (wide range of presentation with most having mild disease and with some needing ICU, ventilation, transfusions/bleeding), Chikungunya (Long Chik, or debilitating arthritis is common and not well predicted), treatment failure (Brucella as well as Tuberculosis). 

As a result of this project, over time, we can scale this to other diseases considered rare diseases by the US NIH: Babesia, Bartonella, Blastomycosis, Chikungunya, Dengue, Infectious Endocarditis, Fascioliasis, Filariasis, Hantavirus, Legionella, Leprosy, Listeria, Meningococcemia, Mumps, Nocardia, Osteomyelitis, Paracoccidioidomycosis, Pertussis, Plague, Psicatossis, Q-fever, Rabies, Rocky Mountain Spotted Fever (RMSF), Rubella, Sennetsu fever, Syphilis, TORCH syndromes, Tuberculosis, Tularemia, Weil syndrome, Leishmaniasis, Non-Tuberculous Mycobacterial (NTM) infection, Mucormycosis, Clostridial Myonecrosis, Hepatitis D, as well as Cholera, Pinta, and Anthrax.

The databases we plan to establish could further be used by other researchers, especially in AI, to build better tools focused on rare disease analysis, and highlight some of the challenges and possibilities of AI in this domain. Our system is intended as a clinical decision support system for these less common diseases, assisting, but not replacing doctors and other providers, especially where access to specialists is limited and where transfer to a higher level of care is quite far and few have experience with the disease. 

Our existing clinical care data and methodology focus on evaluating on Tuberculosis, Legionella, Non-Tuberculous mycobacterial (NTM) infections, Osteomyelitis, Babesia, and infectious Endocarditis. The ultimate goal, however, is to build on databases and explore other, rarer diseases. The work on TB, Legionella, and NTM would include work on computer vision, using X-Ray imaging. Others would focus more on lab values, demographics, vitals, and review of systems. This approach will also equip clinicians in lower-resource settings with rich clinical knowledge to share their experiences.

Patients with the studied illnesses may be more vulnerable from a sociodemographic perspective: they are more likely to be immigrants, have language barriers, have more hurdles in accessing healthcare or lack insurance in the US. With climate change, environment fragmentation and destruction, as well as increased mobility, pathogens causing rare diseases cross geographic borders and become more resistant. In our work, we will analyze how disease affects US and global populations, and ensure that the resulting AI methodology is fair and equitable, especially with regard to more vulnerable communities (here here here). 

Those who have rare diseases do not benefit from the same collected wisdom of years of clinical practice as other more common diseases. Such illnesses by their very name, rare diseases, will not have been seen by family physicians or rural health care clinics. These diseases often affect vulnerable and underserved populations both in the US and globally, for example, immigrants, individuals with low-socioeconomic status, communities of color and indigenous populations. It is our goal to develop an equitable system to better evaluate rare diseases while also bridging digital and resource divides that have undermined medicine and rare disease research.

We are a team who has a broad knowledge of working with resource limited settings and in clinical medicine, while working on AI-based approaches.

Djinaud Prophete is the CEO of Radiokam, a successful teleradiology platform in Haiti, allowing multiple hospitals and clinics to send radiology films for evaluation remotely. This provide rapid turnaround of imaging, which may otherwise delay diagnoses, especially when it is not feasible to have a radiologist on site. Given the crises Haiti has been facing, from an upswing in kidnappings of clinical staff and violence, as well as disruptions from strikes, protests, earthquake in the south of the country, and COVID, off site clinical care, as possible, provides a robust means to maintain continuity for necessary medical care.

Ti Kay has also built on the use of technology to help improve health outcomes in Haiti. The organization started its work in 2010 after the earthquake in Port-Au-Prince. With a team of very dedicated Haitian nurses, Ti Kay an an in patient and outpatient TB unit seeing up to 1000 patients a year from 2010-2014. The work since 2010 has included homebased telemedicine care for those with TB and HIV, who cannot easily access care in clinics because of the severity of their illness and other hurdles. TB itself is considered a rare disease in the US and we will continue to support those providing care for TB. We also have seen many patients referred for TB evaluation whose diagnosis was either confirmed or likely a rare infectious diseases often overlooked or blanketed by more common diagnoses. We hope to expand work for those whose illnesses are often even less able to access care because of the rarity and lack of resources. 

We have rich network of clinicians working in low resourced settings who we hope will be able to share their valuable their clinical knowledge and increase their own learning. We have already met and discussed with doctors and teams in Ghana, Nigeria, and Sierra Leone to expand our insight, as well as the database and potential clinical tool reach. We want to give the opportunity to other clinicians and have tools for handling diseases they are less familiar with and for patients who are suffering from these diseases that are rare and often neglected are able to access the best care, guided by the best evidence.

All of our team members on this prototype have worked on developing AI solutions. The collaborative work, dating back 2 years, has allowed us to share knowledge between those with more clinical or more computer science backgrounds, ensuring tools are effective and useful and never merely an experiment.

We would like to publicize the use of these open databases to bring in more de-identified data from clinicians and further partners. We can establish a larger network of providers and researchers, allowing for the sharing of clinical insight and collected data, and elevate the research of those from lower resource settings.

We can benefit also from continuing discussions on legal perspective of medical aid tools, which can be interpreted differently in different countries.

We would also need more funding to focus more on this work.

We also have the longstanding track record of working with technology as an effective solution where hurdles have often prevented other approaches from working.

While the number of AI-based disease analysis applications is large and increasing to date, there exist relatively few benchmark datasets that are publicly available that present real-life scenarios of imaging and patient data that focus on rare diseases. As a result, it is difficult to understand whether the AI based methods work for these diseases. We therefore will bridge this gap by collecting data worldwide representing diverse populations and conditions, developing AI models to analyze it and engaging local communities to provide better awareness of strengths and weakness of our AI methodology.

We hope that our project will draw attention to applications of AI for rare disease analysis and empower both clinical and technical researchers in development and implementation of rare disease AI, with a focus on underserved populations.

Goal 0: Continue existing provision of care for patients in low resource settings, using telemedicine to provide continuous and home based care and teleradiology for evaluation of patients identified by telemedicine or hospital or clinic based care. These evaluations will continue to identify rare diseases and ensure patients are matched with accessible care as possible.

Goal 1: Selection of evaluation criteria and development of evaluation protocols for rare disease analysis.

Deliverable: Building on our existing prototype for identification and classification of types of tuberculosis and other pulmonary abnormalities on imaging, we will expand the tool to identify further pulmonary syndromes, with the addition of clinical data.  We will combine existing metrics and develop new tools for evaluation of the ability of computational models to work well with rare disease data, which may be of smaller volume or present with a significant domain shift, as compared to more main-stream heath data analysis problems.  To achieve this goal, we will collect and combine data from low resource settings where these rare infectious diseases are more common and where data has not been collected as robustly for AI tools. We will design experiments with existing datasets to define performance evaluations (across patient demographics or sub-populations) in imaging and health records selected by clinicians. The expansion of the prototype will allow for the better identification of rare diseases, but also predictions of clinical severity. Ethical board approvals have been sought and obtained where clinical data is to be shared.

Goal 2: Design of pretext tasks and optimization objectives for improved performance. 

Deliverable: we will design tasks and optimization objectives specific to biomedical data within each domain (health records, imaging) and across domains to better learn underlying data graphs and causality relationships, while simultaneously capturing the data variability. The goal of this aim is to learn cross-domain knowledge resulting in a robust model.

Goal 3: Quantitative and qualitative analysis of results 

Deliverable: The team, drawing on its combined technical and clinical skills, will evaluate methods’ results. We will look at both successful and unsuccessful performances to identify and address any potential systematic errors. We will also analyze performance patterns together with associated patient record and image data to study error patterns for specific patient demographics to design equitable models. 

Goal 5: Publication, code release and public dissemination of results.

Deliverable: we will package and document all code and benchmarks into Python modules and Jupyter notebooks for easy use and access by other users. In particular, we will make the frameworks accessible as libraries (via pip/conda install) and models readily available, similar to the Pytorch Model Zoo setup. We will also maintain a blog about our findings, and share novel ideas with the broader AI/ML and clinical communities in the form of scholarly publications and social media.

Goal 6: Expanding tool's platform

Deliverable: The code will be tested with the Radiokam teleradiology platform in Haiti to better determine the benefit for radiologists, primary clinicians, and patients. The goals will be to streamline diagnoses of rarer diseases, identify markers of clinical severity, and also reduce the heavy workload of clinicians in identifying rarer diseases. The tool would undergo further clinical testing to ensure effectiveness in practice as intended, without unintended consequences like deskilling. This can be a model and also provide mentorship for other clinical settings in Ghana and Nigeria where we clinicians are also looking for accessible tools for radiologists and clinicians.

Using fairness metrics such as equality of opportunity, parity, equalized odds, we will consider how fairly do the computational models treat protected attribute variables of interest. We plan to continuously discuss findings with collaborating clinicians to also identify any other criteria to identify any non-random correlations in results and discover vulnerable sub-populations. 

Rare diseases disproportionately affect underserved populations who often do not have access to healthcare and doctors. AI can fill in some of these gaps, but requires availability of rare disease data and cooperation/trust of patients and clinicians. Such trust needs to be built on existing care of patients with accessible evaluation and diagnostics and treatment. Tools to have practical outcomes with a history of successfully negotiating the roll out and implementation of technology in low resource settings.

Moreover there are relatively few available datasets and studies of AI for rare diseases. By building on our existing engagement with clinicians to collect and share databases, while also being a part of the tool development and continually sharing our collective knowledge for rare disease analysis, while building novel AI tools to improve computer aided diagnosis (CAD) for these diseases. We hope to build such tools using both imaging (X-rays) and available clinical data, building on our work to standardize data dictionaries across various low resource health settings, which will often have access to fewer diagnostic tests.

The technology used to build the prototype and for its further expansion will be heavily reliant on neural networks and AI. We plan to apply techniques such as classification and domain generalization to identify how the computational methods perform and generalize to specific variables of interest in rare disease diagnosis. We plan to use standard analysis libraries such as EHRKit and Pytorch, and develop our own. Using fairness metrics such as equality of opportunity, parity, equalized odds, we will consider how fairly do the computational models treat protected attribute variables of interest. This prototype will be continually modified and adjust in consultation with other clinical colleagues to identify any other criteria to identify any non-random correlations in results and discover vulnerable sub-populations. 

In many AI/ML applications to biomedical data analysis, fast and accurate diagnoses and treatment determinations are key to saving lives, allocating medical resources, and easing patient anxiety. Automated analyses of clinical data from patients can help identify important salient patterns, flag the patient for further evaluation, detect changes, and highlight important features to an evaluating medical professional. AI/ML methods typically perform well on data points similar to the training distributions, while many clinical and biomedical datasets are limited and do not represent the true diversity of patients. Lack of generalization may cause significant inequity in performance across subpopulations and disproportionately target vulnerable communities whose data may be underreported or underrepresented. In this project, we aim to study and develop techniques that are resistant to distribution shifts and generalize across populations, whether geographic communities, cohorts with similar diagnoses and conditions, or other important common characteristics. By directly incorporating patient equity variables and analyzing their effects, we plan to study how the proposed methods perform on vulnerable or underrepresented communities. 

One of our partners, Djinaud Prophete, has set up and runs a teleradiology company, Radiokam, in Haiti. The platform can be expanded to offer further diagnostic capacity for those hospitals in Haiti that contract with the teleradiology company.

This would allow for this tool to have benefit in Haiti and then to offer mentorship for similar opportunities in other countries and expansion, as feasible.

We do collect patient data (radiology as well as additional clinical information). This dataset would not be used for financial purposes and would not be used for profit. Data collection for ethical reasons would not be for sale or for payment.

We have depended on targeted donations and grants, while maintaining low budgets and overhead. The teleradiology company however is for profit and is based on service fees by companies contracting to use its services, allowing for centralized and rapid radiology reads and meeting an important clinical need.

We have worked to build programming that remains sustainable for those suffering from infectious diseases, focusing on low resource areas like Haiti and trying to find ways tech can truly help. We have received a number of grants and donations and can share our tax returns. We have received a Canadian Grand Challenges  award and have been featured in books and newspaper and magazine articles.

Books:

Farewell, Fred Voodoo: A Letter from Haiti by Amy Wilentz

The Big Truck That Went By: How the World Came to Save Haiti and Left Behind a Disaster by Jonathan Katz

Articles include

Barron’s: How to Wisely Give Humanitarian Aid by Jonathan Katz

“The Wounds of a Nation Still Bleed” Amy Wilentz’s ‘Farewell, Fred Voodoo: A Letter From Haiti’ Review by Michiko Kakutani

Haiti: The Compromising Reality by Mischa Berlinski

What's Next Locusts? by Pooja Bhatia

Star-Ledger Columnist Reflections by Bob Braun

and more