Denotia

Diagnosis of rare diseases depends majorly on professional knowledge, clinical experience, and scrutiny, leading to unsatisfactory diagnostic accuracy. Low doctor-patient ratios, high case loads and overburdened diagnostic clinics call for automated, AI-driven interventions. Attempts to discover high-quality imaging biomarkers have had some success.

Denotia can classify MRI scans into subtypes of frontotemporal dementia (FTD) and Alzheimer's disease (AD) in a few seconds. Our model uses a novel deep-learning method that analyzes these scans based on the differences between different structural regions of the cortex, with 86% and 84% accuracy respectively. Once implemented, it can revolutionize patient outcomes worldwide, reducing premature mortality and improving quality of care.

FTD is the second most common form of dementia in people under the age of 65 after Alzheimer’s disease. The mean age of onset is usually given as the late 50s, with an age range of 20-80. However, onset before 40 or after 75 is less common. It is important to distinguish FTD from other neurodegenerative diseases like Alzheimer's disease (AD), requiring different etiologically based treatments and patient counseling as clinical features such as disease progression and survival may differ. Both clinical and imaging criteria must be met to diagnose and classify such conditions. However, due to the lack of automated MRI quantification, the latter is delayed thereby causing non-specific, late interventions and mismanagement.

Diagnostic latency in the case of FTD is 4.1 years. Despite having distinct underlying causes, the overlap in symptoms and features leads to 15-33% of AD cases are misdiagnosed as FTD (discovered during autopsy) in clinical-pathological studies [1, 2, 3, 4, 5, 6]. Moreover, lack of standardization, universality, or uniformity for sensitivity/specificity values for types of dementia in current test batteries leads to misdiagnosis.

We built an AI web application for rare neurological conditions. Our program, Denotia, is an ML-based tool to detect and classify FTD and Alzheimer's Disease real-time.

1. Cortical thicknesses are automatically extracted from uploaded scans using a preprocessing suite. This saves time and resources and prevents most manual errors.
2. The thickness data is transformed into a network graph to serve as the data point for the algorithm.
3. A trained, strongly optimized graph neural network model is used to classify uploaded network graphs based on regional edge-weight variations in the graph.
4. After several rounds of feature extractions are completed in a few milliseconds, the final result is displayed as FTD positive/negative and AD positive/negative, along with the predicted sub-type.
5. Early classification leads to efficient streamlining of interventions and treatment plans. Long-term symptom management and treatment approaches can be finalized thanks to an accurate and expedited diagnosis. Moreover, this can accelerate pharmacological interventions and clinical studies through effective patient selection, stratification, and real-time measurement of outcomes.

Denotia serves a vast age group with dementia, who have the irreversible form of FTD. Early classification is needed for specific interventions and better management. Further, it helps differentiate it from other reversible forms.

Given the potential risks, cerebral biopsies are indicated only if other diagnostic approaches (e.g. MRI) have been insufficient in showing the cause of symptoms, and if it is felt that the benefits of histological diagnosis will influence the treatment plan. CSF biomarkers are also not conclusive.

An appropriate treatment plan in cooperation with the patient party can be set up, and caregivers can be sensitized to the different outcomes early on, the type of dementia classified, and subtype-specific intervention with accurate prognosis.

FTD-specific drugs aren't available, hence the specific subclassification can augment caregiver sensitization as well as disease progression in follow-ups. Often, the condition clinically progresses but cannot be detected radiologically to the untrained eye as the changes are too subtle. Deep learning spots these changes.

• FTD and most neurodegenerative conditions do not have a cure and require very diverse support/assistance in the form of symptomatic treatment. There are several healthcare professionals, NGOs, and therapy centers that try to cater to these needs. However, the space is fragmented, unregulated and thus poses issues of poor quality, user experience, and inaccessibility. 
• Denotia builds on traditional diagnostic methods as it utilizes evidence-based techniques to classify subtypes of FTD without fitting them into a box. The tool assists the clinician in navigating the diagnostic process, reduces operating costs for a hospital (allowing them to process a higher caseload on a priority basis), enabling the integration of the patient into mainstream society. Radiologic technologists can also benefit from its usage.

Frontotemporal dementia affects 2.7 to 15 per 100,000, second only to AD in primary degenerative dementia. Misdiagnosis/underdiagnosis cause late interventions, mismanagement, and improper treatment plan.

We used scans from the USC Laboratory of Neuroimaging's NIFD and ADNI (n = 466) to build an objective web tool that can serve underrepresented populations across diverse educational backgrounds and linguistic capabilities. Neuropsychological battery tests are often inconclusive; a tool like Denotia can help detect neurocognitive conditions up to 3 years in advance.

Since April 2020, Denotia has built a strong community of families, self-advocates, working professionals, and allies. Initially, we developed ADiag, a graph neural network-based tool to diagnose Alzheimer's disease.

Following similar principles, expansion of the model to FTD started in April 2021 and we are currently in the prototype stage with the new version. Currently, we are working with key neuroscience institutes from across India, including the National Institute of Mental Health and Neurosciences (NIMHANS) in Bangalore, Burdwan Medical College, and the Bangur Institute of Neurosciences in West Bengal.

Since May 2021, our focus was refining the algorithm and tracking the effectiveness of our program. We've been in a constant feedback loop with key doctors and industry experts in India. Additionally, we are focused on creating SOPs to effectively scale the program and meet our targets/goals for this year of rolling this out to hospitals and clinics.

Despite fulfilling the highly subjective clinical criteria, many FTD patients receive a delayed radiological diagnosis due to lack of standardization, inter and intra-observer variation, variability of clinical & radiological exposure, and judgment. The majority [40%] get a delayed clinical diagnosis, owing to lack of standardization and socioeconomic impact on cognitive expression and skills.

Denotia bridges this gap by classifying MRI scans into clinical subtypes of FTD and Alzheimer's disease, the two main causes of dementia in adults. This offers several key benefits:

• Reduces the number of follow-ups, time and resources required. Usually, a minimum of 3 visits is required; with Denotia, 1-2 visits (depending on stage and time after onset) are enough.
• Improves reliability and confidence of doctors in reaching a conclusive diagnosis
• Leads up to cost-effective, timely, scalable solutions: Existing AI interventions in radiology focus on critical abnormalities, like stroke, bleeding, major trauma, etc. Denotia prioritizes subtle changes which are likely to be overlooked even if carefully observed.
• Our survey (n=44) revealed that proper documentation and recording can be a huge challenge for progressive diseases like FTD. By providing information from the first visit itself, our tool can ease the transition from paper to digital, and archival of the changing pattern.
• When expanded with more community features, it can bring together experts in the medical community for peer opinions and raise awareness about rare conditions.
• Widespread use can accelerate the finding of additional/new surrogate markers as well as parameters for numerous conditions.

FreeSurfer, GrayNet: used to process T1-weighted MRI scans from FTLDNI and ADNI; first, cortical thickness is extracted, and then this data is converted into network graphs. These graphs have a 2D representation.

PyTorch Geometric: used for the graph neural network model using the 2D format. It consists of two sets, the GraphSAGE layers and dense differentiable pooling layers, alongside two fully connected layers.

Flask: used to create the web application from the Python code.

Bootstrap and HerokuApp: used to deploy the web application to the Internet, where end-users can access it.

Our core product is a deep-learning-powered web tool to classify brain magnetic resonance imaging scans into multiple categories. This includes 3 major subtypes of frontotemporal dementia—behavior variant, semantic variant, progressive nonfluent aphasia (PNFA), and Alzheimer's disease. It is delivered to clients via the Internet/Intranet and can even be set up as a standalone tool, subject to local computational resources.

Usage of DL in computer-aided diagnosis has been studied in depth [1, 2]. Moreover, MRI scans have been used to subclassify FTD in the past. Cortical thickness is a trusted marker to distinguish between AD and FTD subtypes. Lastly, MRI-based strategies have been used successfully to classify these target diseases [3, 4].

We are currently expanding to establish a better pipeline for data acquisition, to include more neurological conditions and imaging tests, as well as additional markers for FTD and AD.

AI algorithms performing feature detection, prediction, and classification can take RDs’ diagnosis to the next level, increasing these figures and uncovering new disease mechanisms and therapeutic targets (Brasil S et. al).

Diehl-Schmidt et. al (2014) reported that "different atrophy patterns for the subtypes are detectable at a group level...However, longitudinal studies show this particular imaging approach is not powerful enough to enable early detection and classification of disease at a single subject level."

Instead of solely relying on subjective clinical symptoms, computer-aided techniques can provide newer perspectives. At-risk populations for neurological conditions are usually non-English speaking as well as minimally literate in developing countries. In such a case, DL-based networks have the ability to perform accurate differential diagnosis of diseases without any hypothesis-based preprocessing.

This technology was accepted at IECBS 2021 and was placed 4th at ISEF 2021 by a jury of scientists and professionals.

Patterns of brain atrophy observed in T1-weighted MRI scans have been used previously to track structural changes in the human brain [1, 2], specifically for computational systems to pinpoint dementia pathologies. Moreover, binary classification using computer-aided diagnosis systems with MRI have been built for both AD and FTD [3, 4]. Few multi-class dementia classifications exist, and we have taken this approach using GNNs. DL-based networks have the ability for differential diagnosis of neurological conditions without hypothesis-based preprocessing, as seen in the case of Denotia.

• Bias in AI: Any ethnic, regional, and/or gender imbalance in the de-identified datasets can undermine a model’s performance on MRI scans of underrepresented groups. We tackle this by adding weights and stratified sampling. Moreover, since our machine-learning model was trained with graphs depicting the relative thickness differences of different cortical regions in patients, the MRI scans are not compared with one another. Essentially, each person acts as their own control and there is thus minimal bias.
• Privacy: The information contained in the hosting database is anonymized using an ID and password protected. Once we finish pilot studies, we will explore safe methods to temporarily store and discard patient information.
• Lack of agreement: Segmentation or diagnosis called for by an MRI can vary across radiologists. Any systemic bias in radiology will persist for models trained on radiologists’ predictions, despite deep learning models dealing with such random variability in ground truth labels. To prevent over-reliance on machine-generated decisions, we will tabulate general symptoms of FTD subtypes and AD, to ensure qualitative analysis. We plan on using this information on the development and addition of a clinical information form to our detection feature, as well as a response box to add any clinical notes.

Akin to medical imaging problems stimulating new developments in deep learning today, AI-aided diagnosis exceeding human performance could inform new discoveries in radiology. 

Currently, we are a prototype and have to rigorously test our MVP to serve patients. Till now, 56 dementia patients have agreed to participate in the first pilot (with around 10 radiologists at 3 institutes as beta-testers). 16 of them have a clinical FTD diagnosis.

India has approximately 800,000+ FTD patients. Conservatively, we assume reaching out to 10% of them in our first year, given the close network of FTD caregivers.

• 1 year: 100,000 scans (adopted at 35 hospitals and diagnostic clinics with ~300 scans monthly)
• 5 years: 2,000,000 scans (adopted at 900 hospitals, clinics, independent radiology centers, and nonprofits with varying monthly scans)

• Expanding our algorithm to detect abnormalities in other modes of imaging, such as X-rays, CT scans, PET, CR/MR angiogram, fMRI, etc.
• Targetting more diseases with distinct markers such as NMOSD and ALS.
• Geoexpansion due to the widespread nature of the disease: FTD and other neurological conditions are global diseases that don't discriminate between borders. Given the universal nature of this disease, we have identified a need for our product across countries like the USA, UK, Japan, Bangladesh, South Korea, and other countries within the EU.
• Automated report writing assistant that pre-populates radiologist templates.
• Peer-reviewed validation studies in collaboration with NIMHANS.

Goal 3.4: By 2030, reduce by one-third premature mortality from non-communicable diseases through prevention and treatment and promote mental health and well-being

Mean survival rates vary widely across different variants of FTD, ranging from 2.5 - 8.17 years. On a long-term basis, we will track the effectiveness of Denotia through any significant increase in survival years post-diagnosis. By providing an essential health service (Indicator 3.8.1), we can prevent diagnostic latency and premature mortality, which is especially common in the case of early disease onset.

Furthermore, quicker diagnostic confirmation of FTD will in turn accelerate medical research and development. Late diagnoses (coupled with rapid progression) make it difficult to conduct large-scale clinical trials - Denotia solves both.

In the short term, we will track:

• Number of doctors that completely adapt to our model (regular check-ins to realize this)
• Number of organizations that adopt our model
• Number of years from the onset of symptoms to diagnosis using Denotia 
• Total number of cases diagnosed
• Number of positive cases diagnosed per FTD variant
• Total number of patients that receive a positive diagnosis

Goal 17: Strengthen the means of implementation and revitalize the global partnership for sustainable development

Institutions and mechanisms to strengthen science, technology, and innovation: Track the percentage of all stakeholders impacted, especially the end-user: We want to benefit the patient with a diagnosis as early as possible.

Full-time: 5

• 2 Computer Science students
• 2 Neuroscience researchers
• 1 business developer

Part-time advisors: 7

Includes radiologists, neurologists, and professor-researchers

ADiag, which uses the technology Denotia is based on, was placed 4th (top 25% of 1920 projects) at the Regeneron International Science and Engineering Fair 2021. This July, Denotia will be presented at the International Electronic Conference on Brain Sciences 2021 to gain expert feedback from an international audience working in this area.

Growing up with a grandmother with dementia normalized fighting for acceptance. Interning at NIMHANS, alongside individuals working to comprehend neurological differences was an eye-opening experience. Vishnu brings his first-hand knowledge to the table and widens the perspectives of the other team members.

Naman, our software developer, has vast experience in the edtech sector. His exposure to underserved communities helps us build equitable solutions for greater access and transparency in clinical settings.

June and Anaya, our technology design consultant and business developer, have targeted their efforts towards creating a seamless and effective user experience. They oversee the functioning and administration of beta tests as well as future pilots.

Aranyo has worked on a culturally adapted screening tool for autism, which has been recognized by the APA. As the son of a radiologist, his knowledge about the challenges faced in the profession helps include every voice from the community and outside in our outreach efforts.

Our advisors include co-PI of the nationwide DSP, professor at NIMHANS, BIN, and other key neuroscience institutes in the country. They will serve as our hosts for pilot studies as well as initial usage.

Our team members are from a variety of lived experiences, united by our desire to democratize diagnostic systems in India and beyond.

Several expert individuals play an important role as our advisors and collaborators in our initiatives. At Denotia, we believe in a do-with-not-for approach and have regular conversations with working professionals and individuals with dementia.

Everyone is responsible for driving their own ideas from end to end. As a team, we prioritize physical and mental health and well-being as well as personal time, and that has reflected how we function. Open and transparent communication channels ensure that everyone can synergistically bring forth better results. Moreover, in order to ensure equitability, each member has equal rights and responsibilities: everyone is held accountable for their portion of duties and everyone enjoys a sense of involvement in the decision-making and management of the venture.

With gender parity across our roles, our inclusive policies will continue while recruiting newer members. We share common goals and priorities, as well as our weekly highlights on team calls and micro-meetings. This helps us give constructive feedback and building team motivation and spirit. We recognize that we cannot do everything on our own. Thus, alongside skills, we also seek expert guidance. Our advisors help us realize the overarching vision of facilitating more efficient and easier access to quality healthcare services. Our leadership team, having witnessed the implications of neurological diseases firsthand, understand, empathize, and therefore hope to revolutionize the way in which such diseases are diagnosed.

No. However, our advisors have diagnosed a sizeable segment of the population suffering from rare neurological conditions such as FTD.

Recently, conditions like moyamoya disease and NMOSD have been found to have distinct radiological surrogate markers. Our work can be generalized to such conditions and this would be greatly accelerated by insights from Horizon's R&D team. Working towards faster, non-invasive diagnostics directly impacts patient needs and symptomatic management.

Placing at this Challenge would give us the credibility to connect with more local partners and government bodies. Moreover, tech partnerships and recommendations will allow us to further democratize access to efficient intelligent tools to aid doctors in reaching the right diagnosis. More so, it will aid the development and addition of treatment plans for rare diseases of a cognitive/behavioral nature.

Solve encourages innovation and human-centered design, which we strive to uphold through our comprehensive all-round approach balancing the needs of all parties (patient and practitioner). Having a clinic-based healthtech solution will allow all beneficiaries complete transparency and efficient help.

- Access to capital is important for us to fine-tune the product and scale up to meet the growing demand. Due to low awareness levels in India, the neurodegenerative space is often neglected by the government and private equity players. Participating in the Horizon Prize provides access to more sensitive, empathetic and global capital. [anaya]

- Our platform can serve as a database and a tool for pharma partners to fastrack drug design and develop solutions for rare neurological diseases. We would appreciate support byHorizon in expanding our tool for NMOSD, which has key MRI features our algorithm can differentiate. Identifying and mapping our KPIs is significant to further improvement of our models. Moreover, we want to conduct micro-surveys and regular check-ins with both our customers (diagnostic clinics and tertiary healthcare centres) and end-users (FTL patients and families) to expand our features and improve overall efficiency in terms of delivery, classification/prognosis, and accurate risk prediction.

- Further, we appreciate the help of specialists in backend development as well as data science. Their assistance will allow us to deploy more community features as well as optimize our prediction widget. Having deployed our website using Heroku, we acknowledge the platform’s limitation in data security, scale, and network performance. Getting sponsorship from a larger-scale hosting service will allow us to process more data and hopefully detect more diseases in the future.

We'd like to partner with MIT's Laboratory for Computational Longitudinal Neuroimaging. Having their intel about FreeSurfer, the essential preprocessing method for our ML model, and the neurogeometry of presymptomatic diseases will allow us to develop a more optimized machine-learning model that can detect neurodegenerative disease faster than existing methods. As our web application is open for clinicians from all countries, this development will improve capabilities worldwide.

MIT has several exciting projects, such as Galaea, with whom we would love to collaborate. All our scans are anonymously stored and consent taken for their usage in further research and development. In exchange for these scans, these research groups could provide us additional insight as well as digital technology support to further complement our offering. Once we scale up, we eventually want to automate some processes, such as cognitive evaluation. Lastly, we are interested in collaborating with Dr. Guillermo Bernal of MIT’s The New Lab. Her project in neurorehabilitative VR headsets is a perfect complement/intervention method to our diagnosis method, which can be developed to read EEG, EMG, and EDA signals.

We love the workDigital Cognition Technologies and MemoryWell are doing to improve diagnosis and future outcomes of neurocognitive conditions. By partnering with them, we could coalesce ideas to target the same underserved population, both clinically/radiologically and in improving the quality of care. Together, we can revolutionize the lifestyle of thousands of middle-aged and elderly at-risk individuals, particularly in developing and underdeveloped countries.