Non-invasive Anemia Detection App (NiADA) with Computer Vision Algorithm
The Technology Innovation
The World Health Organization (WHO) has stated that anemia is a global public health problem affecting 571 million women and 269 million children globally. Anemia exists with non-specific symptoms for patients to ignore long enough to become fatal. Late detection often leads to loss in productivity and even life-threatening conditions. It negatively affects the physical and cognitive development of children and contributes to increased mortality world-wide. The first step for addressing a problem is to provide an easy and accessible way to detect and monitor it, so that appropriate timely preventive or corrective intervention can be established.
The current state-of-the-art process for detecting anemia, Complete Blood Count (CBC) test, falls short in several ways. The test is invasive, requires a hematology lab setup and skilled technicians, is challenging to conduct frequently on children and infants, requires travel to the lab. It is time consuming and often expensive for lower-income communities and daily-wagers and requires consumables and proper medical waste management after the test is done.
Our solution, NiADA- Non-invasive Anemia Detection App, approaches the anemia surveillance problem in a novel way. NiADA has following properties:
● easy to use - it is a smartphone app that takes a photo of the inner eyelid (conjunctival pallor)
● non-invasive, does not require a blood draw.
● real-time - results in seconds.
● accessible - works with or without internet.
● point-of-care solution - convenient as no travel to the lab.
● inbuilt observability - shows history and trend from individual patient to population level.
● scalable - grows with demand.
● evolutionary - a technology architecture backed by and differentiated with a data platform that can incorporate incoming data from various global locations from diverse population.
NiADA’s novel solution introduces a new and effective way for preventive screening of this commonplace and yet silent killer disease. The proposed project aims to develop NiADA by using Deep Learning AI algorithm for automatically detecting presence and degree of severity of anemia from smartphone generated images of a person’s conjunctival pallor (inner area of lower eyelid). In recent years, the use of deep convolutional neural network (CNN) algorithms for medical image analysis and AI-aided medical diagnosis has progressed greatly. Successful examples include brain image analysis, retinal image analysis, chest X-Ray analysis etc.
We extend the previous work in multiple directions.
● Develop a machined learned way for detecting the palpebral conjunctiva region from an image taken by a consumer camera, instead of a special medical device.
● Employ hybrid feature engineering for mapping automatically identified interest-zones to the lab tested Hb result.
This automated learning of features by an algorithm can be very useful for progressing noninvasive procedures (such as use of smartphone cameras) for medical diagnosis and continuous monitoring as part of preventive healthcare. It can greatly reduce visits and the cost of medical care in outdoor or emergency rooms with early detection at home. Being a non-invasive method, it is extremely effective for children.
The Technical Objectives and Challenges
Future Data’s objective is to provide a smart phone-based solution at home and at point-of-care locations, that can be used as an accurate and effective (Mean Absolute Error, MAE within 0.2-0.5) screening and tracking mechanism for Anemia. This requirement adds certain technical risks and challenges as stated below for the project as it requires novel solutions to be developed.
Today, doctors are screening anemia examining the pallor of the conjunctiva, nail beds, palms, and tongue. NiADA detects anemia from the conjunctival pallor images, as
current evidence suggests that it is the most accurate indicator of anemia.
Data collection challenges:
Despite the interest and research in trying to build deep learning or statistical algorithms for developing non-invasive or minimally invasive solutions for detecting anemia from lower eyelid images, projects remain academic as they fall short of using appropriate amounts of data.
Future Data addresses this problem by setting up operations with a few hospitals for daily collection of inner eyelid images and the lab test result (CBC) as ground truth to be used for the machine learning input.
Currently, about 250 images are being collected every day from multiple hospitals. But as it is unlikely for patients to come see a doctor for anemia screening until it has become fatal or become overtly symptomatic, the collected data remains unbalanced.
- 80% of collected data are non-anemic, so to balance the data for training the algorithm, we are employing and researching better data augmenting methods.
- We are planning on getting samples from areas where anemia is prevalent but logistic challenges remain high.
Data quality challenges:
Deep learning algorithms need good quality data for accurate prediction.
Among the daily samples, about 10-15% of images require special preprocessing before it can be used. This is addressed at multiple fronts:
- Establishing a data quality pipeline by human doctors and trained professionals to sort images into different categories.
- As this selective input to model training improves our model accuracy, we are using the above classification to develop a better method for denoising.
- Some medical conditions prevent easy access to the inner eyelid, modified techniques need to be employed.
- Types of consumer cell phones pose great challenge for the model input.
Challenges in building the Machine Learning Model:
Medical image processing has achieved major successes in recent years, but it is difficult to estimate the required data volume for desired (MAE with 0.2-0.5). Moreover, more data is futile once the model reaches a threshold.
- Most medical image processing algorithms require special medical devices. We are adding a layer by using images from consumer devices. This will be a step forward for many future applications we will build to improve public health.
- Deep learning algorithms are big, and the trend has been increasing the number of parameters to improve accuracy, making them unsuitable for smaller devices like mobile phones. Our solution needs to fit into a mobile phone so that it works in rural areas without internet access.
The Market Opportunity
In USA, select demographic groups are victim of this silent killer which tests the limits of the healthcare system and destroys economic productivity in the absence of a non-invasive, point-of-care anemia screening solution like NiADA.
Target population:
· Anemia affects reproductive age (15 -49 years) women and children disproportionately.
CDC-WIC program to provide monthly nutritional support for 12.5 million women from underrepresented groups/races of pregnant women, infants, and children to combat anemia, in all 50 states. The program faces significant challenges in regular monitoring and tracking of anemia and reports 13% growth in anemia in the last 10 years.
· 10% of the growing senior population in the USA suffers from anemia.
· Annually, 37 million patients with chronic diseases (cancer, RA, kidney disease) and 16 million blood transfusion patients can benefit from a non-invasive anemia screening method as they require it frequently.
·
CDC reports about 3 million visits to primary care centers and ER just for anemia, annually.
Beachhead market:
Future Data is based in Utah. Utah has the highest rate of childbirth in USA. This beachhead market is comprised of about one million pregnant women, infants & seniors, and 1.6 million patients with chronic diseases.
Pricing and Total Addressable Market (TAM):
With usage based B2B subscription model, NiADA will be marketed initially through partnership programs at hospital chains like Intermountain Health and through women’s health centers in the Utah valley contacted during market research. Beachhead market TAM can reach 150Million dollar annually. The USA TAM is well over 7.9billion dollars.
The Company and Team
Future Data is pre-revenue. It is on-track for starting validation for the binary classification model to detect presence of Anemia in the next month.
Our founding team:
Mou Nandi, Cofounder & CEO
Mou has over two decades of experience building large, distributed software systems in Fortune 500 companies like S&P and HP using search engines, pattern recognition & image processing and natural language processing-based solutions. Last decade she has spent developing the platform for a legal document management company, NetDocuments, from scratch, building the teams to deliver the platform based on Search Engine and Machine Learning (Both Natural language processing and image processing).
Dr. Jhuma Nandi - MD, Pathology, India, Cofounder and Chief Medical Officer
Jhuma has seventeen years of experience in the field of medicine. She is currently in the USA going through USMLE second stage exams. She has contributed to the treatment of HIV patients in nodal ART (Anti-Retroviral Therapy) centers. During her early post residency (Pathology, India) she played a major role in quality and lab safety in hematology. Jhuma has worked as an in-charge of the hematopathology unit of a renowned teaching institution in India.
Krishanu Banerjee, Cofounder and Chief Data Officer
Krishanu has fifteen years of experience in a series of cutting-edge research for new AI product development in high-tech and energy industries. His experience includes fraud detection through image processing, utility chatbot and, predictive engine for missing payment prediction etc.