AI-ML prediction of Neonatal Sepsis

Neonatal Sepsis kills overs 6 lakh neonates (0-29 days) in India alone every year. Infants born in rural areas of India are particularly affected due to lack of comprehensive diagnosis methods. Critical issues in diagnosis of Neonatal Sepsis are limitations of current culture tests, difficulty to obtain an infant’s blood, delayed diagnosis, antibiotic resistance, low access to healthcare in rural areas and low affordability. Our innovation, PreSco, is a mobile application that can be easily deployed on field to predict outcome of babies with risk of onset of neonatal sepsis. PreSco provides an early assessment of neonatal sepsis by generating a risk score based on the input values. The most important use of the application is that it can stratify babies into low, medium and high risk. Such a classification enables a health worker to administer antibiotics rationally. PreSco is particularly effective in resource-constrained countries and can save lives.

PreSco provides a neonatal sepsis score which is an early indicator of risk of onset of sepsis. This score can be generated multiple times for a continuous assessment of a baby till it is either free from risk or is provided with the right treatment. The risk score generated can be transmitted to the nearest referral doctor / centre along with the list of parameters for a further assessment or expert opinion. In low resource countries like India, rural health infrastructure is not strong. However, at least 50% cases can be handled by rural clinics itself. Every year, India gets about 25 million newborns and close to 10 million of them are located at the bottom of pyramid. The government newborn care system is able to provide care to only about 1 million babies. Our solution is effective in resource-constrained areas. It provides cost advantage, rapid assessment, integrates rural healthcare, frontline healthcare volunteers and doctors with urban network. With an early assessment, frontline health volunteers, rural doctors and rural clinicians can begin pre-emptive assessment, begin related therapy, make multiple assessments, or as a last resort, make an informed decision on antibiotic administration.

Our innovation, PreSco (Predictive Scoring) is a cloud-based machine-learning platform. It is a web responsive application that can be easily deployed on mobiles on the field to predict outcome of babies with risk of onset of neonatal sepsis. It deploys specialized machine learning algorithms that are able to generate probabilities of neonatal medical conditions like sepsis infections, pneumonia, meningitis (broadly classified under the umbrella of neonatal sepsis), in babies through multiple data points collected across various time intervals. A typical challenge in healthcare prediction of events lies in handling multiple records of patients without losing out on information from any single record which we are addressing using innovative machine learning techniques that can handle longitudinal / real-time data. We have developed a prototype of an integrated cloud platform for predictive risk scoring of neonatal sepsis. It consists of a data collection application and machine learning algorithms for three different levels of score generation – 1. Frontline Healthcare 2. Primary and Secondary and 3. Tertiary. A data collection application with multiple screens for various types of health parameters (maternal and infant) has been developed. An ensemble of machine learning algorithms has been used for building the predictive model.

Currently, the per capita cost of providing neonatal care in India is estimated to be around USD 187 while the annual cost is about USD 19,381. Additionally, it is estimated that about 11 to 23 babies receive antibiotics for 1 culture positive. Going by the lower bound, antibiotics can be reduced by at least 10-fold saving USD 143 million by using our machine learning platform. In addition to that, the platform ensures overall cost effectiveness in treatment through reduction in antibiotic usage, hospital stay charges in NICU when the algorithms provide low risk of sepsis by at least 50%. Additionally, culture tests for diagnosing neonatal sepsis take at least 48 – 72 hours. This time period is very critical for a neonate as it can deteriorate very fast because of low immunity. While administering an empirical antibiotic is safe choice, this practice has become rampant and has led to an increase in antibiotic resistance and is killing many babies. Hence, there is a need for a rapid and comprehensive test during the first 24 to 48 hours during which a baby is suspect to have neonatal sepsis. Our platform addresses this issue by running predictive algorithms.

As per United Nation’s Sustainable Development Goal 3.2, neonatal mortality rate NMR for India should be 12 by 2030. It is currently high at 25. Rural health network in India has 82% shortage of specialists. More than 40% of neonatal deaths are reported to occur within first 24 hours of admission due to absence of health facilities at rural areas and subsequent transport to urban centres. At least 80% parents shift or drop treatments for their babies mid-way because of high costs. Through PreSco, we are providing an affordable and accessible platform to address the issue.

We have adopted a novel approach to address the issue of diagnosis of Neonatal sepsis. Currently, diagnosis is made on the basis of blood tests, including culture tests, which are time consuming. Facilities for these tests are not available at all clinics, hospitals, especially in semi-urban or rural areas. It is extremely critical to diagnose neonatal sepsis in the early hours as a neonate’s condition deteriorates very fast. Our platform is based on machine learning. It detects the onset of neonatal sepsis in the early hours by mapping non-invasive parameters and providing a risk score. Bases on the risk score, a health practitioner can decide whether to begin treatment or not. This platform uses the power of machine learning to provide predictive analysis which is in itself the innovative feature of our platform.

Core Technologies - Machine Learning & Cloud Platform.

We are using open source technologies to build our application.

Our TECHNOLOGY STACK is provided below –

Database – MySQL

Frontend - HTML5, Bootstrap 4, CSS 3, Angular 7

Cloud Platform - AWS EC2, AWS S3 BUCKET, AWS CLOUD FRONT, AWS RDS, Python

Application programming interface - Node JS 10.15.3, Express JS 4.16.4, Docker 17.12.0-CE

Research from developed countries has demonstrated high effectiveness in decision making by including machine learning algorithms for diagnosing suspect sepsis at least 4 hours before clinical suspicion in developed countries. However, an existing scoring application, the Kaiser Permanante score developed in the USA uses prior probability for building the algorithm. Since the incidence of neonatal sepsis is high in India, Kaiser’s calculator is not likely to be effective as it does not factor in regional epidemiological variation. Our algorithms overcome this limitation. Kaiser Permanante’s scale focuses on prediction of early onset of sepsis (EOS) using maternal factors. Our approach uses infant and maternal factors for predicting both onset of early as well as late onset of sepsis. Kaiser’s scale is developed at tertiary level. Our algorithms are at three levels and are designed to work in low resource settings with open source software for cost optimization and improving affordability, access and connectivity.

Additionally, open source technologies offer greater transparency and interoperability, along with the advantage of low cost. Most of the current COVID self testing apps launched in many countries across the globe such as the UK, Israel, India etc are based on open source technologies and they have been working well. 

Machine learning algorithms that are able to generate probabilities of neonatal medical conditions like sepsis infections, pneumonia, meningitis (broadly classified under the umbrella of neonatal sepsis), in babies through multiple data points collected across various time intervals. A typical challenge in healthcare prediction of events lies in handling multiple records of patients without losing out on information from any single record which we are addressing using innovative machine learning techniques that can handle longitudinal / real-time data.

Reference - Mani, S., Ozdas, A., Aliferis, C., Varol, H. A., Chen, Q., Carnevale, R., ... & Weitkamp, J. H. (2014). Medical decision support using machine learning for early detection of late-onset neonatal sepsis. Journal of the American Medical Informatics Association, 21(2), 326-336.

Currently – we are piloting at 2 large hospitals

Impact in the first Year – 15,000 babies

Cumulative impact in five Years – 100,000 babies

Goals in the next one Year – In the first year, we plan to highlight on the Pay-per-Use revenue model. This will be achieved through direct sales, communication network, demonstration in health exhibitions.

Goals in the coming five Years – In five years we plan to achieve, close to 90,000 pay-per-use users and 1500 Annual Subscription users. This shall be achieved through targeting an average growth rate between 20% and 30%. 

We anticipate the below mentioned barriers - 

Technical Barriers – Fluctuating Accuracy, Sensitivity metrics for the algorithms developed. Quality of Data Inputs (Source of Truth).

Legal Barriers – Data security issues across geographies.

Cultural Barriers – Training and Facilitation, Low adoption of technology services by healthcare providers, hospitals and governments

Market Barriers - Low adoption of AI-ML technology platforms and solutions by healthcare providers government and healthcare organizations.

Overcome the barriers by the following strategies - 

Technical Barriers – Adoption of Robust Data Collection Mechanism, Testing and Validation Process.

Legal Barriers – Collection of Anonymized and De-identified data. Implementation of security guidelines.

Cultural Barriers – Improve training and usage. Inclusion of local languages.

Market Barriers – Exhibiting, showcasing application at health conferences, hospitals etc to improve generation of leads and conversion to sales.

No

A total of five people work on our team. We have a team of consultants and mentors who advise on a part-time basis.

Avyantra Health Technologies is a healthcare startup founded in the year 2017 and is based at Hyderabad, India. Avyantra’s founders strongly believe in fusing technology and innovation for enabling accessible and affordable healthcare in developing countries. Along with a combined experience of more than four decades in IT, Analytics and Healthcare industries, the team members complement their skills and training and believe in the spirit of teamwork. The core team at Avyantra is supported by a committed team of professionals aligned towards the organization’s vision of identifying, developing and delivering innovative healthcare solutions and improve access of healthcare to the low- and middle-income groups and other marginalized sections of the society in India. Avyantra Health Technologies is one of the 13 start-ups from across the globe to be selected for Unicef Innovation Fund’s 2019 cohort of investment. The company cleared a rigorous due diligence process to qualify for investment to their innovation of machine learning platform for early diagnosis of neonatal sepsis. Avyantra has successfully graduated from the Unicef Innovation Fund’s one-year programme that includes investment and business mentoring. In the last one year, the team forged collaborations with tertiary hospitals and an academic institution to develop the platform. The team met all the milestones and completed the product development successfully and is now moving into the next stage of testing & validation. 

• SCALE THROUGH OPERATIONAL AND FINANCIAL GROWTH BY PARTNERSHIPS WITH GOVERNMENT AGENCIES - Our project is aligned with Ayushman Bharat (Universal health coverage) in India. We also expect support from other partners in this area for further growth.
• SCALE THROUGH PARTNERSHIPS - We plan to collaborate with large hospitals (public as well as private) for data collection and scale up of our platform by enhancing our models and refining the prediction system. We are exploring opportunities for partnering with large software and medical firms and distributors for marketing & distribution of our platform. It is in this area that MIT Solve could offer us support in terms of leveraging their partners for partnerships and collaborations for testing and scaling our platform in other markets across the globe.

Business Model – Pay-per-use and Annual Subscription

Customers – B2B (clinics, hospitals, doctors, frontline health volunteers etc) and B2C (Parents)

We are an asset light company which helps us to keep our costs low. Post-COVID, in India and across the globe, there is a larger focus on digital health and mobile applications using advanced technologies such as machine learning and artificial intelligence in healthcare. We believe that these factors will help us in our growth and sustainability.