Clinical Decision Support System (CDSS) for eSanjeevani

As per WHO 2022 data, India was at 0.7 medical doctors per 1,000 population, besides facing a shortage of trained manpower of nurses and pharmacists in the health system. To address this gap and provide an alternative to conventional physical consultations, the Ministry of Health and Health and Family Welfare, Government of India, introduced the telemedicine platform "eSanjeevani", which facilitates remote consultations. 

The eSanjeevani platform has enabled two types of telemedicine services — doctor-to-doctor (eSanjeevani) and patient-to-doctor (eSanjeevani OPD) teleconsultations. The former is being implemented under the Ayushman Bharat Health and Wellness Centre (HWC). There are over 50,000 operational AB-HWCs in the country. Patients use eSanjeevani by logging in, using a smartphone or web-based application, for general or specialized OPD consultations from enlisted medical institutions and doctors. They can also access eSanjeevani at public HWCs across the country.
The platform facilitates 0.2-0.25 million consultations in a day. We analyzed data from over 60 million consultations conducted on it and found several challenges, such as inaccurate and unstandardized recording of patient symptoms, use of free-text fields, lack of screening tools, incorrect OPD recommendations, limited application of standard treatment protocols, and high patient load. These factors relate to multiple challenges in accurate diagnosis, adherence to standard treatment guidelines, treatment regimens, and quality of treatment. 

The Government of India has included eSanjeevani in IndiaStack, a global digital public good platform, to offer the telemedicine platform to other countries as part of India’s G20 Presidency. Given the scale of the platform, it has the potential to transform primary health care for underserved communities across borders.

Our goal is to leverage AI/ML as a comprehensive clinical decision support system to improve the quality of care and features of the eSanjeevani telemedicine platform. We have a partnership with the Ministry of Health and Family Welfare Government of India, that provides us unique access to the eSanjeevani development environment and mandate to improve it.

To start with, we have developed, integrated, and deployed a SMART Interactive Patient Form (SIPF) into eSanjeevani. The SIPF uses a rule-based logical workflow to enable the accurate collection of chief complaints by engaging patients with relevant questions. It enables structured and efficient collection of necessary patient symptom details, shows patient symptom summaries, connects the patient with the appropriate remote healthcare provider, and facilitates automated recommendations of differential diagnoses to the attending physician. The form is currently available in 13 languages and covers 115 symptoms (and 31 diseases). 

We aim to expand this symptom repository, encompassing over 1000 clinical symptoms (covering 100+ diseases). We aim to develop, integrate and deploy AI/ML models for image-based screening for various conditions such as those related to dermatology. We aim to expand functions of the form to include diagnostic assistance, treatment regimen assistance, treatment adherence assistance, and morbidity risk profiling. At the same time, we are exploring LLM (Large Language Model)

based use cases to improve documentation of notes, and turn the conversation between patients and doctors into electronic health records by extracting information through multimodal inputs. Through our partnership with the Ministry of Health and Family Welfare, we get access to servers of the eSanjeevani so the period from build to deployment is reduced.

Here is an example of a patient journey with our AI powered CDSS in use at a HWC:

The eSanjeevani platform crossed a landmark milestone in February 2023, having served over 100 million beneficiaries. Officially, 57% of this population are women, and 12% are senior citizens. The three most populated states in India - Uttar Pradesh (U.P.), Bihar, and Maharashtra are among the top 10 in terms of adoption. U.P. and Bihar have two of the most backward health systems in the country with the majority of the population living in rural areas. 

The aforementioned numbers illustrate how the telemedicine platform is reaching vulnerable populations and ensuring the last-mile delivery of services in rural areas.  

Our SMART form solution aims to serve overburdened healthcare workers and healthcare providers such as Community Health Officers, doctors, and other staff in HWCs and outdoor patient settings. Given that India’s health system is already overburdened, it collates relevant patient information beforehand, so that doctors can focus on improving the quality of consultations and are supported by differential diagnosis. The solution improves patients’ healthcare-seeking experience by matching them with the appropriate caregivers, reducing time and expenses towards travel, unnecessary tests, etc. It also aims to ensure that the same quality of treatment is delivered across rural and urban settings, establishing standard treatment quality for all. 

Our solution would also have a broader impact on the healthcare system as a whole. By improving the efficiency and quality across the care chain, our solution helps to strengthen the healthcare system and health workers, making it more resilient and better equipped to handle the needs of India's vast population. Additionally,  our approach to open-sourcing our AI solutions, has the potential to create a ripple effect, inspiring others to develop and implement innovative solutions that can have a transformative impact on the healthcare sector.

At a primary level, we have established an AI-Centre of Excellence (AI-CoE) for health, which provides us with proximity to the nodal agency responsible for this platform. Our team has extensive experience in developing AI solutions for social impact problems, and we have strong partnerships with the government and other program partners.

In addition to that, we prioritize human-centered design and family-centric care principles to ensure our solutions are tailored to the needs of the target population. We gather feedback from patients, healthcare workers, and other stakeholders through user research, regular site visits, and program integration to ensure our solutions are achieving outcomes. 

We also collaborate closely with C-DAC, the technology service provider for the eSanjeevani platform, to ensure that our code is easy to integrate with the existing platform and meets the high standards required for deployment in public healthcare. With these key partnerships and our expertise in AI for social impact, we are confident in our ability to design and deliver effective solutions to the target population.

Lastly, we have internal expertise and experience in designing and deploying AI solutions for social impact across domains with a team of 160+ professionals. 

The SIPF is supporting with 0.2-0.25 million consultations every day conducted via eSanjeevani and approximately 4 million over a month.

We are applying to MIT Solve because we believe that our solution aligns with MIT Solve's mission of addressing global challenges through innovation and collaboration. As the only non-profit applied AI institute in India, we occupy a unique position in the ecosystem in the country, specializing in using AI/ML to solve social impact problems at scale. Therefore, in addition to funding, Solve’s resources and global network of organizations, individuals, and experts that use technology for social impact will be immensely valuable to us to learn from. Moreover, we see MIT Solve as an opportunity to contribute to the broader community of social innovators and share our experience in designing AI solutions for the complex social issues and populations of developing countries like India. More specifically, in the domain of healthcare, Solve aims to address inequities among vulnerable populations by leveraging community-driven tech innovation. 

Wadhwani AI’s goals are aligned with Solve in this regard. Our collaborative approach involves partnerships with the government, other nonprofits, academia, domain experts and the private sector at each stage of solution development so that our interventions are human-centric and have a wide impact. 

Our solution is innovative in two key ways.

Firstly, by improving the features of the government's telemedicine platform, we are setting new benchmarks for global technologies, as an organization representing low and middle-income countries (LMICs). This has the potential to change the market by catalyzing other players in the space (especially other LMICs) to innovate and improve their technologies. 

Secondly, our solution strengthens the capacity and efficiency of technology platforms already implemented by the Government of India. By integrating AI/ML-driven solutions on top of existing systems, we are making significant improvements in the delivery of healthcare services, particularly in rural and remote areas where access to quality healthcare is limited. This not only benefits patients and healthcare workers but also improves the overall health system by reducing the burden on healthcare workers, optimizing the capacity of doctors, and delivering quality consultations. 

Overall, our solution has the potential to create positive impacts that go beyond our immediate impact goals, which can inspire and influence others to make similar efforts.

1. Transform the quality of care across the care chain for the most vulnerable populations in India: Women, children, senior citizens, overburdened health workers, and populations in rural areas are some of the most vulnerable populations due to a range of socio-economic factors. Currently, excessive patient loads and a shortage of infrastructure and manpower prevent these groups, especially from accessing quality care. We plan to transform the end-to-end primary care experience by expanding the functions of the SIPF such as treatment regimen assistance, treatment adherence, morbidity risk profiling, and integrating image-based screening and large language models (LLMs) into future iterations. We will continue to work with various state health departments to deploy our solution across all states in India, improving the quality and efficiency of healthcare delivery to overburdened healthcare workers and healthcare providers.

2. Create a global AI public good: Over the next five years, we plan to further expand our reach and impact globally by sharing our solution as a public good. We aim to make our AI/ML technology accessible to health systems and healthcare providers in LMICs through our open-source approach. We believe that our solution has the potential to create a ripple effect in the healthcare space by catalyzing broader positive impacts from others in this space. By strengthening existing technology platforms with our AI/ML models, we aim to set new benchmarks in global telemedicine technology, improving health outcomes for vulnerable populations.

3. Improving national datasets: Our solution would lead to spillover benefits by improving the state of health data collection in the country and structuring government datasets. The Smart Form generates structured data that is crucial in monitoring and evaluating health programs, and will also be beneficial for training new AI/ML models. This data is also used to inform policymakers on health system priorities and to allocate resources. The solution could be expanded to generate real-time surveillance data for tracking disease trends, identifying outbreaks, and monitoring the effectiveness of interventions. With the increased adoption of technology-based solutions, this will encourage the government to invest more in digital infrastructure, creating a more efficient and accountable health system. 

To measure the extent to which the system has strengthened, we will develop and implement a robust monitoring, evaluation, and learning (MEL) strategy. This will involve conducting a quasi-experimental study and using appropriate metrics for conducting baseline and end-line evaluations with statistically significant sample sizes to understand the quantitative and qualitative impact. Indicators can include the number of SIPFs completed on eSanjeevani, the number of errors in completing the form, the total average time that is taken to fill it, usage (defined by ease of using the solution and number of people using it), and coverage (number of people who have access to the solution, additional health facilities that are leveraging eSanjeevani and the SIPF, etc.). These indicators are examples and will be finalized in consultation with the Ministry of Health. Routine monitoring of data, monthly reviews, quarterly interactions with stakeholders, and regular dip-stick studies will enable us to track progress and identify areas that may require a course correction. Relevant and high-quality data will be collected and used to develop and refine the AI models, along with continuous retraining. 

Activity

Immediate Output

Long-term Outcome 

Integrate new image-based screening solutions; enhance the SIPF’s symptom repository; clinically validate symptoms and associated questionnaires in coordination with technical partners

Expansion of provisional diagnosis to additional diseases (currently the SIPF supports the diagnosis of 31 diseases)

Expansion in the list of differential diagnoses and number of health-seeking patients 

Expand the functions of the SIPF such as treatment regimen assistance, treatment adherence, and morbidity risk profiling

Increase in the number of patients completing their treatment and of upward referrals as needed

Increase in the number of patients adhering to the treatment regimen, leading to reduced disease burdens

Integrate new cutting-edge technologies such as LLM in the SIPF

Creation of knowledge-base of symptoms, diseases, and drugs; facilitation of patient-agent conversations towards goals; prediction of differential diagnosis; transcription of doctor-patient conversations; extraction of medical records, etc.

Development of a comprehensive knowledge base for AI which will significantly improve solutions by enhancing differential diagnosis and treatment regimen recommendations, facilitate more organic conversations between patients and healthcare workers/providers

Support the Ministry with implementing the eSanjeevani program

Enhancement in the model’s performance and the solution’s reach

Use of AI-driven solutions on the national telemedicine platform to optimize efficiency and reduce the burden on the health system

The solution involves a combination of expert knowledge and the power of data through AI. It currently has two components - one component elicits the patient health information using an Interactive Form, and is a rule-based model. The rules or the sequence of questions that are being asked in the form come from medical literature and clinical practice. Eg. If someone has a fever, based on clinical practice and medical literature, the patient needs to be asked follow-up questions, such as the duration of the fever, the onset, and other characteristics of the symptoms and associated symptoms. These rules and the questions in the form have been formulated by a group of medical professionals. This medical information elicited from the patient then is passed on to the ML model, which then uses this information to predict a possible list of diagnoses. There are certain rules embedded into the system as well, to safeguard the patient. This is based on the idea that “no output” is better than a “wrong output”. There is also an expert-in-the-loop in the entire system who needs to verify the model’s output before making a decision.

The current ML model uses patient information/symptoms to convert them into a format that is read by the model, called encodings. This is just a representation of text into words, which is then fed to the ML model, identified as gradient-boosting trees (XGBoost). We are also running some experiments in parallel to use deep learning-based models like Biobert as well as LLMs to improve the diagnosis performance. The current XGBoost model outputs a list of diagnoses (called differential diagnoses) along with their scores. We pick up top-N diagnoses based on the model score and display them in rank order to the expert making the final decision.

Our organization strongly believes that DEI is an essential part of our mission and approach. We recognize that healthcare disparities exist across diverse populations, and we strive to create an inclusive healthcare environment for everyone. We take several steps to incorporate diversity, equity, and inclusivity into our work:

Firstly, we have made sure that our interactive patient form is available in multiple languages and covers a wide range of symptoms, which helps patients from diverse backgrounds to access healthcare services without any language barriers.

Secondly, we have also made sure that the providers who use our solution are trained on how to work with patients from diverse backgrounds. We provide them with training on the solutions we design and how best to incorporate them into their workflows while accounting for their nascency in using digital solutions to enhance and unburden their workload.

Thirdly, we conduct periodic reviews and assessments to ensure that our solution is achieving its intended outcomes and that we are meeting our diversity, equity, and inclusivity goals. We co-develop all our AI solutions, and program interventions with the communities through a process of iterative prototyping - collecting feedback and incorporating it. Through the co-development process, we hope to build ownership of the project to facilitate sustainability after the grant period. 

Fourthly, bias is a common challenge in AI and occurs when aggregate results are not reflected in marginalized or protected cohorts. We have experience in this area from our work on predicting adherence to medication in TB. We will use two broad approaches: collecting representative data that adequately captures protected cohorts, and testing and evaluating the algorithms’ performance at key stages of development and deployment to systematically check for bias. Bias removal will be addressed through both data resampling and algorithmic approaches.

We regularly assess our progress and seek feedback from team members and external partners to ensure that we are continuously improving our efforts toward DEI. 

We operate as a non-profit applied AI institute that leverages in-house AI/ML expertise to design solutions primarily for healthcare, education, and agriculture. Our social impact use cases result from landscape analysis, discussions with program partners, and government officials. We rely on grant funding from philanthropic organizations, corporate social responsibility initiatives, and similar sources to sustain our operations. Our approach is program-centric and need-based, and we work closely with partners to identify and address pressing social issues.

A key aspect of our business model is our strong partnerships with the government and other program partners. We collaborate with government bodies to identify use cases that require AI intervention and develop solutions that address the specific needs of the beneficiaries. We develop AI solutions, evaluate them, build capacity for the government, and then transfer them for scaled deployment. For example, we have established an AI-Centre of Excellence (AI-CoE) for health in partnership with the Ministry of Health and Family Welfare. The AI-CoE combines applied AI/ML techniques with a program-centric, need-based, and entrepreneurial problem-solving approach to improve existing technology systems' scope and effectiveness. Through this partnership, we are deploying AI/ML models to expand the functionality and capability of three technology platforms under the MoHFW - eSanjeevani (telemedicine), Nikshay (Tuberculosis), and RCH portal (maternal and child health). 

Overall, our business model is centered around a mission-driven approach to using AI for social impact. We strive to make our solutions widely accessible, and to that end, we open-source all our products, to turn them into digital public goods to maximize their impact.

To achieve financial sustainability, we rely on a combination of funding sources. This includes sustained donations and grants from philanthropies, corporate social responsibility (CSR) funds, and government bodies. We have a strong track record of securing grants from leading philanthropic organizations such as the Bill and Melinda Gates Foundation, USAID, and Google.org, as well as partnerships with government agencies such as the Ministry of Health and Family Welfare in India.

Furthermore, we plan to leverage our open-source approach to drive sustainability by turning our AI solutions into digital public goods. By doing so, we hope to draw contributions from a wider community of developers and users, which would drive ongoing innovation and impact and bring attention to our work within social impact funding networks. 

We are also keeping an eye on new developments in the social impact funding landscape, such as the upcoming social stock exchange (SSE) in India which will be launched by the end of 2023. Getting listed on the SSE will enable us to access more opportunities for grant and CSR funding, and open up access to different types of financial instruments such as social impact bonds. 

Project Name

Donor

The Official title of the grant

Start date

Duration (Months)

Newborn anthropometry

BMGF

Contextual AI in Health + Ultrasound for TB

01-Nov-19

46 Months

Pregnancy risk stratification (PRS)

Contextual AI in Health and Ultrasound for TB

Under 6-years anthropometry

CIFF

AI-Based Anthropometry

01-Oct-22

30 Months

Anthro

Botnar

Smartphone-based anthropometry technology

01-Jan-19

54 Months

Line Probe Assay (LPA)

USAID

Transformative Research and Artificial Intelligence (AI) Capacity for Elimination of TB and Responding to Infectious Diseases: TRACE-TB

10-Jun-20

48 Months

Loss to Follow-Up (LFU)

Cough Against TB

Differentiated TB Care

Integrated Disease Surveillance - IDSP

USAID

TRACE-TB: Strengthening Response to Epidemics (COVID-19)

01-Jul-20

31 Months

E-CollabDDS

CDSS

CottonAce

Tides Foundation

CottonAce

17-Jul-19

54 Months

CottonAce

H&M Foundation

CottonAce

06-Apr-22

12 months