Healthcare is one of the fastest-growing sectors in the country and is expected to reach $372 billion by the year 2022 as per an IBEF report. Along with artificial intelligence, there are multiple factors driving this growth — an increase in government spending on healthcare, rise in the purchasing power of the population, rapid urbanisation causing lifestyle diseases etc.
That said, healthcare in India is plagued by accessibility and affordability issues. India has one of the lowest ratios of healthcare providers and facilities — doctors and hospital beds — per thousand people and rampant opacity issues in quality and price of care. If India is to leapfrog and provide better quality care for its citizens, cutting-edge technologies such as machine learning and deep learning have to be deployed in the healthcare sector.
Sign up for your weekly dose of what's up in emerging technology.
1mg, an online healthcare platform, integrates the capabilities of AI and ML in its platform to deliver superior healthcare-related services at scale. Analytics India Magazine, for this edition of deep-dive, spoke to Gaurav Agrawal, Co-Founder and CTO of 1mg, to understand how the company is using AI and ML.
Founded by Prashant Tandon, Gaurav Agarwal, and Vikas Chauhan, 1mg started as HealthKartPlus in 2012 with an initiative within HealthKart to provide information about medicine like side effects, substitutes, and compare prices of the same products. But, while working closely in the healthcare space in India, the firm discovered there was a huge issue around opacity that was leading to a cost burden on consumers, for the same generic medicine, the price variation was at times greater than 80% between brands, and the consumers were just not aware.
“Our App HealthKartPlus went viral soon after the launch; we saw such encouraging traction, that we decided to pursue it as a fully dedicated opportunity. 1mg became India’s largest health app, and then we kept adding features, learning from our customers, and rebranded it as 1mg in 2015. 1mg is an integrated health app and offers online pharmacy, diagnostics and consultations at scale in one place. In addition, the app also has a ton of digital health tools such as medicine reminders, digital health records, and much more to make healthcare management easier. Our goal is to provide 360-degree healthcare service in a few clicks,” says Gaurav.
Healthcare Solutions Supported By Technologies
1mg offers a wide range of healthcare solutions that have deep AI technological capabilities. Case in point, the firm has a basing recommendation system on a user’s journey through the website/app, which is now in its third year of advancement. As a result, the recommendations are now next-action-prediction based rather than simple “product” recommendations that most early recommender systems provide. Besides, its fulfilment system is another example where there is significant use of technology — software that enables fast picking and automated QC ensures items are shipped on time and without error.
To deliver on their products and solutions, the company leverages advanced technology like microservices-based architecture and uses Python and Ruby on Rails for building services. Its services are hosted on AWS and use various AWS services such as relational database services (RDS) and elastic load balancing (ELB). And for the front-end, it uses a mish-mash of technologies but mostly relies on React.
How Is AI & ML Used
Millions of users visit 1mg every week to fulfil their healthcare needs ranging from information about medicines prescribed to ordering healthcare products and other services. One of the biggest areas of AI/ML for 1mg is the development of a Single Unified Health Repository to get a complete picture of an individual’s health. Most of the healthcare data collected — prescriptions, reports, and consultations with doctors — are unstructured. Non-standardised information is not computer-readable, and hence a number of AI models have been built over the years to convert incoming data into computer-readable, standardised data sets.
One of the key aspects of the above is — Digital Prescription Ecosystem that has been developed at 1mg using image processing as well as a semi-automated digitisation system wherein a trained pharmacist codifies the attached prescription image. The system is built to maximise patient safety and eliminate dispensation errors by accelerating Rx Digitisation using medicine suggestions.
Single Unified Health Repository has also been used to extract patients’ longitudinal health data and further develop disease progression models for chronic conditions that make personalised education and interventions possible.
“While healthcare AI takes most of the attention of our AI team, we are also doing some cutting-edge research in areas of commerce as well. In fact, it will be safe to say that there is no area of 1mg, which has been left untouched by AI,” says Gaurav.
Two key initiatives in which the company extensively leverages AI:
Order Delivery-Time Prediction Engine: The engine is powered by deep learning algorithms that are responsible for high fidelity and accurate ETAs for any part of the order fulfilment pipeline. This helps the firm in reducing order cancellations and improving customer ratings for the company. The system takes into account long-term data that includes both internal and external features — backlog in packing and logistics delays — to build robust and accurate delivery time prediction models.
Personalisation: It is at the core of 1mg’s data science efforts. To enable better recommendations for the visitors on the platform, 1mg has been using state-of-the-art collaborative filtering and transformer models to build its recommendation engine. It uses graph representation algorithms like Meta Prod2Vec to build user and item embeddings, which are utilised in neural collaborative filtering and models like BERT, GPT-2, to enhance personalisation of product recommendations; this has significantly improved the click-through rate (CTR) and product conversions.
Research Paper On AI-Based Differential Diagnosis
1mg also offers an AI-based differential diagnosis at the forefront of patient-doctor conversations. The objective is not only to reduce physician workloads but also to restore the human touch between patients and doctors. The company recently published an article — the probabilistic model for a differential diagnosis: development and validation of a probabilistic model to address the lack of large-scale clinical datasets.
Doctors in the country are overburdened. There is an acute shortage of healthcare force which is projected to worsen in the coming years. The company is striving to develop a model that will
- Make good quality healthcare available to every person in the society even those residing in the remote corners of the country
- Improve the efficiency of the doctors
- Ensure and accentuate the quality of healthcare delivered
The consultation carried out by a doctor to assess our health problem has two main parts: information collection to reach a diagnosis followed by management of the health issue. 1mg’s team figured out that there is a lot of scope of technology-based interventions in the first part of information collection. This information gathering can be further subdivided into history taking (this is the part where the doctor asks questions about our symptoms and complaints), physical examination and investigations. In a large percentage of patients, a medical history alone is sufficient to determine a list of differential diagnosis without the need for investigative evaluation. Thus, the company decided to focus its attention to this part of the problem.
1mg tried different approaches, over two years, and reached the conclusion that the art of diagnosis based on history is not a one-to-one connection but a mix of different possibilities. This gave birth to the idea of using a Bayesian approach to create the probabilistic model for diagnosis.
During history taking, symptoms are medically relevant information obtained from the patient. For the model, 1mg defined values as the objectification of the patient response, for symptoms and their characteristics, elicited by the patient. The possibilities of a response value are discrete and restricted by symptoms. Evidence (e) is defined as a set of symptom-value pairs based on the current complaints of the patient. The universe of all possible symptom-value pairs is E, and every patient presents with a subset of E. The addition of every new symptom results in the probability of disease being modified; that is, the prior probabilities are updated with the addition of new evidence. Thus, by aggregating all the probabilities, the model comes up with a list of differential diagnoses for the current complaints of the patient.
In the study, 1mg used clinical charts of patients (created by Internal Medical specialists) and gave them to six doctors and the model. Diagnosis and differential diagnoses for the patient’s condition given by the doctors and model were compared over predefined metrics. Results demonstrated that the model performed at par with the doctors. Similar studies have been conducted previously too by many researchers, but this was the first time that such results have been seen for a model. “Our study demonstrated the power of probabilistic modelling for medical diagnosis when the local patient and disease profile was kept in mind rather than utilising an off-the-shelf approach. In addition, we believe extensive involvement of practicing clinicians during the development phase is essential for the creation of a solution with demonstrable accuracy and clinical relevance,” explains Gaurav.
Tackling Talent Crunch
Since the quality talent is scarce in the data science field, 1mg invests a lot of time in filtering candidates as the popularity of this field has resulted in a lot of people jumping on the bandwagon without having the passion or the skill set for it. “In our early days, we found a lot of folks were implementing features from libraries but lacked a fundamental understanding of how or why things were designed the way they were,” says Gaurav.
“Most of the problems that we are solving at 1mg are unique and hitherto unattempted. Therefore, our data science team members are expected to be good problem solvers and should be able to innovate. Furthermore, they should have the ability to design experiments to test their solutions and draw credible inferences. We have regular paper reading sessions and data science colleagues are expected to keep themselves abreast of the latest developments in ML and AI. These attributes are in addition to the basic skills that are required: programming abilities, a quantitative mindset and a deep passion for the field. We value the diversity of background and therefore have hired and are open to hiring people who have had experiences outside of healthcare and e-commerce,” he added.
Since the inception of 1mg, it has always believed that technology and data will fundamentally change healthcare delivery. With billions of healthcare data points, 1mg is already working on cutting-edge AI models and will continue to accelerate work in that direction. Besides, the company is committed to enabling new healthcare models such as smart hospitals, outpatient insurance, among others, to make healthcare accessible and affordable. Furthermore, 1mg’s goal is to enable truly integrated outpatient care in the country. Such a plan will provide a patient with continuity of care both online and offline and across all their healthcare service needs. “We strongly believe that products and technologies are our core strengths at 1mg, which will continue to deliver superlative value to a healthcare consumer,” concludes Gaurav.