The Role of Conversational AI in Healthcare
Conversational AI is a solution that incorporates AI techniques such as Natural Language Processing (NLP) and Machine Learning (ML) to automate conversations with end users. They go above and beyond rule-based answering engines to understand the context, analyze the text and speech, the intent behind the questions and also consider the user’s preferences. Machine Learning enables the system to learn from the data it receives as and when it’s available and continuously improve its performance.
Conversational AI bridges the gap between humans and software even further, engaging in human like conversations and providing replies to users just like a human being would.
Conversational AI systems have been widely implemented in several industries including Banking, Retail, Marketing, E-commerce and others. But no other industry is riper and provides more impact creating opportunities for conversational AI than Healthcare.
If done right, such AI systems can unburden the maddeningly hectic lives of healthcare workers and have an extensive impact on human lives. Here are some interesting use cases for Conversational AI in healthcare and their important requirements.
Information Dispersal and Symptom Checker:
One of the most basic implementations of a Conversational AI is that it understands the symptoms that are worded in the user’s query and provides tips, related articles and information about the probable cause, thereby helping patients to get their first directive without any delay.
Virtual Assistants designed and authorized by recognized medical institutions can also play a key role in dispersing critical medical information in a timely manner that can prove to be invaluable. Especially in times of a pandemic, when getting in touch with a physician is really difficult, the quick dispersal of information about the virus, its symptoms and preventive measures would be of the essence. A high performance IVA providing a seamless experience and would essentially cut down the waiting line from hospitals.
Such Conversational AI systems and their impact will be multifold in developing countries having a huge population, with a disproportionate number of physicians and patients.
Patient Screening, Triaging and Escalation:
Conversational AI, apart from disseminating the basic medical information, can also assist people in need of emergency medical assistance based on their symptoms.
This is achieved through a specifically designed diagnostic tool. For example, a simple case of nasal congestion can be addressed by the patient himself through the information provided by the IVA. But if the user describes sudden numbness or weakness in his face or arm, especially in one side of the body, then it could be the sign of a stroke which warrants a hospital visit. The IVA relays the information to the user and can also provide the nearest list of hospitals with their emergency hotline contact numbers based on the location of the user. In a pandemic situation such as this, triaging IVAs could make a huge difference to the already overwhelmed hospitals by performing effective first level screening of patients.
Automated Appointment Booking:
Booking appointments with a specific doctor at a hospital used to be through phone or an email to the hospital’s customer service team. This transitioned to online bookings through the hospital’s website.
Now, the same use case can be implemented through a conversational AI. Users can converse with the bot regarding their symptoms and based on the information, the bot suggests a list of doctors available in the user’s locality, thereby enabling the user to book their visit all through the IVA’s chat interface. The IVA can even list doctors based on specialty that relate to the symptoms being discussed and provide a link to the doctor’s profile so that the user can verify the doctor’s credentials and patient ratings/reviews, before booking an appointment. Hence, educating prospects about the various diseases, precautions and vaccinations that are required and subsequent lead generation can happen within the same conversation.
Automated Billing and Registration
Healthcare Institutions can use IVAs to integrate with their backend billing and insurance claims management systems. This integration allows healthcare providers to quickly generate invoices for payments and eases patient interaction with the billing department. Health insurance providers, can also use chatbots to automatically answer questions related to coverage, claims, and procedures — giving much needed breathing space for the agents to concentrate on difficult cases and more productive work.
Automated Prescription Management
Patients spend a considerable amount of time online researching the medication prescribed to them. IVAs can be used by pharmacies to answer common medication related questions such as drug composition, drug-drug interactions, recommended dosages based on age and weight, side-effects, and more. They can even be used as an inexpensive and unique networking tool between pharmacies and manufacturers.
Patient engagement and follow ups
Post treatment engagement using conversational AI is a lesser known use case but one that is really valuable. IVAs can assist in following up with recently discharged patients to guide them through recovery, provide timely reminders for medication, help them to maintain a routine and a healthy lifestyle, monitor their vitals on a regular basis and in general nudge them in the right direction and keep them steady on their path towards recovery. Such a conversational AI implementation results in healthier patients, reduces patient readmissions and spares the hospital and its staff the time to address more serious patients.
Hospital Staff Scheduling
The use cases of the Conversational AI are not limited to the patients but also cross over to the hospital staff assisting them in the scheduling processes.
Intelligent IVAs can ably assist hospital administrators to access patient medical history helping them to provide appropriate directions to make the patient’s next hospital visit worthwhile. IVAs, if fed and trained with the required data, can play an important role in reducing the headache that hospital staff experience with scheduling of shifts and efficiently managing the workload of hospital staff.
Important Requirements of Conversational AI in Healthcare:
Discovery, Analysis and Evaluation of Solutions
Before getting down to building a conversational AI, there should be a clear establishment of the business case. Discussions with relevant stakeholders to identify the problem statement and evaluating different solutions is an essential activity. How is the process that the institution wishes to automate AI being done now? What is the target state expected out of the AI? What is the probable ROI expected through the AI implementation? These are questions for which answers should be worked out. Estimation on the average monthly number of queries and how many of these queries would be repetitive and whether automating these tasks would result in significant cost savings for the institution are things that need to analyzed.
Implementation, Refinement and Continuous Improvement
First comes the data preparation which can be from EMRs, information from the doctors, physician notes, logs, FAQ sections and other sources available within the institution. But the data from real users is the game changer. This can be obtained from chat logs with customer support executives, emails, phone calls with physicians and other staffs, messaging platforms and others.
Then comes the design of the conversation flow. This step involves the mapping and curation of all the possible answers the chatbot can give out. This can range from simple answers to ambiguous questions that involve complex workflows. The process contains several steps of validation to qualify and classify a query before eventually providing the exact answer that the query requires.
Testing of Conversational AI
Testing of an IVA differs from the traditional testing of other software, because here, the process is continuous and improves with each iteration. One way of testing it will be to validate whether all the training data in the NLP model is predicted and correctly handled by the IVA. For example, the query: “When should I book an appointment with Dr. Thomas for my follow up?” should be filed under the Patient Appointments intent and the model should provide the date and time of the next appointment along with the status.
Further tests can be as follows:
Cross-Validation tests the ability of the IVA to handle data that is different and new from what was used to train it. The K-fold and Leave-one-out cross validation techniques are commonly used for cross-validation tests.
Blind Tests are used to test the IVA with utterances, record and compare its answers with the expected corresponding correct answers.
Randomized Log reviews involves random checks of the chat logs and validation of whether the IVA was correct with its predictions and if the answers provided solved the user’s queries.
When it comes to testing conversational AI, it is important to stay within scope at first. The effectiveness of the testing is directly dependent on the data sets chosen. Trying to weave in context into the questions making them scenario based is a good approach. Include questions that would require the AI to fetch data from an integrated system and then process an answer. Queries that mimic what the real time users would ask is another critical part of IVA quality assurance. It is important to test the AI in batches, handling a significant portion of the sample data daily for about a week or two and observing if there is an improvement of 10 to 20 percentage points with each iteration. Tracking the coverage ratio, which is the relation between how many unique questions were asked by the users and how many was the IVA trained for, is a necessary activity. If the ratio is in the high 70s to 80s then it means that the training has been efficient.
Access to training data
Unlike other industries, access to data that is essential to train the bots to make them progressively better may not be readily available in a healthcare institution. Healthcare data is mostly unstructured, meaning most of the data is not labelled or properly classified, which is essential for NLP as it relies on labelled data for training the bot. Some of the processes that the institution wishes to automate might have been carried out through unofficial channels such as phone calls, emails and messages till now and might not even be documented. So it is imperative to build a robust data bank before jumping in to building an IVA. If the knowledge bank of the AI isn’t wholesome, then its performance would not be as the users wish it to be.
High Stake & High Impact
No other industry that has conversational AI in place has higher stakes than healthcare does. What the AI deals with, could potentially be a matter of life and death. It is important to design and train IVAs to prioritize accuracy and knowledge above any other essential traits.
Get the best out of Conversational AI through integrations
When a conversational AI is designed as a stand-alone entity it is not far away from being just another rule based bot. What makes it far more effective, intelligent and human-like is the fact that it is integrated to other data systems within the hospital such as EMRs, CRMs, Omni channel systems and calendars to improve the workflows and to form a thriving information ecosystem. Such integrations will enable the AI to transcend from being just an intelligent bot that automates patient handling and screening to becoming a wholesome concierge that is even capable of assisting hospital staff needs. Integration can be comprehensive in the form of back-end coding. Alternatively, hospitals can opt for low code integration, which costs lesser and takes significantly lesser time to build, resulting in 10 times higher ROI than a full-fledged backend integration. By being able to access patient medical history, the conversations of the AI with the patient would be much more personalized and relevant. Hospitals might store such data on cloud, or on premises or in a hybrid model which would directly dictate the hosting requirements of the AI.
Ambiguity in terminology
Healthcare has an ocean of data and thousands of terminologies. The usage of these terminologies might vary from person to person. There could be a big gap between the words the user uses to define the symptoms and what the symptoms really are. Hence training the IVA to provide the most appropriate replies or ask further relevant questions to confirm the cause is trickier. For example, when a user discusses with the AI about “flu”, they could be referring to “common cold”, “fever” or even “diarrhoea”. Such scenarios require clear and significant disambiguation.
Varying KPIs of Private and Public Healthcare Institutions
Public and private healthcare institutions would have different KPIs and the AI traits should be prioritized accordingly. Private hospitals would want to prioritize patient satisfaction and the quality of care, whereas public hospitals would be more intent on helping their physicians and nurses to handle more patients and provide satisfactory services. Public hospitals usually expect the conversational AI to screen patients and provide preventive care in order to reduce the workload on their staff and subsequently serve patients better. So it is important to discuss with the stakeholders and design the AI with the unique KPIs in mind.
Medical Data Security and Privacy
Ensuring the privacy and security of consumer data is critical in any industry but more so in the Healthcare industry due to its high stakes. If they the AI systems are hacked into or vulnerable to data exposure, the organization could face hefty fines from suing patients and governing bodies. So the design of the AI should certainly consider foolproof and robust security measures to prevent any mishaps.
The Future of Conversational AI in Healthcare
Frost & Sullivan predicts that within the year 2025, 90% of the U.S. Hospitals would have implemented AI and IVAs to improve the quality of care and save more lives. The company also predicts that with AI coming into the picture, there would be a reduction of 50% in the cost of treatment and positive outcomes would be up by at least 30–40%. Accenture forecasts that through AI chatbot and IVA implementations U.S. healthcare can save as much as $150 billion in annual savings within the year 2026.
Author: Mohan Bharathi Srinivasan
