Ever since Alan Turing devised this famous test to check if machines could converse just like humans, there has been a steady progress towards developing one. Rapid growth in computing capabilities and data storage has led to new and ingenious artificial intelligence (AI) techniques that enable machines to learn with minimal human supervision.  

The past few years has seen even more innovations in chatbots that can automate and engage in human-like conversations with a user. These conversational AI systems have been applied to a number of industries including banking, retail, marketing and others.   

Healthcare is an industry that is ripe for so many use cases of conversational AI. If implemented correctly, these systems can have an enormous impact on human lives, healthcare workers and the medical field.   

 We have put together this guide to give you a comprehensive understanding of conversational AI in healthcare – from the various use cases, how they work, how to plan their implementation in your organisation and what further advancements we can expect in the future.

Contents:

  1. What is Conversational AI
  2. Conversational AI Uses Cases in Healthcare
  3. How Does conversational AI Work? 
  4. Conversational AI Challenges in Healthcare Institutions
  5. Conversational AI Strategy for Healthcare Institutions
  6. The Build or Buy Question
  7. Hosting – Cloud, Owned or Hybrid? 
  8. The Future of Conversational AI in Healthcare
 


What is Conversational AI 

The terms chatbots, virtual assistants and conversational AI agents are often used interchangeably. While they are all related and refer to the same technology in general, it is useful to distinguish them clearly for clarity. 

Chatbots are applications that automate chats. A user can ask a chatbot and receive an automated reply with no human intervention. Chatbots don’t necessarily involve artificial intelligence. In fact, the first incarnations of chatbots and even most of today’s bots use pre-defined, rule-based programming to deliver replies to queries.  

They work using guided paths presented to users in a button-click format. Users select their preferred options from the choices and deviations from the pre-defined paths are not allowed. They also have limitations in understanding users’ languages, customisations, integration into other systems in the organisation and may not have sufficient security or backup measures.  

Conversational AI refers to solutions that employ a variety of AI techniques like Natural Language Processing (NLP) and Machine Learning (ML) to automate conversations with users.

These go beyond mere rule-based answers to analyse text and speech, understand  intent and context, generate responses and continually learn from queries in order to carry out actual conversations with a user like a human. 

Back to contents


Conversational AI Use Cases in Healthcare 

Technologies like artificial intelligence and robotics are helping us progress to the healthcare of tomorrow. Specifically, conversational AI solutions have the potential to make life easier for patients, doctors, nurses and other hospital and clinic staff in a number of ways. Let us explore them one by one.

1. Symptom Checking and Information Dissemination 

One straightforward application of conversational AI in healthcare is in the dissemination of medical information. Think how when we want answers to burning health related questions, we often have to sift through jargon-heavy medical content. Chatbots can make this a more frictionless experience and provide answers to questions as if you are having a conversation with a doctor or physician.

 

symptom checker chatbot

For example, a parent can enter the symptoms or conditions of her sick child into the chatbot and receive basic first aid tips and guidelines. Or someone with an irritation in their throat can be directed to possible conditions that the bot identifies based on symptoms along with related articles and tips.

2. Medical Triaging and Escalation of Emergency Cases 

Sometimes the symptoms may point to the need for immediate escalation and emergency attention. Being able to identify these high priority cases and directing them to an ambulance is crucial. This can be accomplished through a step-by-step diagnostic tool. For example, a simple case of nasal congestion can be addressed with a health tips article. But a case where a user is experiencing a sharp pain below the breastbone could demand emergency attention and a call for an ambulance. 

medical triaging bot

In health crisis like the COVID-19 pandemic, such use cases could make a big difference to overwhelmed healthcare institutions. Conversational AI could be put to use to help prioritise patient screening and triaging treatment. Users can enter their symptoms in to the bot and be directed to take safe actions fitting the severity of their situations.  

KeyReply launched our own COVID-19 chatbot module to help users perform self-assessment to free up clinicians’ time to focus on infected cases. 

3. Doctor Appointment Booking  

Patient queries can also be more transactional in their intent. This is seen when they try to book medical appointments at a hospital or clinic. Chatbots can guide the patient through the necessary information through a conversation and eventually complete the transaction by confirming, rescheduling or cancelling appointments.  

Think of how you booked appointments for medical check-ups at your local hospital a few years ago. You would either call the hospital’s customer service team directly or email them. Their website may even have had a feature where you could check the calendar and find a convenient timeslot. 

 Today, you can use chatbots to get this done, just as if you were talking to an assistant. All you need to do is type into the bot what you want and through a series of conversations, you will be able to find and book your health screening appointment. You can even search for doctors by their specialties to book appointments with your preferred doctor, all in the conversational interface.  

doctor appointment chatbot

4. Patient Engagement Chatbot  

A lesser known but incredibly valuable use case of conversational AI in healthcare is in keeping up the engagement with patients. We have seen how bots help users diagnose and book appointments for treatment. But the post-treatment engagement is just as important if not more.  

 patient engagement bot

Patient engagement bots can help hospitals check in on the patients health and monitoring their vitals regularly after treatment. They can alert and remind them on taking medications in time and nudge them in the right direction to avoid any relapse. The end result is a healthier patient on the one side and healthcare staff with more time to spare for serious treatment cases in the hospitals on the other.  

5. Patient Information Retrieval and Scheduling 

The use cases in healthcare are not restricted to patients. Care teams can use chatbots to admit new patients and retrieve the personal info of previous patientsHaving access to patients’ medical history at their fingertips can help care teams provide appropriate directions and personalised treatment to make the visit a fruitful one for the patients.  

 patient info retrieval bot

They can also review their patient load throughout the day and assign nurses and doctors to various departments and shifts through reminders and push notifications. 

Back to contents


How Does Conversational AI Work? 

Chatbots work by analysing and processing user input and matching it with the most appropriate response from a database of answers. How they accomplish this is what distinguishes the simple bots from the artificially intelligent conversation agents.

Rule-based Bots  

In their early forms, chatbots were simple bots that pushed out automatic notifications. Or at best, they were systems pre-programmed to respond to queries based on specific rules. If users asked questions outside the prescribed set of questions, the bot wouldn’t understand or would provide a general response. End of story.  

Conversational AI  

The most advanced solutions on the other hand employ artificial intelligence techniques like NLP and ML. These help the bots understand context, intent and the user’s personal preferences. Machine learning is what enables the system to learn from data as and when available and improve its performance. Today’s solutions also come with personalities of their own in a bid to simulate a human conversation with the user.   

Natural Language Processing 

Natural Language Processing refers to a branch of artificial intelligence that deals with the analysis of natural or human language data by machines. Humans have evolved a unique capability over millennia to develop languages as a means to communicate information and ideas. The true complexity of human language is incomprehensible, with its differences across geographies, dialects, nuances, tones, context, accents and unique traits in specific domains.  

natural language processing

Machines don’t have an evolutionary history comparable to humans. Thus, it is a monumentally difficult endeavour to try and make machines understand language. To a machine, all of human language is unstructured data. Natural Language Processing uses algorithms to extract rules in human language to convert it to a form that machines can understand. 

nlp classifier

The above diagram approximates how NLP classifier works. A first step converts text from words into binary vectors, with each line representing the definition or meaning of the word. Then the sequence of word vectors is computer into a matrix representation of a sentence. Next, the matrix is compressed to distil a summarised version containing only the crucial information. And lastly, the classifier outputs a predicted target representation.  

Machine Learning 

Machine learning refers to a more general set of techniques to enable machines to look at past and current data and optimise for the best processes that lead to the right results. This can in general be categorised as either supervised or unsupervised. In supervised learning, the training data is labelled, while in unsupervised learning, it is not and the system has to study the data set to discover an underlying structure in order to make predictions.  

Intent 

No matter how questions are phrased, there is always an intention behind the query. A broad subject matter that the question is about. This is what is known as intent. Consider these 4 questions: 

  • "I want to book an appointment for a health screening”  
  • “How to book a health screening?”  
  • “Medical check up appointment booking”  
  • “Health check up appointmnt. how to book?”  
 

All 4 are different variations of the same essential question or action that the user wants answered – to book a health screening appointment. All 4 questions can be addressed with one type of response. So, grouping these questions under a single Intent allows for the bot to easily identify a user's intention and in turn give a relevant response.  

Entities 

Entities are groups of keywords that one their own may mean distinct things but all belong to a same category.   

For example, synonyms, abbreviations and jargon terms fall under the same entity.   

  • “doctor", "medic”, “surgeon”, “clinician”, “dr.” – all these words can be said to fall under the “physician” entity.  
  • “health screening”, “medical checkup”, “premium screening” –  are terms that can be said to fall under the “health screening” entity  
  • “fever”, “cough”, “blocked nose” – all could fall under all under the “symptoms” entity  

 Different brand names for a particular product could also refer to the same entity. For example, “Master Card”, “VISA", “AMEX” all fall under the “Credit Card” entity. 

Entities provide more context to an intent and thereby helps bots address more scenarios with just one sentence structure. In effect, they help bots scale up the scope coverage with the same model and amount of training data. 

Examples  

Examples refer to the different ways in which the same intent can be expressed by different people. In the above example of booking a health screenng appointment, the 4 variations correspond to 4 examples.   

Examples could also include variations of the same intent but with spelling mistakes, improper sentence structure, short forms, slangs and grammar errors.   

For example, the following queries also point to the same intent.   

  • Possible to reschedule health screening appointment? 
  • Can change health check up appointment? 
  • Medical check up appt how to change?  
 

Thus, examples are used to train the bot to recognise the different ways a specific intent may be expressed, so that it can provide the right response.   

To help train the bot effectively, it is important to collect real user data or as close to how real users would ask in every day chatbot queries.  

Conversational AI Platform 

A conversational AI platform is a software that helps you build, maintain and improve the chatbot. It generally comprises of a graphical user interface (GUI) with the capability to analyse and process data, deploy machine learning models and algorithms.   

The platform should also have the functionality to improve the system and deliver business insights based on bot data analytics. Subject matter experts and business stakeholders will also have the flexibility of updating dialogs and correcting responses as and when necessary.  

When the user asks a question, it goes through the NLP engine or brain, which quickly processes how to return a response. If no response can be found, there is generally a fallback layer comprised of knowledge from FAQs. If even this stage does not produce a response, the bot passes the question back to a live agent. The consistent training of the bot by clearing conflicting responses and adding more examples is what makes it smarter and more intelligent over time.

 

keyreply conversational ai platform  

The rise of messenger apps like Facebook, WhatsApp and LINE has contributed to the growth of these platforms. Not only do these apps have features to double up as chatbot platforms but they also have API kits that vendors can use to integrate into their own platforms.   

Testing 

A key part of conversational AI is testing. Unlike in traditional software, testing is not a one-time activity in the case of conversational AI systems. Instead, it forms an essential component of how these systems work and improve over time. 

There are many techniques that can be used for testing. One way is to ensure that all the training data in the NLP model is itself correctly predicted by the model. For example, if the utterance “How do I file a claim for my medical insurance” is under the intent “Claim_Medical_Insurance”, the model should correctly point to this intent.   

There are three other types of tests:  

  • Cross-validation: This tests the model’s ability to predict new data that is different from what was used to train it. So the bot will be tested with queries outside of the set of examples that were used for its training. The K-fold and Leave-one-out cross validation (LOOCV) are two common techniques.  
  • Blind tests: Blind tests involving testing the model with utterances and comparing the model’s response with the actual corresponding correct answer. Each one is marked as to whether the prediction was correct or not.  
  • Randomised manual log review: These involve human evaluators going through a chat log to mark whether the bot responses were correct and helpful to address the user queries.  

It is through an ongoing iterative testing process that the performance of the bot can be improved.

 

Read more about the 7 Golden Rules of Testing Conversational AI Solutions

 

3 Desirable Qualities in Conversational AI 

Depending on the use case, it is desirable for conversational AI agents to have one or more of these qualities. 

  • Knowledgeable – The bot should be good at fetching the right info from the databases it has access to, and returning to the user with a correct response. At the end of the day, users want to get things done more than anything so this is one quality that is good to have in abundance. 

knowledgeable ai bot

  • Engaging – Even if it is obvious that the user is conversing with a bot, it is good to give the bot a certain personality. Not only is this helpful in providing a good user experience, it can also be an opportunity to promote the company brand. If it makes sense for your brand, jokes, anecdotes, quips, small talk and chit chat – all are welcome here. However, it is not ideal to have too much of this in your dataset in case it overshadows the main content that it is being answered when the user has actual business queries. 

engaging ai bot

  • Empathetic – Just like in human to human conversations, it makes a big difference if the bot can put itself in the user’s shoes when responding. If the answers are too factual and devoid of any warmth, it may address the user’s queries but nothing more. In fact, it can even turn away the user who might prefer to speak to a human the next time. Think of how you as a consumer would want to be greeted and engaged with. 

Note that rather than being specialised in one single quality, a good conversational agent should be able to seamlessly blend them all into one cohesive conversational flow. 

Back to contents


Conversational AI Challenges in Healthcare Institutions 

While the mechanisms by which they operate may be similar, the same conversational AI solution may not be applicable across diverse industries and uses casesIn the healthcare industry, there are specific challenges to address which will dictate how organisations go about implementing a chatbot.  

1. Limited Access to Training Data 

The data needed to train a bot may not be readily available in a healthcare institution. It is an industry which has traditionally been slow to adopt technological innovations and digital transformation. This could be due to the emphasis on human to human interaction (patients expect to be treated in person by doctors), the higher levels of risk and compliance regulations.   

 Consider the types of data that would be ideal to train a bot. Common queries around location and operating hours aside, users could ask about medical procedures, health screening, symptoms and matching doctors and could even share their personal info.   

 Even if such data is recorded and documented, they may not be labelled. Labelling is necessary for any NLP system to extract meaning and establish relations between words and entities.  To complicate matters, some of the communication that needs to be automated may be carried out through unofficial channels like personal messaging or email. These may not be documented or labelled.   

2. Differences in Symptom Descriptions and Medical Terminology   

The healthcare industry is somewhat unique due to the vast medical terminology it uses. Specifically, there could be a big gap between the language of user’s queries and the correct medical terms corresponding to those queries.  

For example, when a user asks about “flu”, they could be referring to “common cold”, “fever” or “diarrhoea”. Such scenarios require significant disambiguation. 

 Moreover, the terms that a bot most frequently encounters could vary between geographical regions, societies and even among individual healthcare institutions. For example, in some conservative societies, people may want to consult a doctor as soon as they discover symptoms. In other societies, they might be inclined to wait to see if the symptoms subside before even thinking about reaching out to a hospital. The result is a slight difference in the most common queries that might be entered for symptoms. These are broad generalisations but important nonetheless for conversational AI systems to account for.  

3. The High Impact Nature of Scenarios and Use Cases

The common use cases in finance, retail entertainment or sales and marketing involve topics that are relatively harmless. Getting wrong or inaccurate responses from time to time will not have a huge impact. Think how you interact with a chatbot to enquire about the procedure to open a bank account online or checking out a product from an e-commerce site. If the bot is unable to help you complete the transaction or if it takes you to the wrong product page, it does not signal the end of the world. 

But in healthcare, where it is often a life or death matter, the stakes are much higher. A parent could be enquiring about the right treatment for her injured child or a user might be in need of urgent emergency care for a stroke. In such high-impact scenarios, chatbots may have to prioritise accuracy and knowledge over other traits like personality. 

4. Differences in KPIs Between Private and Public Healthcare Institutions 

Even in the healthcare industry, the priorities and KPIs could differ based on the individual institution. Private institutions might prioritise patient satisfaction and high-quality care more, especially for the Executive and Premium packages. They will be interested in KPIs around leads and awareness among users on related treatment services and elective surgeries. 

In contrast, public hospitals generally place emphasis on enabling their nursing teams to handle more patients and provide satisfactory experiences for patients. Managing the workload of healthcare workers and optimising costs will also be high among their priorities. Most importantly, they will aim to shift resources towards preventative care in order to reduce the load on their staff so they can serve patients better. 

Conversational AI platforms and vendors will therefore have to work with the hospital management and IT stakeholders to design solutions with their unique KPIs in mind. 

 5. Integration with EMR and Other Systems   

A conversational AI solution that is an island in and of its own is hardly better than a rule-based bot. To get the best out of the solution, it needs to be integrated into other internal systems within the hospital to form an information ecosystem.  

To make a crude analogy, think of the Apple ecosystem. The consumer devices – iPhone, iPad, Macbook and the Apple Watch – by themselves may be of immense convenience. But it is the connected ecosystem comprising all these devices that enable features like the smooth Handoff, calendar, podcast and iBook syncing, fitness data sharing and so on.   

In healthcare institutions, access to electronic medical records which include patient profiles, previous treatments and allergies make a big difference. By integrating into these systems, the conversational AI can provide users and patients with more relevant and personalised responses.  

Note that in hospitals such critical data might be stored on premise, on the cloud or in a hybrid model. This directly dictates where the conversational AI platform will need to be hosted.   

 6. Patient Data Privacy and Security 

Protecting customer data and ensuring privacy is an important consideration in any technology adoption, irrespective of the industry. In healthcare, this is even more critical for various reasons. First, the use cases are more high impact as discussed above. Data that could be key to life or death decisions must not go public.   

Secondly, access to such critical data can enable by third party agents could cause embarrassment, be it intentional or not. One of the earliest publicised applications of big data involved a case of a parent being targeted with pregnancy ads for his teenage daughter. Although the validity of this story is still not certain, the issue remains 

Lastly, healthcare being a service that is universally accessed, the patient data could also include health details of various influential and political figures. Leakage of such data could find their way into hackers and bad actors who could use such data for nefarious purposes.   

7. Clinical Protocols and How They Differ Across Hospitals

Unlike other industries, there are certain protocols and standard operating procedures that have to be followed in every interaction with a patient or customer. These cannot be circumvented and there is no room for improvisation either, as this could lead to legal and regulatory consequences.  

This also ties into the “philosophy of care” practiced in the region and even in the specific hospital. Due to societal, cultural and economic differences, the attitudes towards healthcare may differ between countries and regions. And this often directly translates into the clinical protocols adopted in the region and hospital.  

As mentioned in regards to the medical terminology above, patients in the U.S. may be inclined to wait for a time period before they consider getting checked. This could be either due to the general expensive nature of healthcare services in the nature or the prevailing attitudes among the population towards healthcare or both. In contrast, people in Singapore generally try to book appointments and get checked up at their hospitals as soon as they start observing symptoms. The healthcare institutions in these regions therefore differ in their philosophy of care and therefore in their adopted clinical protocols.  

Thus conversational AI systems have to take into account these protocols when designing their dialogue flows to cater to the needs of the population. If necessary, AI techniques should take a backseat here and only be applied outside of these procedures. 

Back to contents


Conversational AI Strategy for Healthcare Institutions

 Despite the challenges that are unique to the industry, healthcare institutions can get all the benefits of a conversational AI solution by approaching it with the right strategy. This involves 3 key phase - Discovery, Implementation and Refinement, and Integration.  

1. Discovery 

Before doing anything, it is important to establish a business case for deploying the conversational AI solution. This involves getting the relevant stakeholders together to identify the problem statement and evaluating potential solutions.   

It helps to conduct an examination of the current state and an expectation of the target state, along with the corresponding ROI calculation. Estimate the average monthly volume of queries that need to be automated, the portion of these queries that are repetitive and whether automating these will result in significant cost savings.   

 This is where private healthcare institutions might set objectives and KPIs in relation to leads and revenue while public hospitals do the same for their costs and investment optimisation targets.  

It also helps to extrapolate the current state to what the next three years would look like. Not only will this align the various teams on the KPIs to measure but it can also help figure out the progressive categories of tasks that can be automated within the organisations – from the mere repetitive “irritants” to the more complex “strategic” ones which bring maximum value to both the customer and to the organisation.   

 2. Implementation and Refinement 

Once the discovery phase is completed, the implementation starts with the data gathering and preparation stage.    

  • Data preparation 

Data used to train the bot can be collected from various sources within the healthcare institution. Organisational structure, info on doctors and physicians, key specialisations of treatment, FAQ sections, internal wiki documents can be helpful.   

But what really makes a difference is data from real users. This could come from previous chat logs, email enquiries and other unofficial channels of communication such as personal messaging apps. If these are still insufficient, it helps to rope in the customer-facing teams to get a better idea of how customers phrase their enquiries so that the bot is trained on data that is as closely representative of real-world queries as possible.  

During data preparation, examples of real user queries are collected and their intents and entities labelled. Aim to collect at least 10 to 20 examples for each intent to help the bot understand queries comprehensively.  

  • Dialog flow 

Once the data preparation is done, it is time to set up the flow of the conversation. This step involves mapping out and curating all the possible answers that the bot can return. The answers can range from simple direct answers to more ambiguous questions involving more complex workflows. These often contain several content nodes or steps to qualify the question and lead the user to a specific intent.   

3. Integration 

This is also the stage where the bot is integrated with other systems like electronic medical health records, CRMs, omni channel systems and calendars to improve workflows. Such integration is what takes the application from being just an intelligent bot towards becoming a full-purpose concierge that addresses the needs of more internal teams in addition to patients. 

 For this to happen, the internal healthcare systems have to be open and ready to integration. Such an integration can involve a comprehensive back-end coding with the involvement of the vendor's software engineers. Alternatively, it could be achieved through a low-code integration which does not need coding support. Low-code development can be an attractive option for hospitals with limited budget as it can result in nearly 10 times the ROI of a back-end integration. 

To give you an idea of the difference in timelines, consider a normal integration of a chatbot to an appointment system. It involves the basic features like creation of the appointment, checking appointment status and cancellation info the appointment. This simple use case can take around 3 to 4 weeks to integrate. It will also require multiple resources including project managers and developers from the vendor and the IT teams working together. This may not be possible if basic off-the-shelf solutions are preferred over a vendor with expertise. A more full-scale concierge system which integrates to multiple systems within the hospital, would require more resources and a longer timeline that extends in to months.  

A low-code approach can accomplish the same basic appointment feature integration in 2 days, and will also bring down the timeline for a full-fledged solution. Correspondingly, the overheads and cost in this case will also be lower.    

4. Testing  

As described above, testing is a critical stage in ensuring that the conversational AI works as intended and improves over time. Thus organisations should approach this part of the adoption carefully. The most important thing to keep in mind is how conversational AI systems differs from traditional software. Unlike traditional software, conversational AI solutions are not rule-based programs but complex systems that employ probabilistic models to learn from training data to make predictions. With this in mind, there are some key guiding principles to follow during testing.  

  • Stay within scope at first: Keep the testing to within the scope of the training. Just because you can enter anything into the AI interface doesn’t mean you should. 
  • Prepare a good test set: The efficacy of the testing depends on the data set chosen. Try to prepare queries that are scenario based, descriptive, representative of real questions asked by users and which have well-defined answers.  
  • Test-Analyse-Action: Adopt an iterative approach where trainers go through a statistically significant sample size of user query data daily for one or two weeks. In every iteration, look for improvements of 10 to 20 percentage points.   
  • Keep an eye on the coverage ratio: The analytics you measure to track the improvements in performance have to include Coverage Ratio. This gives you an idea of how many of the questions that users asked were the bot trained for. If the ratio is around 70-80%, it means the questions are well selected and representative of what real user ask. If it is much lower, then you should revisit the questions and replace some to represent real user questions more accurately.    

Back to contents


The Build or Buy Question 

 A question that many organisations face in their digital transformation journey is that of whether to build technology solutions within the firm, using their own resources or to buy the services of a qualified vendor. Aside from the usual considerations like cost, vendor reputation and time commitments, the answer also depends on these other factors.   

1. Skills and Knowledge 

It is a fact of reality that not all institutions will have highly skilled technology teams and expertise within the firm. Firms in the financial services, retail, higher education, marketing services and IT services verticals generally have a higher adoption of technology solutions. Such firms may therefore already have an in-house talent pool of data scientists, developers, UX researchers and engineers. Forming specialised teams that work on conversational AI solutions is a reasonable strategy for them. In general, it takes a team of at least 20 to hundreds of highly skilled researchers in an AI lab, such as that of Lenovo, to achieve a certain acceptable level of performance.    

Healthcare institutions and other smaller enterprises may not have such a level of technology expertise in-house. In fact, hospitals may already have a large and complex ecosystem of mission critical systems to maintain and may not want to take further technology risks with AI R&D and software development. They might be better of buying the services of a vendor so they can focus their resources on upgrading and maintaining their core systems instead.  

2. Domain Expertise

 Healthcare institutions can be expected to have the necessary domain expertise inside their organisation for obvious reasons. This might make building in-house an attractive option. However, they will still have to rely only the data sets that they have access to, in order to train the conversational AI.   

Conversational AI platform vendors, especially those experienced in working with multiple healthcare institutions, will generally have built up a specialised knowledge database in this domain. Leveraging this extended domain knowledge may help the bot cover a larger scope of queries and achieve a higher accuracy. 

3. Multi-Language Support 

The language used by patients and users of a healthcare chatbot is also a deciding factor. If the hospital operates in English-speaking regions or where the languages used have numerous data sets, developing ML and NLP models for conversations can be manageable. But if the patients will converse in languages like Thai, Vietnamese or others in Africa, which do not have a lot of data sets available, bringing in a vendor with experience in these low-resource languages can be a better option. 

Back to contents


Hosting – Cloud, Owned or Hybrid? 

Once the decision has been made on whether to build in-house or use the services of a vendor, the next decision is around the hosting of the solution. You could host the bot on-premise, in the cloud or go for a mix of both.    

Hosting on-premise involves provisioning dedicated storage and physical servers. This gives you more control over the security and privacy of the data. You will therefore also take on the risk of maintaining the solution and ensuring continuous application delivery. A private cloud option does away with the need to have dedicated physical storage by offloading to the cloud while still ensuring security.   

A public cloud implementation involves hosting the solution in an off-site cloud, usually offered by a provider like Amazon or Microsoft Azure in a monthly subscription plan. This option comes with the convenience of quick deployment and integration, but could compromise security.  

A hybrid option allows you to get the best of both worlds, with some sensitive workloads hosted in the private cloud while offloading less critical workloads on to the public cloud. You will still need to classify the services you want to deploy in each based on the accompanying risk.   

Here are a few additional considerations to keep in mind when choosing the type of hosting.  

1. System Limitations 

When working with a vendor, system limitations and lack of availability of partners and resources could prevent a full cloud or full on-premise hosting which leaves the hybrid model as the best option.    

2. Cost Considerations 

While it may be tempting to think that a physical server or data centre deployment would be cheaper, there are other issues that could ramp up the costs over time. Firstly, there is often a minimum server hosting requirement. Moreover, it is not easy to scale as this would require purchasing more hardware which turns out to be more expensive. In a cloud-based model, the pricing is dynamic and based on resource consumption. This means you pay more if you need bigger sizing, and less if there is no need to.  

3. Timelines 

On-premise (private cloud or local server) deployment requires more time due to various factors. To start, getting security compliances and clearance might take time. If the existing systems are old, even simple file transfers could take hours or days. And in case of any system incompatibility, some additional rework might be required to ensure that the chatbot solution fits and is deployable. 

4. Database proximity  

To ensure that the data extraction and analysis is smooth, the database servers should be close to where the chatbot solution is hosted. Ideally this should be just milliseconds away from the server hosting some of the core scripts. 

5. Data privacy 

The hosting option is also affected by the local data transfer and privacy restrictions. Each region or country will have its own regulations. In healthcare, data security is of paramount importance. Hence, it is important to work with a provider who show proactive steps to ensuring compliance to industry standards. One well established guideline will be the Health Insurance Portability and Accountability Act (HIPAA). It regulates the protection of privacy and security of health information.   

Companies who are compliant have written policies, conduct training, monitor and enforce standards. They also have designated compliance personnel and respond promptly and take corrective action to offences. 

Back to contents


The Future of Conversational AI in Healthcare 

In a rapidly evolving technology field like artificial intelligence, it is hard to predict what the state of affairs will look like in a few months, let alone a few years. Just think back to the year 2010 (before the explosion of convolutional neural networks) and see how far we have come today. Or compare 2010 to the year 2000 when the idea of AI was still in the domain of science fiction more than every day technology solutions.  

However, judging by some of the trends in the field today, demands for new use cases in the industry and recognising some concerns from the global community, we can guess what the future of conversational AI in healthcare will look like.   

Virtual Care in a Post COVID-19 World 

While there was a slow progression towards digitising healthcare services and introducing more virtual care technologies, the COVID-19 pandemic has accelerated this transformation. It has essentially changed how people perceive care and how healthcare institutions plan provide it.  

In the pre-COVID era, many healthcare providers could not completely break away from providing care physically. At the bare minimum, they had to provide customer service through phone calls. But as governments around the world ordered people to stay home, the daily operations of multi-million-dollar contact centres, especially those that are hosted on-premise, were instantly thrust into disarray.  

Some enterprises were able to manage this sudden shift since they had some form of digital customer servicing channels like live chat via instant messaging tools like WhatsApp or their web site or app. This was especially helpful in catering to customers and employees at home who saw an increased utilisation of live chat services by to 2 to 3 times the previous volumes.   

Conversational AI solutions are already being deployed by governments and hospitals across the world to do a basic level of patient triaging and screening.  

virtual care post covid

We can expect a continuation of these behavioural shifts even after the pandemic subsides, including more technologies that deliver virtual care in a low physical touch era. Telemedicine tools like video consultations with doctors, remote health monitoring apps and physical robots providing care to infected patients are some examples.   

 This does not mean a complete replacement of the human element in healthcare. Such technologies and robots should be seen as working together with humans as essential pillars of a dream team. The robots can take care of repetitive, non-patient facing, time consuming and high-precision tasks so that humans are freed to focus on actual care and improving patient satisfaction.  

wearable health

 

Outpatient Journey 

As hinted at above, the engagement with patients after their treatment is extremely important. In the future, we will see more hospitals placing more emphasis on preventative care. The day to day operations of healthcare staff revolve more around treatment than prevention. If they can spend more time on prevention, they are effectively minimising the chances of patients coming in, and thereby able to spend more time on more serious cases. After treatment, patients can also often relapse into a condition and end up back at the hospital in a worse condition than before, for more intensive treatment. 

Coupled with the growth of wearables and IoT devices, conversational AI systems will enable hospitals to care for patients in their homes before they even have a need to visit. This will free up the care teams who can focus on treatment for the more critical cases and emergencies in the hospital.   

AI-Enabled Counselling 

So far, we have only looked at physical ailments, chronic diseases and injuries. But we should not ignore the growing problem of mental health. <stat on global change in mental health and bots and services that address these.

Just like outpatient care, we can hope to see more conversational AI systems doing the bulk of the first layer of emotional support. This could be in the form of notifications, daily check-ins and gamification of positive habits. The more extreme cases can then be handled by qualified psychiatrists.   

Empathetic AI 

So far, the use cases of conversational AI have been aimed at automating repetitive tasks effectively. But healthcare is not just about effectiveness. Compassion, empathy, humanity and care are all attributes that are essential in any healthcare service provider. Take the roles of nurses and doctors. Their job is not simply to diagnose, prescribe medication, set up the equipment for treatment and help patients take their medications. A big part of their job is to make patients feel like they are cared for

Conversational AI systems have come a long way from being monotonous robots. Today the advanced systems have interesting personalities embedded into them and are sounding more human every day. There are even therapy bots, physical robot teddy bears and toys that have emotional care and compassion as the goal rather than effective automation of tasks. But these are still quite basic, predominantly aimed at children and not able to carry out extended conversations. In the future, as AI systems get better at automating repetitive tasks with better accuracy, the next frontier will be in perfecting the humanity part of these bots.   

 

Back to contents

 

Find out how KeyReply can help transform your healthcare institution with conversational AI.

Request For Demo