<img src="https://images.squarespace-cdn.com/content/v1/58f645cae58c62dbeca6bc2c/1605285458326-ZA7QV1XGTOP6BVIXZM0G/34.jpg" alt="34.jpg" />
by Vagelis H. 12/15/2020
Anybody who has worked on creating chatbots for healthcare providers (or any other goal-oriented chatbot) soon realized that there is no magic AI. In commercial chatbot platforms, AI is used in limited aspects of the chatbot, such as intent classification (e.g. patient has billing question vs. tries to make a new appointment) or information extraction (e.g. map “tomorrow” to a date). For more details on AI and its limitations for healthcare chatbots, read our previous articles on bot capabilities and bot technologies.
In all commercial platforms, the chatbot creator has to manually define some type of flow diagram (or slot filling), which must try to anticipate as many as possible inputs from the patient, while advancing the conversation towards the chatbot’s goal (e.g., make an appointment).
SmartBot360 has been working with a large number of healthcare providers and organizations since 2018, building and refining chatbots to attract new patients or serve existing ones. Here are some lessons learned, which can help in creating more effective chatbot flows:
Allow free text input only if you have free text processing. For example, if your chatbot can only understand “yes” or “no” for a specific question (e.g. “are you a new patient”), and any other input will lead to “sorry, I didn’t get it,” don’t allow the patient to input anything else. This could be achieved by selectively disabling the free text message box during some parts of the bot.
The most common “uninvited” question is “do you accept insurance X?” If you are a healthcare provider, no matter what the chatbot’s welcome question is, e.g., “what’s your name,” many patients will immediately ask about insurance. The reason is that they don’t want to waste their time interacting with the bot and then find out they are not covered. So, every healthcare chatbot must have a mechanism to catch such questions. The simplest way is to have a text-contains component that matches terms like “insurance”, “blue cross”, “anthem” and so on.
Appointment (re)schedule. Close second after the insurance-related questions are appointment rescheduling questions e.g., “I want to reschedule the appointment I have tomorrow morning.”
List of services. Another common category is questions about the offered services, e.g., “do you do liposuction?”
Bot or human? Another popular question in healthcare bots, as in all bots, is something like “are you human or bot?”
Live chat. A large number of patients may also ask to speak to a human at any point.
A key challenge to address all the above cases is that these patient utterances can occur at any point during the chatbot flow, which we refer to as “off-script” utterances. Some chatbot platforms may allow designers to define different intents for each of these patient requests. A problem with that is that it breaks the flow of a goal-oriented chatbot. Another approach, employed by SmartBot360, is to augment flow diagrams with a knowledge base of common questions and a language generation module.