EIT Health Germany Series 3: The Impact of Voice Tech in Healthcare (Julia, Hoxha)
Voice tech is one of the tech areas with high potential to optimize healthcare processes for providers and ease chronic disease in management for patients. In reality: How far are we from futuristic ideas where everything would be operated with the help of voice? How do innovators reduce the risk of misunderstandings in designing voice technologies? What does the development of human-like bots look like?
Voice technology has tremendous potential in healthcare but is still in its infancy.
“Voice technologies used for transcriptions reduce administrative tasks and the burden on physicians and medical staff. Solutions, where physicians can dictate data into the EHR, have gained considerable use and traction by today,” says Dr. Julia Hoxha, co-founder and CEO of Zana - a European startup that provides healthcare organizations with the technology to design and to deploy their own chatbot and voice assistants. We are still not at the stage where we would only use voice as a tool for navigating healthcare software and technologies.
Voice technologies are predominantly used in combination with texting. Other existing applications of voice technology include the use of this technology for increased patient engagement, improved patient adherence to treatment, and better modes of interaction.
Digital voice assistants for healthcare
Zana offers healthcare organizations a platform to design and deploy their own chatbot and voice assistants. One of the company’s lead products that aim to become a prescribable app reimbursed by health insurance in Germany is a patient companion for patients with heart failure. Zana’s mobile phone chatbot and voice assistant are used to improve adherence to treatment by encouraging patients to report their symptoms and vital signs more accurately. Users or patients are able to ask questions, and the patient companion within the mobile application answers them. The interaction and responses of the digital companion are based on national patient guidelines for heart failure, which are fed into the system. Additionally, another patient companion app for COPD & Asthma is undergoing validation through a clinical trial in the Netherlands.
How can AI understand humans?
One of the most frustrating experiences when using voice assistants is if the user has an accent and the voice assistant misunderstands the message the speaker was trying to convey. The solution to that is structuring data, using domain knowledge and ontologies. “Natural language understanding consists of translating speech to text and then understanding the intent of the user. This is a continuous struggle and technology needs continuous self-improvement in the domain. Let’s take the example of the companion for heart failure. We know that the question at hand relates to symptom tracking of heart failure. We already have a very well-crafted database of possible symptoms. During the speech recognition process, if something like fatigue comes up and it's confused with another word, we already have a predefined acceptable set of options which we can basically pre-filter and make a better prediction of what is the correctly classified meaning in a given case. Structured knowledge and fine-tuning by using domain knowledge is very important,” explains Dr. Hoxha, whose previous research work and know-how is transfered into the technology now applied at Zana. If the first problem is to make sure an AI app gets understanding right, the second challenge is, how will an AI assistant be as emphatic as possible to a different set of users?
How do you make sure that conversational AI chatbot or assistant is as human as possible?
Zana has a 2-step approach to this challenge. First, they gather as much data as possible through openly available resources such as patient forums. Their in-house conversation design team works with a network of doctors to consolidate data related to heart failure reporting. “For example, let's say we have around 13 different symptoms in acute heart failure. To enable the bot an accurate understanding we have put together a data set of almost 2000 different formulations for these 13 symptoms,” explains Dr. Julia Hoxha.
The second part of the process is the system’s self-learning and improvement over time. The more the system is used, the higher accuracy it achieves. Throughout the process linguistic and psychology experts are included in the design to adapt to different user personas.
What are the hardest things to tackle in conversational AI?
Technology.
Technology hasn’t reached maturity yet in speech recognition or natural language processing to have high accuracy in classifying different formulations or different dialects of the clinical language.
Dialogue management approach.
“Right now, most platforms rely on a logical/rule-based approach in crafting the conversations. There is further potential and need for more dialogue data and advanced algorithms with machine learning-based approach to improving the conversational experience,” says Dr. Hoxha.
Market transfer.
Healthcare is a highly regulated industry. A lot of regulations are already in place but even more, are missing. “Many standards are not well established yet. What is the evidence for an AI system so that we can get the credibility and gain user trust?” mentions Dr. Hoxha.
In 2022, Zana joined an exclusive group of promising early-stage European healthcare Start-ups supported by the EIT Health Investor Network. Following an in-depth vetting process, Zana has been rated as “Gold Company” in the assessment performed by expert evaluators and the investment committee.
CEO Dr. Julia Hoxha believes this will bring the startup additional visibility to attract investors: “The added value of the network is that it includes almost everyone in the European healthcare scene. Being a part of the investor’s network and being rated as the gold company will hopefully ease access to investors and raise awareness about the company to foreign investors as well.”
