A Conversational AI Chatbot is Helping New Mothers Keep Healthy (#6)
Research in Translation
Introduction
Welcome back to ‘From the Computer to the Clinic’ - a newsletter about computational biology and its contributions to biomedical research.
In this newsletter, we explore how computational biology research can drive clinical progress. By sharing success stories in one disease area or domain of research, we aim to inspire the use of these successful approaches for other diseases and research areas also.
If you haven’t already, you can subscribe to this newsletter, or share it with friends and colleagues
Part VI
This is part VI of this series - if you missed parts I-V, you can find them on the home page of this newsletter.
The starting point of this series was the story of an AI chatbot trained to address the questions and concerns of new mothers after they leave the hospital. The chatbot was developed by the company Memora Health, implemented in the Penn Medicine healthcare system, and is currently in use by many real mothers caring for their newborn children. The Penn doctors helping to implement the system have taken to calling their chatbot Penny. Other centers will likely adopt new names as they partner with Memora and other companies to bring AI applications into the clinic - but the fundamental principles of how these chatbots work and why they are useful will remain the same.
Penn’s Healing at Home program is a growing success. Both Memora Health and Penn Medicine emphasize the benefits of the chatbot for patients and providers. New mothers are getting very specific responses to their maternal needs without the intervention of doctors, who can commit more time and resources to caring for mothers still in the hospital (Memora mentions on their website that over 70% of patient messages are addressed by the AI chatbot without the need for care team intervention).
Penn Medicine, in their report, note that the program is currently expanding beyond Philadelphia and into the rest of the Penn Medicine system. They estimate that once the program has expanded throughout the system, more than 16,000 new mothers will benefit per year. If this kind of system were implemented in hospital centers across the country and worldwide, millions of new mothers could benefit (by the most recent statistics, there are over 3.5 million births a year in the US). The chatbot is still being improved for those already using it and soon to adopt it, with a Spanish-language version, and enhanced support for mothers who give birth preterm and/or are suffering from postpartum depression.
It’s hard to say exactly what kind of code the chatbot is running under the hood because the software is proprietary. But it is certainly taking advantage of the developments in machine learning and automated response generation that we have discussed throughout this series. At the heart of modern AI chatbots are the deep neural networks that were introduced in part III. Over the past several decades, neural network architectures have been adapted to accommodate text as input (in the form of word ‘embedding vectors’), and structured so that the network, when it sees each new word in a sentence, can remember the ones that came before (see part IV). This capacity for memory was initially made possible with the advent of recurrent neural networks, and is now accomplished by more complicated architectures like long-short memory networks (LSTMs) and transformers. ChatGPT runs on the transformer architecture, and it is not unlikely that the Memora Chatbot is using it as well.
As we saw in part V of this series, the power of large AI models depends not only on their underlying architectures, but also on the quantity and kinds of data that are used to train them. In the 2010s, researchers were able to draw on vast amounts of data from social media applications to train powerful response generating AI chatbots. By the end of the decade, researchers were combining multiple forms of data, and training chatbots to not only produce human-sounding text, but also recognize named entities (like restaurants) in messages and speak about them intelligently in their responses.
It is not entirely clear how the Memora chatbot was trained, but the developers likely drew (with consent) from examples of conversations between new mothers and their clinicians, possibly stored in the hospital’s medical record + communication software (many hospitals use an Epic software system). From this training data, which this article about Penn’s Healing at Home Program notes was carefully curated by clinicians, the AI chatbot learned how to respond appropriately to common maternal health concerns and child care questions, such as “Is the swelling in my feet normal?” or “How soon should I feed my baby after their last meal?”
These capacities - to remember past conversation, produce human-sounding responses, and have some background knowledge of the medical concepts mentioned in messages - are all critical for a healthcare chatbot like Penny to be a successful digital health tool. New mothers are most likely to use the system if it is engaging and personal. It should be able to remember the mother’s name (and that of her new child/children). And when it’s being used for a multi-part conversation, it should keep track of what questions and answers have come before so it does not repeat itself or forget relevant information. It also needs to have some knowledge about the appropriate feeding schedule and other needs of new infants or the causes and remedies for swelling, bleeding, and other symptoms that new mothers experience. Perhaps most important, it needs to know when it is not up to the task. Users of the chatbot should always have the option to send a message directly to their clinical team, and the chatbot itself should recognize when the user is referencing symptoms that are potentially more serious and get the clinical team involved. Penny appears to be covering all of these bases quite well.
There is no shortage of companies developing AI chatbots for healthcare applications - but the successful integration of an AI chatbot into one of the nation’s major medical systems is something more rare. Other hospital systems can learn from Penn Medicine’s Healing at Home Program as a shining example, using AI chatbots like Penny not only to help new mothers, but also individuals recovering from major surgery and coping with mental health conditions.