Making healthcare more accessible to all through AI

Anyone involved with the healthcare system knows that doctors spend an enormous amount of time reading and writing documentation. In many cases more time than they spend interacting face to face with patients. This leads to inefficiencies, burnout, and ultimately declining availability of healthcare.

But we can use technology to help. Over the last year we have seen a revolution in artificial intelligence techniques processing language, answering questions and generating text. This technology is a natural fit for many of the challenges that come up in medical documentation.

Google Cloud and MEDITECH have been working together to leverage Google's state of the art machine learning models to process EHR data and streamline tasks for physicians. The first result of that collaboration is a system that creates a draft hospital course narrative for discharge summaries. And it will be revealed this week at the MEDITECH Live 2023 event.

This system automatically processes long clinical charts corresponding to a hospital stay, identifies key information, and drafts a summary. The summary is then displayed to a physician who reviews it, and can decide to include it in the discharge note or edit it.

I believe the right path to deployment of AI in critical environments such as healthcare is through AI + human systems, in which we have solid safeguards in place, instead of trying to drive full end-to-end automation.

At the same time, the potential impact of a system like this is massive. A clinician who is processing a discharge spends half an hour to an hour reading the clinical chart, identifying key information by hand and assembling it into a note. Much of this work can be done with very high accuracy by an AI. And then reviewed in a few minutes by a human expert.

Of course, building a system like this requires substantial technical work. Large language models (LLMs) have become much better over the past year, but they are still not good enough to achieve the level of accuracy necessary for this task with a simple prompt. Moreover, the LLMs need to be deployed, maintained and integrated with other complex infrastructure.

A variety of techniques are available to prompt engineer, fine-tune, and retrain models. As well as novel system architectures combining ML models with retrieval for summarization tasks. Some of these techniques were used for example to train Med-PaLM, Google's large language model specialized in medical question answering.

There are more than 30 million hospital stays in the USA alone every year. Imagine the impact of saving over 30 million hours of work for doctors across the country. This means more time spent in front of patients and reduced healthcare costs for everyone.