Is Generative AI The Future of Equitable Healthcare?

For years we’ve been hearing about the potential for AI to revolutionize healthcare and health outcomes in a number of ways. As we continue to see an impressive and rapid growth of generative AI, we are starting to understand just how significant of an impact it may have on our healthcare. Recently, Mass General Brigham published findings in npj Digital Medicine demonstrating the ability of finely-tuned generative AI models to identify 93.8% of patients with at least one adverse social determinant of health (SDoH), compared to the only 2% determined by official diagnostic codes.

As it stands, most electronic health records (EHRs) are equipped with codes to record social determinants of health (SDoH). However, there is a breakdown as it’s not common practice to capture SDoHs on intake or check-in forms, or when registering at a clinic or hospital. SDoHs include safe housing, transportation, pollution, income, and access to opportunity, among a myriad of other factors. 

The correlation of SDoHs to health outcomes has largely been overlooked and under-researched, but with the emergence and increased promise of precision medicine SDoHs are receiving increased awareness. However, a number of challenges are arising, including the need to balance documenting patient details, while combating alert fatigue for practitioners using EHRs. Enter generative AI, which is proving to have the capability to read EHR records and transform how social determinants of health are identified and documented. The result is being able to equip medical professionals with detailed adverse SDoH information on a patient, so that they can deliver more targeted care. 

In Mass General Brigham’s recent study, they applied specialized large language models (LLMs) to electronic health records to identify under-documented social determinants of health. Their results are extremely promising and have shown to vastly outperform any existing method. 

Building Large Language Models

Large language models are a type of generative AI capable of developing an understanding of the semantic meaning of text. In other words, they can interpret natural human language. To do this, the model is trained on a representation (embedding) of a massive volume of example text, typically scraped from the web. By seeing millions or even trillions of examples of words, the model can recognize not only the frequency of words but also their contextual meaning. 

As an example, a large language model will be able to discern from the surrounding context whether the word “right” means the opposite of “left” or is a synonym for “correct.” Large language models are currently best-known as the technology that powers chatbots like ChatGPT, but they serve as general-purpose models that can be tuned to perform a variety of tasks such as recognizing and extracting meaningful entities, such as, in this case, social determinants of health.

What About Bias?

As with any healthcare advancement, generative AI has been scrutinized for its potential to perpetuate bias. Since large language models learn from a massive amount of data, they can quickly internalize and perpetuate the existing biases, prejudices, and racism from which they are trained. Knowing this, the team carefully fine-tuned and monitored the algorithmic biases to ensure that the models were not making existing health disparities considerably worse than they currently are. As a result, these specialized large language models demonstrated reduced bias, enhancing their utility in diverse patient populations.

In conclusion, integrating generative AI with electronic health records could offer innovative opportunities to identify and address social determinants of health. If we can monitor algorithmic bias, and train these large language models with good data and ongoing research, generative AI holds immense potential to transform healthcare delivery and promote health equity.

Sign up for our newsletter to stay up-to-date and get early access to the latest insights, trends, and breakthroughs in drug discovery, powered by cutting-edge data and analysis from industry-leading experts.