How the latest technology breakthrough is making a splash and its potential to transform healthcare
Everyone is talking about generative AI. Nearly overnight it has become the hot new innovation, promising to reshape our society and economy and driving investment in new companies leveraging this breakthrough technology.
James Currier, a serial founder and General Partner at venture capital firm NFX, believes we’re witnessing the beginning of a wave of innovation similar to the internet and mobile phones. Sequoia Capital prophesied that generative AI could create “trillions of dollars of economic value” and predicted a wide range of applications, from novel artwork to marketing and sales copywriting to automated code generation and beyond. The media and general public are eyeing this trend as well, with The New York Times recently proclaiming, “A Coming-Out Party for Generative A.I., Silicon Valley’s New Craze.”
The gold rush has already started among investors and corporate acquisition departments hoping to get in early on the next big thing. Jasper, a year-old copywriting assistant, recently raised $125 million at a $1.5B valuation. Huggingface raised $100M at a $2B valuation. Stability AI raised $101M at a $1B valuation. Inflection AI raised $225M at a rumored post-money valuation of $1B. And OpenAI, who raised more than $1B from Microsoft in 2019, is rumored to be raising additional cash with a $25B valuation.
Generative AI is a stochastic process, providing unique outputs each time it processes a given prompt. If you have not seen these tools in action, the results are striking.
Below is what DALL-E, OpenAI’s generative art API, provides as output for our simple prompt:
“Painting of heart surgery in the style of Picasso.”
Generative AI can create more than just images. One of the most important classes of application is text generation. Large language models like GPT-3 can perform a variety of tasks like answering natural language questions, summarizing material, and completing sentences from a prompt. For example, starting from “The symptoms of COVID-19 are,” GPT-3 continues with:
similar to symptoms of the flu. You may have a fever, cough, and shortness of breath. You may also have a runny nose, headache, sore throat, and extreme tiredness. The symptoms can appear two to 14 days after you have been exposed to the virus. Call your doctor if you have these symptoms.
GPT-3 can also generate high-quality synthetic scientific text. Based on the prompt, “Patients with diabetes are at higher risk for death and cardiovascular outcomes than the general population. We investigated whether the excess risk of death and cardiovascular events among patients with type 2 diabetes could be reduced or eliminated,” GPT-3 continues:
Design, setting, and participants
This study used the US Food and Drug Administration’s Mini-Sentinel program to examine the association between type 2 diabetes and mortality, myocardial infarction, and stroke among 6.5 million adults aged 18 years or older with type 2 diabetes.
Patients with type 2 diabetes were identified using claims and administrative databases from 12 health plans.
Main outcomes and measures
Death, myocardial infarction, and stroke were identified using claims and administrative databases from 12 health plans. Results There were 6.5 million adults with type 2 diabetes with a mean (SD) age of 61.2 (13.3) years and a mean (SD) follow-up of 4.0 (2.0) years. During the follow-up period, there were 17,746 deaths, 5651 myocardial infarctions, and 4985 strokes. Death rates were higher among patients with type 2 diabetes than the general population (adjusted hazard ratio [HR], 1.86; 95% CI, 1.80-1.93). Myocardial infarction rates were higher among patients with type 2 diabetes than the general population (adjusted HR, 2.76; 95% CI, 2.41-3.15). Stroke rates were higher among patients with type 2 diabetes than the general population (adjusted HR, 1.65; 95% CI, 1.41-1.92).
Conclusions and relevance
Patients with type 2 diabetes have a higher risk of death and cardiovascular events than the general population. The excess risk of death and cardiovascular events among patients with type 2 diabetes could be reduced or eliminated.
The above example shows that GPT-3 not only learned and replicated the style of scientific journal publication accurately, it also provided very believable content consistent with the original prompt.
The underlying neural network architectures that power these massive systems (transformer-based neural networks) are complex to build and expensive to train, so companies like OpenAI, Cohere, Huggingface and Stability AI are working to make generative AI accessible with tools and APIs that simplify usage and reduce costs.
Syntegra: Generative AI for healthcare
While flashy examples, such as the above art produced by DALL-E, capture the public’s imagination, other potentially more impactful applications have received less attention. Healthcare in particular is a vertical where generative AI can reduce the friction of data access, reduce physician burn-out and help automate manual and time-intensive tasks.
When we founded Syntegra more than 3 years ago, we identified generative AI as a solution to a widespread yet persistently unmet need: how to share individual-level medical data in a way that maintains all of its statistical patterns and utility, but guarantees patients’ privacy. Rapid, low-burden access to healthcare data opens up huge opportunities for researchers, life science companies, insurance providers and digital health companies to drive innovation in precision medicine, analytics and clinical decision support, and ultimately to accelerate advances in patient care.
Syntegra’s generative AI technology learns from and replicates any structured data — such as EHR, claims, registries or clinical trials. Our algorithm then generates patient-level synthetic records that are “realistic but not real” — accurately representing the clinical patterns present in the original data without the risk of exposing private information about real patients.
And we are only getting started. There are many opportunities to expand use of this foundational technology in healthcare. For example, in real-world data being used for research, Syntegra’s generative AI can be used to fill in gaps (also known as “missing values”), remove known biases in datasets, and help create more accurate training sets for machine learning models by expanding small or rare cohorts.
Despite Syntegra’s progress, we are still just scratching the surface of what could be possible in healthcare with generative AI.
One potential use case is using generative AI to improve physicians’ workflows. Clinicians spend excessive time with voluminous EHRs, impeding doctor-patient relationships and driving burnout. Generative AI is quite good at summarization, and could be used to review the medical record and provide a succinct summary relevant to the patient and the treating physician at the point of care. Over the next few years, expect an explosion of provider workflow tools leveraging generative AI, allowing more time to be spent with patients, as well as other yet-to-be-conceived use cases.
New drugs and devices take years and billions of dollars to make their way through lengthy clinical trials and regulatory approval. What if generative AI could be used to partially simulate clinical testing, leading to shorter time to market and lower costs for new treatments?
Given its leadership role in this emerging field, Syntegra will be both a participant in and an enabler of the generative AI revolution. Just as GPT-3 captures the essence of natural language, so do Syntegra’s tools capture the essence of medical knowledge. Syntegra is helping spread the power of massive transformer-based neural networks to many research and commercial medical applications.
Note: everything in green is generated by AI, and pink is prompt text
Authored by Syntegra co-founders Ofer Mendelevitch, CTO, and Michael D. Lesh, MD FACC, CEO