A view of the future from 2017: 'omics, AI, and precision medicine
I find it useful to start with a clear vision for an ideal version of the future to guide and motivate my work, written in as much detail as possible. Building that future then becomes a matter of reverse engineering.
What follows is my attempt to paint that picture in 2017. The broad strokes of this vision continue to guide me today.
We live in a world where Google knows what you plan to search within two keystrokes, Siri listens more attentively than your own children, and Facebook can pick your face out of a crowd.
Although the new-found abilities of artificial intelligence sometimes border on eerie, their benefits are undeniable. Indeed, we have come to rely on them. As our ability to generate data exponentially increases, these systems provide an indispensable augmentation to our own intelligence, drawing out the meaning we need, when we need it.
In no field do we stand to gain more by this augmentation than in medical practice. Medical decision-making is often carried out by weary human minds asked to grapple with staggering complexity under tight time constraints with lives at stake.
Further, we can now generate massive datasets like the sequenced human genome, rich in important medical information that is far too nuanced for human minds to comprehend. We celebrate small victories as we come to understand parts and pieces of such massive datasets, but their full value will be realized only when we apply artificial intelligence to help us draw meaning from the set as a whole; when we need it, where we need it. In what follows, I hope to convey a clear picture of how this world may look and what key steps will be necessary to get there.
The World of Big Data, Artificial Intelligence, and Precision Medicine
Imagine yourself 20 years in the future. You get sick and go to see your doctor. At the clinic, a nurse takes a sample of your urine, blood and feces. These samples are sent to the laboratory, where they are rapidly analyzed to determine the concentrations of thousands of chemicals in your urine and blood, and the identities and abundance of thousands of microbes in your stool.
A physician interviews you and performs a physical examination. Other imaging and laboratory tests are ordered as needed.
All these data are placed in a database with the rest of your medical data, including your genome.
Artificial intelligence applications learn the patterns of health and disease from your data and data from millions of other people. These AI systems discern the patterns in your data that explain your current illness, and they generate a report for your physician that describes the pathological process underlying your current problem. They also point out the interventions that will be most effective for someone with your specific genetic background and current physiologic state.
Your doctor discusses these interventions with you to determine what will work for you. If some intervention won’t work as proposed, your physician notes your constraints and asks the application for the next-best options, based on your constraints.
By receiving the best intervention for you as your first intervention, you avoid all the harmful side-effects, the lost time, and the waste of trial-and-error medicine. Society conserves its healthcare resources to provide the same high-quality care to others.
In this future, it is a basic right to have your medical data securely stored on government-run servers accessible by your medical providers through privately-owned electronic medical record systems. Private businesses develop artificial intelligence systems to learn from the data, helping clinicians to derive meaning from the vast amount of data generated on each of their patients.
From Here to There
Now let us return to present day. The key technologies required for this vision of precision medicine are already routinely used in research, but we must surmount a major hurdle before their full potential can be realized in the practice of medicine: the AI systems that will learn from the database will not start providing real value until there are many thousands to millions of entries in the database. Thus, there must be an initial investment in database infrastructure, big-data analytics, and the transfer of existing medical records in standardized format into the database. There must also be a large cohort of people from the general population who are willing to place their medical data in the care of these servers. No existing private healthcare systems can support such a far-sighted investment*. Yet, once these systems are in place, they will drastically reduce healthcare spending in the form of ineffective treatments, missed diagnoses, over-testing, and over-treating.
Once enough entries exist in the database, AI systems will begin to provide high-quality diagnostic, prognostic, and therapeutic answers to our clinical questions. When the value of these insights exceeds the cost of performing big-data analytics, healthcare systems will begin to upload patient data at their own expense as the standard of care. As we approach this tipping point, government subsidies for big-data analytics can be weaned off. In perpetuity, the government need only provide secure servers with standards for the format of medical data entered. Any medical record system that interacts with the government database would have to follow the same high standards of privacy and security that are already in force for today’s electronic medical systems.
This database would also become the world’s most valuable scientific resource, holding deeply informative answers to important questions in a variety of fields. Health-impacting chemical signatures in blood or urine could be tracked over time, mapped across the country, and associated with specific behaviors. The tremendous work that has gone into understanding the impact of individual genes could be easily expanded by AI systems to appreciate the nuances of gene-gene and gene-environment interactions on a scale that would exceed the capacity of the human mind to comprehend.
At the time I wrote this vision, I didn't know about two developments:
An incredible piece of legislation that will safely and responsibly open up access to medical data on a scale we have never seen before (more on that in another post)
Federated learning, which is an approach that allows AI to learn from data where it is, without the need to gather it all into a central repository.
What do you see for the future of medicine over the next 5-10 years? Let us know in the comments below!