The Healthcare AI Triad: The Future
Exploring how technology and culture could affect the healthcare AI triad
This month we’ve reviewed the Healthcare AI Triad of Algorithms/People, Compute, and Data. All of these aspects need to be considered when we discuss the future of AI in medicine, and thoughtfully approached to protect patients while encouraging innovation. Today we’ll look toward possibilities for the future of the triad.
Data
Synthetic Data
Synthetic data, or data created by AI, bypasses many of the thorniest issues in healthcare including:
Data privacy, since the patient records created aren’t real
Data bias, since records from traditionally underrepresented groups can be created synthetically to provide a more statistically robust result
There are many synthetic data generation (SDG) approaches, but often:
Their methods are not well-described which makes them difficult to repeat and validate
They’re not validated, or only superficially validated.
Some of them have elderly men who get pregnant, etc
There’s currently no definitive benchmark or standard for validating synthetic data
In the future, synthetic data may become sophisticated enough to provide ways to fill in gaps in AI training data and give researchers a way to test hypotheses without accessing PHI.
Data captured by AI
Ambient scribes: One underappreciated aspect of ambient scribes may be its ability to capture and analyze large amounts of clinical data that we didn’t have access to previously. Data from ambient scribes may prove more accurate than the copy-and-pasted mess in the EHR, and possibly provide some unexpected insights.
Wearables and remote patient monitoring: Wearables will likely generate a ton of data. Whether it's incorporated into training sets or analyzed will depend on the developer and financial incentives, since the data are usually “owned” by the maker of the software product instead of a non-profit hospital or government entity.
Patient consent and compensation for data
There are some creative ideas out there about how to get consent for PHI use by AI and to compensate the individuals rather than letting the companies profit from the data. These include using Blockchain to pay patients, or even using a “collective data management” system similar to the “collective rights management” system in which musicians are paid when their songs are used.
Physician consent for data collection
Physicians will need to advocate for:
Requiring consent to be monitored
The ability to refuse to have their productivity and performance data used without their knowledge
Increased participation in systems that involve large amounts of data collection from physicians
Compute
Increased parallelization is the theme with both quantum and neuromorphic computing, which are on the leading edge of compute technology.
Quantum Computing
Everyone remembers (maybe?) from physics that:
“unmeasured quantum states occur in a mixed 'superposition', not unlike a coin spinning through the air before it lands in your hand”, and the information of a qubit (analogous to a ‘bit’ in classical computing’ is in a state of
Qubits are the quantum equivalent of a classical computer bit
Superposition represents a combination of all possible configurations of the qubit.
Groups of qubits in superposition can create complex, multidimensional computational spaces.
This allows quantum computers to run many simulations at the same time (parallelization)
Quantum has several problems that would need to be solved including:
Not a lot of hardware suppliers like high-end lasers, cryogenic electronics, superconductor cables
The hardware has to be kept very cold (just about at absolute zero) to ensure accuracy, which introduces a lot of room for disruption
“Quantum computers are relatively well-understood in theory, but despite billions of dollars in funding from tech giants such as Google, Microsoft and IBM, actually building them remains an engineering challenge.”
Neuromorphic chips
Neuromorphic chips are built to mimic the concept of synapses and neurons, and are massively parallelized, meaning that they can do a lot of operations with much less power.
Currently all the public and private models use traditional silicon and metal oxide semiconductor technology, but researchers are trying to find ways to use less conventional materials
There has not yet been “a machine learning algorithm/application combination for which neuromorphic computing substantially outperforms deep learning approaches in terms of accuracy”
“Neuromorphic chips have been built with existing technologies, but their designers are hamstrung by the fact that neuroscientists still do not understand what exactly brains do, or how they do it.”
Algorithms/People
Algorithms
Already, we’ve seen smaller models like Llama from Meta that are close to behemoths like GPT in performance. It’s not clear how many more enormous foundation models we’ll need; it’s possible that we can fine tune for the vast majority of cases with smaller models with less compute. And if we have data that are high quality, we can reduce the size of themodel even more. A recent example is AlphaFold, which can predict how proteins will fold and decreases the time it takes researchers from months (?) to days (?) only trained for a few weeks on 16 TPUs, and cost in the 5 figures instead of the millions that GPT-3 cost to train.
People
CSET states that “Improving algorithmic efficiency and overcoming parallelization bottlenecks in training are difficult problems that require significantly more human expertise than simply purchasing more compute. This suggests that the path towards continued progress in the future rests far more on developing, attracting, and retaining talent than merely outspending competitors.”
People may actually be the most difficult part of the AI triad. Ensuring appropriate skills training that starts at a young age and continues through college is a multifaceted and expensive proposition. Some countries are trying; Finland is attempting to teach 1% of its population AI fundamentals.
In the short term, some AI policymakers recommend immigration reform to help the US stay competitive and ensure enough knowledgeable workers to maintain US dominance in AI. Because this requires political will, it’s also a less straightforward solution.
Clinical AI Triad of Death
After discussing the very sophisticated approaches that AI may use in the future, I want to discuss a more immediate issue: the threat of non-adoption of AI. I love this Triad of AI Death in Clinical Practice because it addresses the real issues that AI will need to overcome relatively quickly to be successful:
Poor quality data
Ethical and legal issues
Lack of educational programs
As physicians, we need to advocate to mitigate those issues in order to provide our patients with access to better care and ensure a positive practice environment. Physicians can help address these barriers. Don’t let healthcare AI die.
So far this month we’ve discussed:
The future of the AI triad
Hope you’ve enjoyed this series. Let me know what you think about the future of the Healthcare AI Triad in the comments, or email me at machinelearningformds@gmail.com