In her upcoming book, Wealthcare: Demystifying Web3 and the rise of Personal Data Economies, Brigitte Piniewsky, MD describes how technology will disrupt the business of healthcare while actually improving health. Web3 is a buzzword for new decentralized tech not controlled by giants like Meta and Google. Also central to the book, machine learning and artificial intelligence (ML/AI), such as ChatGPT, are addressed for their ability to digest vastly larger pools of data than humans (or even teams) can process.
ChatGPT blends technologies like Google search and human-edited content, such as Wikipedia entries, along with a human language model and communication skills such as threaded conversation (chat). The result, whether a written research review or a painted picture, seems intelligent because it is presented in a familiar, human way.
Medicine, like physics, is a science that processes physical measurements into hypotheses and more-or-less useful predictions. Weather is a well known example of a non-medical science that uses the largest computers running constantly improving prediction models to process growing numbers of physical measurements. As the sensor networks expand in both scope and quality and the models improve in sophistication through a combination of ingenuity and chance, forecasting improves. But ingenuity is not strictly a human trait and chance can masquerade as ingenuity. Because computers can test far more hypotheses than humans can, the adage “It’s better to be lucky than smart” is actually true for ML/AI. Modern AI vastly surpasses the human ability to take chances and, properly harnessed, can look positively ingenious.
But unlike physics or weather, medical AI depends almost entirely on measurements linked to an individual human. Health records are only the tip of the personal information iceberg. 80-90% of the critically relevant data - social determinants of health (SDOH) - are not in a hospital or government database. SDOH includes employment, family, education, social interaction, location, access to food and nutritional status, services and purchases. Almost without exception, these measurements are private and typically accessed through an individual’s smartphone, web browser, or some other personal “user agent”. Google and ChatGPT can’t see them and almost none are posted publicly by individuals.
So, your smartphone knows most everything about you. But how do you put it to use? There are a few choices:
Don’t put it to use. Health care improvement will thus continue to be slow and very expensive. Disparities will continue. Private interests, such as data brokers, who “take” disjoint pieces of your data without notice will continue to use it for their benefit, not yours.
Share the data itself with one or more institutions. Which institution would you trust to hold your lifelong health records and SDOH? What would keep them accountable for using the data only for good?
Choose and install apps that will process the data locally without sending it out and then share only the insights. When many others like you allow the insights to be combined, medical science and policy can advance while preserving individual privacy.
Let’s call this third choice federated machine learning1. Federated because the data you and others contribute lead to insights that can advance both individual interests as well as medical science. Ideally, it’s ML/AI that you own, running for your benefit on a smartphone or cloud service that you pay for. Like a good doctor, your ML/AI should be working only for you.
Sadly, federated ML/AI is not quite ready for prime time. Even when it becomes available, your smartphone operator may not allow the software to load because it poses too high a security risk. Who would you trust to add software to your phone that has access to “everything? How could you be sure that the software is working only for you?
To summarize, artificial intelligence is currently learning only from public data; the same kind of data that you might find with a Google search and Wikipedia. ChatGPT and others like it are challenging how we teach, see art, and do democracy. However, to transform how we do health care and health policy, ML/AI will need access to very sensitive data that describes us and our lives in intimate detail. As a society and as individuals, this is uncharted territory. The benefits may be significant but the risks are paramount. Future posts will discuss paths to a solution, including technical advances and potential regulation.
In the meantime, check out Brigitte’s short book available Feb 7th on Amazon: Wealthcare: Demystifying Web3 and the Rise of Personal Data Economies.
Our team of physicians and engineers at HIE of One is currently prototyping Trustee®, a patient-controlled health record designed to support federated learning for clinical, research, and public health. With Trustee®, the decision to share information, whether for medical care or research, resides with the patient, not a data broker, EHR, or other third party. If patients choose to sell their data directly to an end user or indirectly through a federated learning initiative, they can specify the acceptable payment and privacy guardrails. And with Trustee®, the physician or research end user will know the information is accurate and complete.
For more information about Trustee®, contact Adrian Gropper, MD info@trustee.ai.
I asked ChatGPT: What is federated machine learning?
Federated machine learning is a distributed machine learning approach where multiple parties train a machine learning model on their own data, without sharing or pooling the data. Instead, the model is trained locally on each party's data and the trained models are then combined to create a global model. This approach allows parties to retain control over their own data while still benefiting from the collective knowledge of the group.
Well said Adrian! There is a huge hole in the data used to generate health intelligence, now with new tools we solve for the gap while ensuring individual's maintain control over their private details.