Privacy Preserved AI for HealthCare

Remember our example of John the refugee? As we think about how the app supports the five stages of his journey, we also need to think about how it supports public health analysis. I introduce Jane, who is a public health planner in San Paulo, Brazil. Jane wants to know where sick people are. So as John is engaging with his app and indicating his symptoms, Jane can start seeing the map of where the sick people are.

As John starts informing his close contacts through the exposure and notification app, Jane should be able to see how the contagion is spread, the people who contact John, and how they are traveling in some other parts of San Paulo. As John takes medications, he indicates through the app, which medications are working or not, and his current symptoms. Jane can take the aggregate data and realize which medications and which interventions are most effective. What’s more, as John gets medication on certain days and in certain parts of the city, Jane and her partners can see if this implementation conforms to equity and other ethical concerns.

Our system architecture has three components: Capture, analyze, and engage. For capture, we are building no-peek computational software. For analysis, we use privacy-preserved machine learning and NLP. For engagement, our platform extends to the vulnerable population. As we'll explain later, we also support paper credentials where individuals can simply collect QR codes along the way that are cryptographically secure as well as an SMS-based chatbot platform. John shares his medical history, his test status, what kind of tests he took, what kind of virus variant was recorded along with this test, and his vaccination status on the app.

We can capture John's symptom diary over several days, including how he recovered along with his treatments and medications. If John has tested positive, we also want to capture the Bluetooth exposure keys. In addition, significant sensor data from John's phone is also being collected, of course, all with no-peek privacy.

On the other hand, Jane is very concerned about the four goals we discussed earlier, which are understanding the prevalence and spread of illness, the effectiveness of treatment, and analyzing equitable distribution.

Our engagement platform allows John to see what kind of exposure alert he should receive. He understands what symptom trajectories look like over time. He can also see what the regional information is, and he can understand his own personal risk score based on activities he is planning.

The four main technologies that we use in no-peek primarily are minimum upload exposure notification, secure multi-party computation, differential privacy, and federated or split learning.

Last updated