Data, Data, Data

No bricks without clay. Provenance is everything.

1. Who or what produced the data?
   A clinician, a patient, a sensor, or a model?
2. Why was it recorded?
   Care, billing, self-tracking, or research?
3. Under what conditions?
   At rest, exercising, in a clinic, wearing a cheap wearable that fails on hot and humid days that your aunty gifted you for Christmas?
4. What transformations happened?
   Filtering, averaging, imputation, classification…?

Where does it come from?

People can log things to personal diaries (digital or otherwise.) You can divine a lot of health information from other unstructured personal artifacts as well (e.g. text messages, photos of food, mood surveys.) This is really about their lived experience and is subjective by definition.

They can have smartwatches and other wearables that capture all manner of things: Heart Rate, SpO2, physical activity (steps), sleep, blood glucose, blood pressure. The nice thing with these is that repeated measurements can make personal baselines (and changes to them) visible. Note that this includes medical devices like insulin pumps and Continuous Glucose Monitors (CGMs) for those with diabetes. Anything electronic that generates a signal based on your physiological state.

They then engage with the healthcare system which generates its own clinical records (EHR.)

There are several other sources: you can get genetic screening. Where you live (e.g. in a highly polluted city), what you do for a living (e.g. exposure to loud noise), what you make, your race, gender, quality of healthcare access, how these evolve, and so on, are all ‘data’ that can be used to build representions of your state of health.

Together, these modalities form a Personal Health Data Ecosystem. It would be wise to note that no single source will give you a full picture of someone’s health and well-being. Heck, you can’t say this even if you combined all of them. We’re beautiful and complicated ♥️

What are some problems with these modalities?

We’ll do a few.

• You may write to your personal diary regularly. Or when there’s something happy or sad to write about. Or when there’s something to write about. Or you may get busy. Or you may not recall something accurately (especially many days after it happened.) Or you may not want to write about it. Or your diary might get lost. Self-reporting is subjective, noisy, and is shaped by recall, your willingness to talk about things, who/what is interpeting what you’re talking about.
• You may charge your wearable at night. Batteries degrade over time. Your aunty may have gifted you shitty wearable that’s not as good as an Apple Watch or FitBit and a hot day might lead to missing/incorrect readings.
• All your clinical/EHR data is collected when you’re inside the clinic and not after you leave. It is encounter-centered.

Fragmentation and Integration

The central question is this: How do we turn someone’s fragmented/scattered health records into something useful without claiming more than what the evidence supports?

Missingness

The Big Considerations

Measurement, missingness, error, inspectability, access, privacy, governance

How do Agents access data?

Well one common way is you get the agent to read the resources (fancy word for files) on your computer or via some upload. But that might not be a robust way to do things.

So there are two chief structured/standard ways:

Model Context Protocol. It specifies (a) Tools (b) Resources and (c) Prompts in a standardized way that agents can use to interrogate upstream sources of data to do their thing.