Contextual Personalization: AI's Personalized Approach to Health

I agree: Tracking and more tracking

 When the cookie was first developed in 1994, Lou Montolli at Netscape consciously was avoiding advertisers from using his solution to make websites remember visitors to track users. Fast forward to today, on average each website you visit tracks your behaviour on it with 10-20 odd cookies. No wonder the EU wanted data residency and tightened the noose around developers with regards to privacy. If that has resulted in better privacy is yet to be ascertained, but it has certainly resulted in poor UX - you having to grant permissions and accept more terms to be tracked before you carry on with your online shopping.

In the healthcare context, you roam around with your baggage of cookies in the form of multiple case histories. Consider the example of getting second or third opinions for a rare auto-immune disorder – a John Doe would be lugging his bag full of histories, investigations, scans to a new physician every time he isn’t comfortable with his Dx or Rx. All of these could comfortably reside in a .txt file equivalent of a cookie for each institution or provider you interact with. Of course, provided you click on “I agree”.

The overhead of storing this additional information is negligible, in fact most EMRs and EHRs even in non-reimbursed markets do this systematically. Hl7 messaging has defined the protocol for transmitting these records as well. Then why do we feel Digital Health 1.0 has failed us?  There was a slew of Blockchain startups that promised this, but have largely had little impact. But as a user, and as a developer, we need to empathise with our Montolli moment, we need better tracking in Healthcare.

X =/= Y:  Interoperability

Of all the complex technical jargon that gets thrown around in any conversation about Interoperability, I’ve seen the two biggest challenges with

1.        Legacy systems are still in use

2.        Data being fragmented in siloed systems.

 While infrastructure modernisation needs it’s own snailish pace and we’ll get there, however, in 2024, it amazes me that we label it as a challenge when the average tech savvy user walks around with 2-3 forms of personal compute with them (Phones, Smartwatches, Laptops, Tablets – Long live the ruled diary).

Every entrepreneur at some point in their pitch deck might have mentioned that they would eventually solve for siloed data. But the incentives at a systems level never align for them to achieve progress. The fascinating exceptions of-course are where the payor and providers are the same (eg. NHS in the UK or Kaiser in the US). In out of pocket markets, getting data out from either a payor or a provider would require ₹₹₹ - something only a policy change could affect. But overcoming these and other such challenges will be crucial over the next few years. Only then can we have elaborate context to hyper-personalize healthcare.

We do have some developments in terms of stacks – eg. India now has ABHA which could be the single source of truth going forward. Perhaps this could be another UPI moment where we also export this?

WIIFM??: Context Matters

Every LLM trained by either a Tech behemoth or a tiny company with access to GPU clusters has ingested medical texts. The clinical heuristics have been “Seen” by most models. In fact the standard question to test for Diabetes Management “Can I eat a mango” includes the repertoire of pros, cons, exceptions and what not. But for the average user, that is an extremely generic response. Let alone, being actionable.

As a developer I love to train JEDi with more personal context for every user, so that every response ends with delight. Finetuning gets you there too, but only to an extent. RAG gets you there, but only to an extent. But where we stand today, hyper-personalised responses are still limited. And it’s something we invest actively at Fitterfly in terms of our research. 2024 will see more of these evolve into standalone QnA systems or integrated into Expert Systems. With input costs for inference also coming down, it would be interesting to see how far we could go. I'm certain we will have quicker resolutions or happier paths in user journeys (eg. Diagnostic pathways becoming quicker like searching on Perplexity vs Google).

Conclusion

We need all 3 to occur for an ideal personalised healthcare experience – Better tracking done ethically, out of the box interoperable solutions, and personalisation designed ground up into AI models. You prefer a concierge because they understand your preferences. You prefer a restaurant because the staff remembers your choices. Our GPs/ Family Physicians are great because they remember and respond based on the context of them seeing us grow old.

The question to ponder here then becomes “How might we overcome systemic barriers towards contextual hyper-personalization?”