UDLCO: AI healthcare complexity and hyperlocality

 Introduction:

Hyperlocal has been described as "information oriented around a well-defined community with its primary focus directed toward the concerns of the population in that community"

Also been described as referring to "the emergent ecology of data (including textual content), aggregators, publication mechanism and user interactions and behaviors which centre on a resident of a location and the business of being a resident"

Hyperlocal content has two major dimensions: geography and time. 

More recently, hyperlocal content has evolved to include GPS enabled internet integrated mobile applications which score highly on both the geographic and the time dimensions. They are capable of delivering content that is relevant not just in a community but relevant right down to the individual within a geographic area that can be measured in meters and blocks not towns and neighborhoods. They are also capable of delivering content relevant at very short timescales such as seconds or minutes not just days or weeks.

https://en.m.wikipedia.org/wiki/Hyperlocal

• "Communities can play a key role in addressing health equity issues through hyper local healthcare delivery models that tailor interventions to address specific local needs and culturally sensitive contexts. 
• Success in delivering on population-specific needs depends on health system’s ability to triangulate data from multiple sources, respond to needs with innovative and timely interventions in collaboration between communities and partners"

Unquote : https://kpmg.com/be/en/home/insights/2023/04/ps-hyper-local-community-healthcare-mode.html

UDLCO summary:

The conversations below begin with global discussion around AI healthcare as well as 2.0 healthcare and their limitations that can be tackled through hyperlocal data capture and feedback (blended with AI) content delivered along with an illustration of the same from a user driven healthcare hyperlocal data capture and feedback learning platform called PaJR.

UDLCO Keyword glossary:

http://userdrivenhealthcare.blogspot.com/2023/11/glossary-of-user-driven-healthcare.html?m=1

Conversational Transcripts:

[12/11, 17:49] SPAI Health: A poignant reminder on how hype-cycle works (and abysmally falls). Technology is a great tool but use-cases need to be hyperlocal. There is no one size fits all, especially when it comes to healthcare.

[12/11, 18:39] rb: Thanks for this interesting comment!

I was taking some real patients for our workshop with engineers on "tech end user driven healthcare' here and this will be a good opening line!

[15/11, 06:57] AT AI Healthcare: Here is a research publication on *Medical Appropriateness*.

https://www.springerprofessional.de/en/great-ai-in-medical-appropriateness-and-value-based-care/26487210

[15/11, 07:54] AT AI Healthcare: Both *Medical Appropriateness* and *Medical Necessity* are necessary for *Value-based Care*.

[15/11, 08:06] S: This is possible, but is a complex issue. Of course using SNOMED it is definitely possible to design such a system. One that is practical and is able to do it in real-time. Raise appropriate alerts and warnings. To be effective, the algorithm should be running in the background, constantly checking for inconsistencies and errors. Needless to say, the design is complicated.

[15/11, 08:20] AT AI Healthcare: You are absolutely right - this is highly complex. Therefore, we used *Neuro-Symbolic AI* to address this complexity real-time. 😊

[15/11, 08:32] S: I did some research of my own. AI vs coded approach. In healthcare, due to legal and ethical issues, it is imperative that errors are avoided or kept at a minimum. We medicos subscribe to the dictum of _primum non nocere_ - firstly, do no harm. AI models have an inherent error - mostly the training & validation data-related. Coded provides precision. It is binary - either true or false. That is why there is no alternative to using something like SNOMED. The large vocabulary, all coded by their concept definition using semantic triple and description logic, provides a wide coverage of the many information that gets collected during any clinical encounter.

[15/11, 09:31] ATAI Healthcare: I use *GREAT* AI. It stands for Generative, Responsible, Explainable, Adaptive, Trustworthy Artificial Intelligence.

[15/11, 10:19] S: You have my very best wishes in convincing the medicos.

[15/11, 11:00] ATAI Healthcare: My last comment on this subject. I don't try or even need to convince a medico. I use medical data to prove my models. Medical being safety critical domain, I know - if my model works on medical, it works in any other domain. Like if you can drive safe without using horn in Bangalore, you can drive anywhere in the world. 😊

[15/11, 11:05] Rkb: I completely agree with the later statement, if u can drive in India u can drive anywhere in the world 😀

[15/11, 11:12] rb: Provided you can pass their licencing exams, which appears to be easier to pass in India!

[15/11, 12:51] S: Having worked with both AI/ML and codes, I am firmly convinced that without codes CDSS is not advisable. Many times diagnosis is only possible on post-mortem. CDSS is used in treatment/management for which high degrees of precision are required.

[15/11, 12:55] ATAI Healthcare: I shall deliver lecturers on *"Architecting Secure Software Systems"* and *"Cybersecurity Cognitive Graph"* at NITK Surathkal at the *Five Days Faculty Development Programme On Software and Systems Security*, 9th to - 13th December 2024" (https://www.nitk.ac.in/upcoming_events). 

There are some vacant seats. If you know any faculty interested in Cybersecurity, please pass on this message.

[15/11, 15:31] SPAI Health: Interesting discussion. Both Dr.A and Dr.S I guess are medicos, but they seem to be so knowledgeable that I suspect they have a PhD in Computer Science too in the side🙂👍! 

One key thing we are realising in the AI community is that LLMs have limitations in specific domains, as they are optimised for freeform reasoning from large, disconnected datasets (case in point is the failure of Babylon to make the cut, which we discussed recently). Rule based use of logic (expert systems) exists from the days of IBM Watson and is perhaps best suited for healthcare but then it lacks the ability to take on board the latest and evolving scientific thoughts. There is also a hybrid track currently upcoming which tries to marry these two ends of the spectrum using domain specific knowledge graphs to upgrade and focus LLMs capabilities. There is also a need to go hyperlocal due to the wide variations in datasets across demographics. To some extent using RAG along with Edge AI is a potential solution.

[15/11, 15:58] S: Dr AT is a _bona fide_ computer scientist. Works in hardcore AI applications. Health is one of his areas of interest. I am a medical professional and health informatics specialist, majorly into data standards, interoperability and analytics. 🙂

[15/11, 21:37] rb: "From the doctor's summing up, Ivan Ilych deduced the conclusion that things looked bad, and that he, the doctor, and most likely everyone else, did not care, but that things looked bad for him. And this conclusion impressed Ivan Ilych morbidly, arousing in him a great deal of pity for himself, of greater anger against this doctor who could be unconcerned about a matter of such importance.”

Unquote

https://pmc.ncbi.nlm.nih.gov/articles/PMC3249844/

[16/11, 07:10] ATAI Healthcare: Here is Fig 8-1. Diagnostic approach to hemolytic anemia from Benjamin Djulbegovic's *Reasoning and Decision Making in Hematology* published in 2992.

[16/11, 08:31]rb: Excellent illustration of the global limitation in rule based algorithms in tackling clinical complexity (for those not handling real patients hyperlocally) further enhancing the point @⁨SPAI Health⁩ was trying to make. Ben assumes that if coombs negative,  HS would be the most likely diagnosis probabilistically as he's a US citizen but try telling that to an Indian or someone from Dhaka: https://pmc.ncbi.nlm.nih.gov/articles/PMC2975852/

[16/11, 08:34]rb: Many of our patients in our resource limited rural medical college are coombs positive but we don't jump to a diagnosis of AIHA as we know that it could be false positive and largely we still have to weigh everything on the basis of the patient event timeline (aka history) as well as clinical signs, external or internal.

[16/11, 08:55]G: In 1983, my consultant Prof SB was discussing using weighted scores for automated diagnosis. He asked around for symptoms and signs for possible appendicitis and said a more than a 10 score is final. The conversation went like this for the patient we were observing 

Pain abd ——1 

Fever ——1 

High TLC  ——1 

Loss of appetite  ——1/2

Vomiting ——1 

Tender rt Iliac fossa  ——10

Somewhat clinching the diagnosis

[16/11, 08:55] G: Unfortunately AI can’t examine for tenderness

[16/11, 09:10]rb: Yes it can't capture data unless fed. Not long before it grows up to robotic arms

[16/11, 09:11]rb: Couldn't resist sharing this happening in our home health PaJR right now!👇

Learning summary spoiler: AI LLMs are more likely to err about medicine facts on the side of caution lest they damage their customers egos!

Those who don't understand the script can ask meta AI to translate it for them.

Morning PaJR session:

[14/11, 11:36] PaJR moderator: Apnake baki der group a add korle apni unader keo eta ki bhabe kora jete pare sheta dekhiye dite parben?

[14/11, 19:12] Patient Advocate 63M Metabolic Syn: কোনটা ? AI LLM?

[14/11, 19:12] Patient Advocate 63M Metabolic Syn: 

জেলা শহরে চারটা নাগাদ একটা মশলা ধোসা খাওয়া হল।

[14/11, 19:12] Patient Advocate 63M Metabolic Syn: ছটা নাগাদ প্রাতঃভ্রমণে বের হয়ে পৌনে সাতটা নাগাদ গন্তব্যে পৌঁছে কিছুক্ষণ বিশ্রাম নিয়ে সাড়ে সাতটা নাগাদ চায়ের দোকান। ওখানে এক পেয়ালা চিনি ছাড়া দুধ চা ও কিছুক্ষণ পর এক বাটি দেশী ছোলার ঘুগনি খাওয়া হলে আটটা নাগাদ বাজার ঘুরে বাড়ি ফিরতে ফিরতে প্রায় পৌনে নয়টা। তারপর স্নানাদি সেরে সকালের খাওয়া সেরে ও ওষুধ খেয়ে কিছুক্ষণ বিশ্রাম নিয়ে এখন চলেছি জেলা শহরের দিকে। ফিরতে ফিরতে বিকেল হয়ে যাবে।

[14/11, 19:22] Patient Advocate 63M Metabolic Syn: এইমাত্র বাড়ি ফিরে এলাম।

[14/11, 22:11] PaJR moderator: Hain

[14/11, 22:15] Patient Advocate 63M Metabolic Syn: একটু শিখিয়ে পড়িয়ে নিলে অসুবিধা হওয়ার কথা নয়।

[15/11, 04:29] Patient Advocate 63M Metabolic Syn: চারটায় ঘুম ভাঙ্গার পর কিছুক্ষণ যাবত আপনাদের মেসেজগুলো দেখছিলাম। পাঁচটা নাগাদ শয্যাত্যাগ করব।

[15/11, 10:09] PaJR moderator: Apnakei sheta amader baki patient der jonye korte anurodh roilo jar jonye apnake unader group a add kore dicchi

[15/11, 17:25] Patient Advocate 63M Metabolic Syn: চেষ্টা করলাম, একবার দেখে নিয়ে জানান, আপনি যা চাইছেন সেটা হল কি না।

[15/11, 17:33] Patient Advocate 63M Metabolic Syn: স্থানীয় কিছু অনুষ্ঠানের কারণে সকাল সাড়ে দশটা নাগাদ বের হয়ে বাড়ি ফিরতে ফিরতে প্রায় সাড়ে চারটা। মাঝে এক -দেড়টা নাগাদ প্রথমে এক কাপ চিনি ছাড়া দুধ চা ও অর্ধেক স্যাঁকা পাঁপড় ও কিছু পরে এক গ্লাস লাল চা খাই। তিনটা নাগাদ নেমন্তন্নের খাওয়া একটা হোটেলে:

ভাত

কাঁচা শসা

লেবু

ডাল

পাঁপড় ভাজা 

মাছের মাথা দিয়ে বাঁধা কপি 

মিষ্টি কুমড়োর তরকারি 

কাৎলা মাছ ভাপা

পাঁঠার মাংস 

চাটনি ও

একটা বড় রসগোল্লা।

[15/11, 21:44] AI : ঠিক আছে! তাহলে খাবারের তালিকা হলো:

- পুরি/লুচি - ২টা (প্রসাদ) 🍞

- শসা ও গাজর - স্যালাড 🥒🥕

- তিন-চার টুকরো পাঁঠার মাংস 🍖

- আলু লাউ দিয়ে বুটের ডাল 🍲

[15/11, 21:44] AI : এখন খাবারটা আরও জমজমাট হয়ে গেছে! পুরি/লুচি তো দারুণ, আর পাঁঠার মাংসের সাথে আলু লাউ দিয়ে বুটের ডাল হলে তো স্বাদে কোনো কমতি থাকবে না।

[15/11, 21:44] AI : স্যালাডও স্বাস্থ্যকর এবং তাজা স্বাদ যোগ করবে।

[15/11, 21:44] AI : এটা একটা ভারী এবং পুষ্টিকর খাবার হবে! উপভোগ করো! 😊🍽️

[15/11, 21:45] AI : Let's break it down based on your updated meal:

1. *পুরি/লুচি (২টা)*: Approx 300 kcal (150 kcal each)

2. *শসা ও গাজর (স্যালাড)*: Approx 15 kcal (5 kcal for cucumber and 10 kcal for carrot)

3. *পাঁঠার মাংস (৩-৪ টুকরো)*: Approx 200 kcal (depending on the portion size)

4. *আলু লাউ দিয়ে বুটের ডাল*: Approx 200 kcal (this can vary based on preparation)

[15/11, 21:45] AI : *Total Estimated Nutritional Information:*

- *Total Calories*: Approx 715 kcal

- *Protein*: Approx 30g (from meat and dal)

- *Carbs*: Approx 100g (from puri, vegetables, and dal)

- *Sugars*: Approx 5g (from vegetables)

- *Fat*: Approx 30g (from puri and meat)

- *Fiber*: Approx 10g (from vegetables and dal)

[15/11, 21:45] AI : This meal is quite hearty and provides a good balance of carbs, protein, and healthy fats. The addition of salad adds some freshness and fiber, which is great!

[15/11, 21:47] Patient Advocate 63M Metabolic Syn: এ তো এক মহা বিপদে ফেললেন ! এ তো সবই বলে দারুন/ভালো। !!!

[16/11, 08:42] PaJR moderator: হ্যাঁ দুর্ভাগ্যবশত এআই এলএলএম আমাদের একাডেমিক নার্সিসিজম খাওয়ানোর জন্য ডিজাইন করা হয়েছে!

Yes unfortunately AI LLMs are designed to feed our academic narcissism!

https://24fpatient.blogspot.com/2024/10/63m.html

[16/11, 09:29]rb: Our current tryst with AI LLMs in our healthcare education and practice user windows can be summarised as below (albeit experiential anecdotal opinion): 

AI is like a new boisterous team member, kiddishly brilliant in places and yet a complete let down when you start to think you may begin depending on it!

When you are awed by it's brilliant analysis of study data that it has itself looked up you realise later that it's confabulated the paper and the references itself!

When you are beginning to praise it for its good analysis of a real paper URL you have provided, you realise it's changing the numbers every time you keep asking it repeatedly in different groups!

When you are happy it's able to do a marvelous thematic analysis of your thesis data you realise it's again fudged some numbers here and there!

Creative commons license: https://commons.m.wikimedia.org/wiki/File:Diagnosis_By_Radio,_1925_(telemedicine).webp#mw-jump-to-license