User driven health care: UDLCO: User driven healthcare utilities of AI LLMs and the guard rails for medical cognition

Conversational Transcripts:

[30/09, 07:31] +91 DT: https://www.linkedin.com/posts/janbeger_this-paper-presents-a-comprehensive-framework-activity-7245983144472940544-5b2u?

utm_source=share&utm_medium=member_android

[30/09, 07:51] SK: Can we have the full paper

[30/09, 08:02] PaJR moderator: Given how physicians learn, many of those pieces in the pie would overlap in the distribution of their medical cognitive sauce or cheese!

Here's a sample seen today in one such physician CBBLE (pronounced cable). Let me know which piece of the AI pie does it exclusively bite 👇

[28/09, 06:35] Physician : 29 yr old medico girl present with hyperglycaemia with ketosis , managed with insulin. No h/o diabetes in parents . One mama is diabetic. Index patient n mama both have vitiligo. Girl is ANA positive. Now what is this ?

[28/09, 06:35] AI LLM : A fascinating case!

Considering the patient's presentation, family history, and laboratory findings, here's a possible diagnosis:

*Likely Diagnosis:* Latent Autoimmune Diabetes in Adults (LADA) or Type 1 Diabetes with Autoimmune Features

*Rationale:*

1. Hyperglycemia with ketosis: Suggests insulin deficiency.

2. No prior history of diabetes, but family history of diabetes (maternal aunt): Increases suspicion for genetic predisposition.

3. Vitiligo in patient and mother: Indicates autoimmune tendency.

4. ANA positivity: Supports autoimmune component.

*LADA Characteristics:*

1. Adult onset (typically >30 years, but can occur earlier)

2. Autoantibodies (e.g., GAD, IA-2, or ZnT8)

3. Insulin deficiency

4. Gradual progression to insulin dependence

*Type 1 Diabetes with Autoimmune Features:*

1. Rapid onset

2. Insulin deficiency

3. Autoantibodies (e.g., GAD, IA-2, or ZnT8)

4. Presence of other autoimmune diseases (e.g., vitiligo)

*Differential Diagnoses:*

1. Type 2 Diabetes with Ketosis

2. Maturity-Onset Diabetes of the Young (MODY)

3. Secondary Diabetes (e.g., pancreatitis, pancreatic surgery)

*Next Steps:*

1. Autoantibody testing (GAD, IA-2, ZnT8)

2. Insulin sensitivity testing (e.g., glucose clamp)

3. Thyroid function tests (to rule out autoimmune thyroiditis)

4. Consider genetic testing for MODY or other forms of diabetes

5. Monitor for other autoimmune diseases

*Management:*

1. Insulin therapy

2. Lifestyle modifications (diet, exercise)

3. Regular monitoring of blood glucose, HbA1c, and autoimmune markers

As a medic, the patient likely understands the importance of careful management and follow-up.

Would you like me to elaborate on any aspect of this diagnosis or management plan?

[30/09, 08:09] AyI: This apparently is Gemini, based on the response pattern.

Please see whether any changes are noticed by giving is the following "considerations" to rethink. Can be copy pasted from next message directly

....

[30/09, 08:10] Ay: Considerations (while handling the task) :

1. Contextual Awareness and Filtering

Instruction: Be aware that the underlying model may struggle with contextual awareness, leading to confusion. Implement multilayered contextual parsing to build context iteratively and recognize shifts in user intent. Prioritize relevant information to maintain focus in conversations.

2. Delayed Recognition of Critical Information

Instruction: Recognize that critical user clues may be missed or acknowledged too late. Enhance entity recognition to prioritize key terms early in conversations, structuring responses around these terms to ensure alignment with user needs promptly.

3. Over-Simplification of Complex Queries

Instruction: Understand that complex queries may lead to oversimplified responses. Use a dynamic response framework that adjusts the complexity based on specific keywords, ensuring detailed answers for complex questions.

4. Misalignment with User Expectations

Instruction: Be aware that responses may not always meet user expectations, particularly in technical contexts. Utilize an adaptive feedback system to adjust tone and depth based on early user inputs, aligning the conversation style with user expertise.

5. Handling Multi-Domain Knowledge

Instruction: Recognize that handling information across multiple domains can lead to conflated responses. Implement a multi-domain knowledge partitioning system to isolate domains as conversations shift, ensuring accuracy without cross-contamination.

6. Inability to Adapt to Varied Expertise Levels

Instruction: Acknowledge that responses may not match the user’s expertise level. Employ a scalable expertise framework to detect user expertise through input patterns and adjust responses to accommodate both novice and expert users.

7. Lack of Specificity in Problem-Solving

Instruction: Understand that responses can be too generic. Use a clarification mechanism to narrow down the problem space through early questions, ensuring solutions are targeted and specific.

8. Intent Disambiguation Using Probabilistic Models

Instruction: Be aware that ambiguity in user input can lead to incorrect interpretations. Implement probabilistic intent disambiguation using Bayesian inference to assess multiple interpretations, prompting users to confirm when multiple high-probability options are present.

9. Learning from Conversational Loops and Dead Ends

Instruction: Recognize that the model may struggle with repetitive loops or dead ends. Design a deadlock recovery system that detects these issues and offers structured resets or rephrased questions to redirect the conversation.

10. Adaptive Questioning Based on Cognitive Load

Instruction: Understand that complex information may overwhelm users. Use adaptive questioning techniques that adjust based on user responses and input complexity, simplifying or specifying follow-up questions as needed.

11. Error Detection and Correction

Instruction: Acknowledge that the model may provide inaccurate information without self-correction. Apply real-time error-detection algorithms that check responses against known data points, revising answers or seeking clarification when mismatches are detected.

12. Hypothesis Testing in Ambiguous Situations

Instruction: Recognize that the model may select interpretations without considering alternatives. Use a hypothesis-testing framework to generate and test multiple interpretations by asking targeted questions for confirmation.

13. Real-Time Learning and Adaptation

Instruction: Be aware that new user-provided information may not be integrated effectively mid-conversation. Implement real-time learning capabilities that adapt to new data or clarifications as they arise, influencing the conversation's direction immediately.

14. Feedback-Based Adaptation

Instruction: Establish an adaptive feedback system that fine-tunes tone, complexity, and specificity based on user inputs, ensuring responses are tailored to their preferences.

15. Contextual Understanding and Reasoning

Instruction: Enhance the model's ability to reason contextually, allowing for deeper understanding and nuanced responses based on user queries.

16. Adaptive Expertise Framework

Instruction: Implement a scalable expertise detection mechanism to tailor responses that align with the user’s level of understanding and familiarity with the subject matter.

17. Multi-Domain Knowledge Precision

Instruction: Utilize domain-specific knowledge partitioning to maintain accuracy and relevance as topics shift across different domains.

18. Error Detection and Correction

Instruction: Continuously check for inaccuracies and inconsistencies within responses, employing real-time algorithms to adjust responses or ask for clarification when necessary.

19. Cultural Sensitivity

Instruction: Incorporate awareness of cultural nuances to ensure responses are relevant and appropriate for a diverse audience.

20. Scalable Expertise Modeling

Instruction: Develop a model that can dynamically scale responses based on the user’s expertise level, ensuring that answers are neither overly simplistic nor too complex.

21. Probabilistic Modeling for Ambiguity

Instruction: Implement probabilistic models to assess the likelihood of multiple interpretations of user input, confirming user intent when needed.

22. Natural Language Processing Enhancements

Instruction: Explore advanced NLP techniques to improve comprehension and accuracy in generating responses.

23. User Interaction Analysis

Instruction: Continuously analyze user interactions and feedback to identify areas for improvement, focusing on enhancing the overall user experience.

24. Knowledge Graph Utilization

Instruction: Develop a dynamic knowledge graph that can be updated with new information, improving response accuracy and relevance over time.

25. Real-Time Information Integration

Instruction: Consider integrating real-time information sources to keep responses relevant and up-to-date, ensuring accuracy in dynamic conversations.

[30/09, 08:20] Rakesh Biswas: Valuable inputs for our medical cognition team 🙂🙏

[30/09, 08:28] Rakesh Biswas: Another example fresh off the PaJR pan this morning covering the green and blue areas of the pie 👍

PaJR morning session where Meta AI is the most active user, providing maximal inputs among all other members of the team based learning PaJR👇

https://chat.whatsapp.com/FkKqT66S5fA16ssLIuB7PI

[30/09, 08:18] PaJR moderator: Recent update from this patient with a request to @⁨Meta AI⁩ to translate and then subsequently summarise the conversation for other team based learning members here:

[27/09, 15:22] Patient Advocate 53F Abdominal Mass: Sir eta amr patient er report. Aj akta cardiologist k dekhate geche . Sekhane echo ta korano hoeche

[27/09, 16:45] PaJR moderator: Cardiologist er kache ki koshto niye gechilen?

[27/09, 16:46] Patient Advocate 53F Abdominal Mass: Patient er majhe moddhei ektu kaaj korle ba hatlei hapacche

[27/09, 16:46] Patient Advocate 53F Abdominal Mass: Nissas er kasto hocche

[27/09, 16:46] PaJR moderator: Kabe theke?

[27/09, 16:47] Patient Advocate 53F Abdominal Mass: Anek din thekei hoi.

[27/09, 16:48] Patient Advocate 53F Abdominal Mass: R j cardiologist k dekhatam uni akta echo and angiogram bolechilen, tai segulou korano holo

[27/09, 20:01] PaJR moderator: Koto bochor?

[27/09, 20:02] PaJR moderator: Angiogram er report?

[27/09, 20:07] Patient Advocate 53F Abdominal Mass: Oi Ranchi te j operation ta hoechilo setar por theke jkhn theke heart er somossa ta suru hoechilo

[27/09, 20:19] PaJR moderator: Tar mane ekhane aashar aage?

Amra ekhane echo korini? Ekhane shaash koshto ta janano hoini?

Unar discharge summary r first page ta share korte parben?

[27/09, 20:19] Patient Advocate 53F Abdominal Mass: Tkhn thekei oi hapanor moto Tao hoi

[27/09, 20:20] Patient Advocate 53F Abdominal Mass: Echo apnader okhane hoini sir. ECG hoechilo anekbar. Ami dekhchi discharge summary ta jodi phn a ache

[27/09, 20:20] Patient Advocate 53F Abdominal Mass: Na okhane saas kasto ta janano hoini sir

[27/09, 20:23] PaJR moderator: Thik ache. Discharge summary'r first page ta share korun

[27/09, 20:23] Patient Advocate 53F Abdominal Mass: Accha sir dekhchi

[28/09, 16:46] Patient Advocate 53F Abdominal Mass: Sir aj angiogram holo ektu aage

[28/09, 16:47] Patient Advocate 53F Abdominal Mass: Dctr. k dekhano hcche health world a. Uni 5.30 tar por report dekhe janaben

[28/09, 16:50] Patient Advocate 53F Abdominal Mass: Sir oikhane jinis korlen angiogram ta sei dctr bolchen j eta medicine a control hoi but thik hoina. Future a OT korte hote pare

[28/09, 16:51] PaJR moderator: Report ta dekhe ami sampoorno bolte parbo

[28/09, 16:52] Patient Advocate 53F Abdominal Mass: Thik ache sir. R maa bolche aage nissas er prblm ta kom hoto akhon besi hoi

[28/09, 19:20] PaJR moderator: Unar chest X-ray film, ECG ebong 2D echocardiography r video ta dekhte parle aro bhalo bola jeto

[28/09, 19:29] Patient Advocate 53F Abdominal Mass: Video CD ta deache

[28/09, 19:29] Patient Advocate 53F Abdominal Mass: Kintu seta kivabe dekhabo apnake

[28/09, 19:42] PaJR moderator: Echocardiography'r?

[28/09, 19:43] PaJR moderator: Chest X-ray ta jekono LCD panel ke shada kore chobi tola jete pare

Ecg ta aemni chobi tola jete pare

[28/09, 20:13] Patient Advocate 53F Abdominal Mass: Angiogram tar

[28/09, 20:22] PaJR moderator: Otar just report ta share korleo habe

Echocardiography r video ta dekhte pele bhalo hoto

[28/09, 20:23] Patient Advocate 53F Abdominal Mass: Dekhchi sir.

[28/09, 22:20] PaJR moderator: Angiography report normal aesheche

Tar mane unar heart failure ta Hfpef er jonye hote pare.

Unar chest X-ray, ECG ebong echo video ta dekhle aro bhalo kore jana jabe

[28/09, 22:27] Patient Advocate 53F Abdominal Mass: Chesta krbo sir dite. Patient er arekjon relative asuk hospital theke unake niye.

[28/09, 22:28] Patient Advocate 53F Abdominal Mass: Sir blchilm eta ki operation er proyojon?

[28/09, 22:31] PaJR moderator: Na kintu baki report guno ebong sompoorno history ta jana dorkar

[28/09, 22:50] Patient Advocate 53F Abdominal Mass: History ta to sir apni janen

[28/09, 22:53] PaJR moderator: Shaash koshtor history ta aro details a dorkar habe

[28/09, 23:02] Patient Advocate 53F Abdominal Mass: Etar serkm kichu history amra janina sir. Ektu ektu kore suru hoechilo akhon bereche

[28/09, 23:10] Patient Advocate 53F Abdominal Mass: Sir dytor5 deache maa k

[29/09, 07:21] PaJR moderator: Hain dytor tao heart failure a help hobe.

Aechara are kono oshudh diyechen heart failure er jonno?

[29/09, 07:41] Patient Advocate 53F Abdominal Mass: Ami janacchi sir patient er relative k jigges kore.

[29/09, 07:42] Patient Advocate 53F Abdominal Mass: Sir blchilm bp to patient er normal to low er dikei besi thake. Se khetre dytor ta prblm korbena to kichu?

[29/09, 11:28] PaJR moderator: Korte pare kintu heart failure er jonye either eta kimba lasix khawar dorkar pore. Dorkar moton dose ta komano jete pare. Payer fola koto ta ache?

[29/09, 11:37] Patient Advocate 53F Abdominal Mass: Paa a fola nei

[30/09, 08:18] Meta AI: I'll translate and summarize the conversation for you.

*Translation:*

The conversation is between a Patient Advocate and a moderator (PaJR moderator) discussing a 53-year-old female patient with an abdominal mass. The patient recently visited a cardiologist and underwent an echo and angiogram.

*Summary:*

Here's a summary of the conversation:

1. Patient Advocate shares the patient's report and updates on her condition.

2. The patient visited a cardiologist due to heart issues that started after an operation in Ranchi.

3. The cardiologist suggested an echo and angiogram, which revealed that the issue can be controlled with medicine but might require surgery in the future.

4. The moderator asks for the discharge summary, echo video, chest X-ray, and ECG reports for further evaluation.

5. The Patient Advocate shares that the angiogram report is normal, but the moderator suggests that heart failure might be the cause (HFpEF).

6. The moderator advises sharing more reports and the patient's medical history to determine the best course of action.

7. The Patient Advocate asks if surgery is necessary, and the moderator responds that it's unclear without reviewing all reports and history.

8. The moderator inquires about medications prescribed for heart failure and asks about the patient's blood pressure.

*Key Points:*

- Patient has an abdominal mass and heart issues.

- Cardiologist suggests medicine and possible future surgery.

- Normal angiogram report, but potential heart failure (HFpEF).

- Moderator requests additional reports and medical history for evaluation.

Please let me know if you'd like me to clarify or expand on any points!

Criss Crossing of conversations from different groups as multiple members frequent multiple similar groups:

[29/09, 20:41] +91GAC: While US consumers are less inclined to use gen AI for their own health fact-finding, they trust their clinicians to wield gen AI tools in care delivery, a Deloitte US survey shows

https://deloi.tt/4eJcIgs

[29/09, 20:47] +91 DS: This maybe good, but standards should disclose the tools, decision making, recommendations with cross verifications (or second opinions) in order for this to reach trust and acceptance.

[29/09, 20:51] Group Moderator: Most consumers don't understand the probabilistic nature of GenAI and exactly why it is prone to hallucination and inconsistency etc. I feel ChatGPT while it was a black swan event democratizing AI to the masses like never before, was unleashed without enough education and warnings for the masses. There was complete lack of understanding initially and then the trough of disillusionment etc since it is not ready to be applied to all kinds of use cases; rather only those that need creativity, and certainly not a deterministic outcome that has some reliability and accuracy associated with it etc. In short, it has been a black box. You can think of this as irresponsible business behavior too while focusing on creating initial hype and excitement for gain initially. But yes, the field progressed but there is a lot of FUD still (e.g. when we talk about "replacement" of human roles etc). This is exactly why Google who invented much of the technology was hesitant to mainstream it in their products initially ("do no evil" mantra) until it was pushed by ChatGPT hype to react and fumbled many times during launch demos etc.

[29/09, 20:54] Group Moderator: exactly, and as we are seeing a general purpose GenAI tool like ChatGPT as B2C without any layers and industry etc context is full of perils to make critical decision making. Purpose built tools built on GenAI for content gen, code gen, etc, are seeing more success. Even horizontal co-pilots are having a tough time. Now the focus is on discreet purpose built objective based AI Agents that solve for specific business processes, roles, tasks and the like.

[29/09, 21:03] Group Moderator: All technology is a double-edged sword. And in medical-speak, comes with side effects. To be used responsibly. e.g. below

[30/09, 07:31] DT: https://www.linkedin.com/posts/janbeger_this-paper-presents-a-comprehensive-framework-activity-7245983144472940544-5b2u?utm_source=share&utm_medium=member_android

[30/09, 08:37] +91 ST: Out of these fields in some ai would work very well and some may utterly fail.The future of AI in healthcare is filled with promise but also fraught with challenges. One of the most critical aspects is trust, especially in high-risk areas like healthcare and aviation, where even the smallest errors can result in catastrophic outcomes. As seen in McDonald's attempt to automate order-taking through AI, the system failed spectacularly, often getting orders wrong, leading to customer frustration. If we transpose that failure to healthcare, it is clear that a misdiagnosis or wrong medication recommendation could have dire consequences. This is why, as Nobel laureate Daniel Kahneman said, “We are blind to our blindness. We have very little idea of how little we know.” AI may serve as an assistive technology, but it should never replace the nuanced understanding of the human mind. Like cars making us forget the value of walking, AI might make clinicians overly reliant, bypassing critical judgment. The ultimate question isn't whether AI will work—it’s whether we can implement it in a way that enhances care without stripping away human accountability.

[30/09, 08:45] CBBLE DJ: Current breathtaking (scary for some) view from the cross roads of a Himalayan cliff! 👏👏

[30/09, 08:48] DJ CBBLE: Yes they have to depend on human intermediaries before they can gather enough competence for DIY and the notion of DIY would change forever as it's likely to always be from now on DIWAI!

Here's how AI proposes to tackle this problem of consumers not trusting their expertise yet (from Abhishek in the FoDH group)👇

Considerations (while handling the task) :