Clear Practice Consulting

Home
United States
Plano, TX
Clear Practice Consulting

Website

Jeanne Duke Consulting empowers healthcare professionals to navigate the complexities of generative AI.

ClearPractice helps mid-market healthcare organizations ensure that every critical decision can be clearly explained, justified, and defended under audit, legal, and regulatory review — before someone external starts asking hard questions. By providing expert guidance on everything from foundational prompting techniques to mastering the intricate landscape of governance and compliance, we help you harness the power of AI to innovate and overcome challenges within your practice. Our goal is to equip you with the knowledge and tools needed to create effective, compliant solutions that drive your professional success.

08/06/2026

The Framework: Why Each Decision

Here's what emerged when I sat with these questions:

A defensible dataset starts by defining the scope. Parker's published works, his sermons, essays, and addresses, are the core. That's primary material. That's his actual voice.

Then comes the question of context. Do I include critical responses from his era? Yes — because Parker was a polemicist. He argued. He was responded to. A model trained only on his affirmations would miss half the texture. Including dissent means the model understands not just what he said, but what he was arguing against.

Secondary scholarship, the biographies and scholarly overviews, goes in differently. Not as primary training material, but as reference. The model needs to know who Parker was even if it's primarily conversing in his voice. That requires judgment about which secondary sources are credible and useful.

And then: what do I leave out? Spurious attributions. Unpublished fragments where the source is unclear. Anything I can't defend. Not because Parker is fragile, but because the model's reliability depends on it.

This isn't about size or comprehensiveness. It's about intentionality. Every document that goes in should have a reason it's there. Every absence should have a reason too.

That's where the real architecture lives.

08/05/2026

The Collaboration:

I asked Jim, my husband and a retired history professor, what kind of content he'd want to use in dialogue mode — what would be useful, what would be worth modeling.

I half-joked about pulling together a festschrift from his published work. He laughed and said no — but what about Theodore Parker?

I knew nothing about Parker. Nineteenth-century Transcendentalist, radical reformer, Unitarian minister. Jim has spent years with this era and these thinkers. The suggestion forced me to confront something real: I couldn't just say "yes, let's do that" without thinking through why.

Why Parker, specifically? What makes his work a coherent dataset? What counts as primary material? Do I include critiques of him from his contemporaries, or only his own words? Do I need secondary scholarship — biographies, overviews — to give the model context, or does that dilute it?

Suddenly, the architecture questions were secondary. The real work was deciding what defensible curation looks like.

08/04/2026

The Speed Bump: You can't build an LLM without first asking two questions — and they're harder than you'd think.

I started down the path of creating my first fine-tuned small language model, thinking the hard part was the architecture. Wrong.

The first speed bump: What actually warrants an LLM? Not every knowledge domain needs one. Some questions are better answered by search. Others by a simple retrieval system. I had to get specific about what I was building for — what kind of thinking, what kind of nuance, what kind of conversation. No vague answers.

The second, trickier speed bump: How do you decide what to feed it? Which documents matter. Which voices. Which perspectives. Do you include everything you can find, or are there principled reasons to leave things out? That's not a technical question. That's a curatorial one. And I realized I didn't have a framework for it yet.

That's where I got stuck.

07/29/2026

Healthcare organizations are exceptionally good at preparing for the audit they had last time.

A new benchmark study tested the two most commonly cited "explainable AI" methods, SHAP and LIME, on real clinical prediction tasks. The result: interpreting one year of mortality predictions for a mid-size population would take an estimated 300 hours for SHAP and 64 hours for LIME.

That is not a documentation trail. That is a backlog waiting to become a liability.

If your vendor's pitch includes the word "explainable," ask a more specific question: explainable using what method, tested for faithfulness on a task like ours, and feasible to run at our actual patient volume?

The exposure is not the absence of an explanation. It is an explanation you cannot afford to produce when someone asks for it.

Before it becomes external, it is worth knowing the answer.

07/28/2026

Here is a finding worth sitting with: in a recent benchmark of clinical AI models, adding an attention layer to a predictive model did not cost any accuracy. It gained something else instead: a faster, more faithful path to explaining every prediction the model made.

The method, called Chefer, combined a model's attention with its gradients and outperformed five other interpretability approaches across mortality, length of stay, and complication prediction, while running roughly 15 times faster than the leading alternative.

The takeaway for anyone evaluating clinical AI tools: defensibility is not something you add after the model is built. It is a design decision, made at the architecture stage, that determines whether your organization can answer "why did it decide this" with confidence or with a shrug.

Ask your vendor which stage that decision was made at.

You remain in control of that question if you ask it early enough.

07/22/2026

A question worth asking your AI vendor this quarter: when you tested this tool on synthetic patient data, did that data include the rare conditions, or only the common ones?

A recent benchmark of five synthetic patient-data generation methods found something researchers had suspected but rarely quantified: every method performed well on common diagnoses and degraded sharply on rare, clinically severe ones. The industry habit of simplifying diagnosis codes to make modeling easier hides exactly the cases where a wrong clinical decision carries the most risk.

Aggregate accuracy numbers can look strong while the rare-but-severe cases, the ones most likely to end up in a legal proceeding, are the ones the model has never really seen.

Documented, defensible testing means asking what happened in the long tail, not just what happened on average.

07/21/2026

Good news, with a catch attached.

A 2026 benchmark study tested five methods for generating synthetic patient data used to train and validate clinical AI tools. On the privacy question, every method held up. None showed meaningful risk of exposing real patient information.

The catch: privacy and clinical realism are not the same test, and passing one says nothing about the other. A dataset can be fully privacy-safe and still fail to represent the patients your organization actually treats.

If your organization is relying on synthetic data anywhere in your AI validation process, "it passed privacy review" is only half the documentation you need. The other half is whether it was reproducible, published, and tested under the conditions that matter most, not just the ones that are easiest to model.

Defensible in practice, not just on paper.

11/12/2025

𝗙𝗥𝗘𝗘 𝗪𝗘𝗕𝗜𝗡𝗔𝗥

𝙍𝙚𝙖𝙙 𝙄𝙩 𝙩𝙤 𝙈𝙚: 𝙂𝙡𝙤𝙗𝙖𝙡 𝘿𝙞𝙜𝙞𝙩𝙖𝙡 𝙋𝙖𝙩𝙞𝙚𝙣𝙩 𝙄𝙣𝙛𝙤𝙧𝙢𝙖𝙩𝙞𝙤𝙣

𝗧𝗢𝗠𝗢𝗥𝗥𝗢𝗪: November 13, 2025, at 8 am CST

𝗥𝗲𝗴𝗶𝘀𝘁𝗲𝗿 𝗛𝗲𝗿𝗲: https://link.jeanneduke.com/sp/f5679417a9a

11/11/2025

𝗛𝗼𝗻𝗼𝗿𝗶𝗻𝗴 𝗢𝘂𝗿 𝗩𝗲𝘁𝗲𝗿𝗮𝗻𝘀 𝗧𝗵𝗿𝗼𝘂𝗴𝗵 𝗖𝗼𝗻𝗻𝗲𝗰𝘁𝗲𝗱 𝗖𝗮𝗿𝗲

This Veterans Day, let's recognize not just the sacrifice of those who served, but also commit to supporting their ongoing wellness journey.

As someone who has worked in both healthcare and ministry, I've seen how vital it is for veterans to maintain a consistent connection with their medical teams. The VA's secure messaging system through My HealtheVet has been a blessing for many veterans who no longer need to play phone tag or wait on hold to reach their healthcare providers.

In our digital age, this secure system works like email but protects sensitive health information, allowing veterans to send messages anytime and receive responses within three business days. They can ask health questions, request prescription refills, and manage appointments - all from the comfort of home.

For our veterans in remote areas or those with mobility challenges, VA Video Connect and the ATLAS program offer life-changing access to care through secure video connections. These technologies remove barriers of distance and physical limitations, ensuring no veteran feels forgotten.

This Veterans Day, I invite you to share this information with the veterans in your life. Sometimes the greatest gift we can give is helping someone access the care they've earned through their service.

As we read these words together, let's remember: behind every healthcare technology is a human connection waiting to be made. Our veterans served us with dedication; now these digital bridges help us serve them with the same commitment.

Address

Plano, TX

Website

https://www.clear-practice.com/, https://clear-practice.com/

Alerts

Be the first to know and let us send you an email when Clear Practice Consulting posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.