Latest / Logosofia no Dia a Dia / LLM Councils for Post-Gallbladder Care

Transcript

LLM Councils for Post-Gallbladder Care · Logosofia no Dia a Dia

• [Prompt] custom: I've shared some of my medical challenges. I'm sure more times than anybody was ever hoping to hear | Hosts: corn, herman [Response] Corn: Daniel sent us this one — and it's personal. He had his gallbladder out seven years ago, and ever since he's been dealing with bloating, gastritis, weight gain, the whole cascade. He's seen doctors, tried to piece it together, and the problem is that no single specialist owns this condition. So his question is whether something called an LLM council — running the same problem through multiple AI models, each prompted as a different specialist, then synthesizing their outputs — could actually help someone like him build a coherent treatment plan when the medical system keeps handing him fragments. Herman: This is not a small club he's in. Post-cholecystectomy syndrome hits somewhere between ten and forty percent of patients after gallbladder removal. You're talking about a massive, underserved population. Meanwhile, the cost of running multiple AI models has dropped to the point where a patient can spin up a council of four or five models for pocket change. The timing on this question is oddly perfect. Corn: It's also a systems problem that maps almost too neatly onto a systems solution. You've got bile acids that were supposed to be stored and released in pulses now just dripping continuously into the intestine. That causes diarrhea in up to twenty percent of patients. The same bile can reflux upward into the stomach and cause chemical gastritis — which antacids can't fix because the irritant isn't acid, it's bile. You lose the bile's role in fat digestion, so you bloat. You lose its role in metabolic signaling, so you gain weight on the same diet. And then there's the psychological layer Daniel mentioned — going from someone who loves food to someone who thinks about eating in purely chemical terms, what can I tolerate without symptoms. Herman: What strikes me about his prompt is how clearly he's thought through the architecture. He's not asking whether AI can replace his gastroenterologist. He's proposing two specific council designs. One is role-based: prompt one model as a gastroenterologist specializing in bile acid malabsorption, another as a nutritionist, another as a psychologist who handles chronic illness. The second is geography-slanted: take the same generalist prompt but run it through models trained on different regional data — DeepSeek with Chinese medical literature, Claude with Western, and so on — to surface treatment approaches that might not be common in the patient's own country. Corn: The synthesis layer is where it gets interesting. You don't just collect four opinions and call it a day. You feed all the outputs into a separate model — or the same model with a chairperson prompt — and ask it to flag contradictions, reconcile conflicting advice, and produce something a patient can actually hand to their doctor. Herman: And Daniel's point about running this at recurring intervals is smart in a way I don't think most people would anticipate. Medical knowledge moves. A council you run today might surface different evidence than the same council run three years from now. You're not building a static second opinion — you're building a longitudinal tool that tracks your symptoms and the literature in parallel. Corn: There's a tension here though, and I think Daniel feels it. He's not in the "AI will replace doctors" camp. He's in this middle space where he sees the additive value — no single human can keep up with the literature across gastroenterology, nutrition, psychology, and pharmacology simultaneously — but he also knows the risks. They miss rare contraindications. They can amplify the biases baked into their training data, like the well-documented tendency to underdiagnose pain in women. Herman: We should be upfront: as of mid two thousand twenty-six, there is no peer-reviewed study testing LLM councils for chronic disease management. This is uncharted territory. The closest tools we have — Ada, Buoy — use single-model decision trees. They're validated but rigid. An LLM council is the opposite: flexible, generative, and completely unvalidated for this use case. Corn: The episode is really about two questions. One, would this actually work for a condition as tangled as post-cholecystectomy syndrome? And two, if you're a patient listening right now, how would you build one? Herman: Let's start with the condition itself — because the reason it needs a council in the first place is that losing a gallbladder is not just about losing a bile storage tank. It's a disruption that ripples through multiple systems, and the medical system has no single owner for the fallout. Herman: You lose your gallbladder, and suddenly bile has no place to park. The liver keeps producing it — about a liter a day — but instead of being stored and released in concentrated pulses when you eat fat, it just drips continuously into the duodenum. That constant drip is the root of most of what goes wrong. Corn: The duodenum is not designed for a constant bile bath. It's meant to see bile in coordinated bursts. So you get irritation, you get diarrhea — bile acid diarrhea specifically, which is different from the kind you'd get from a stomach bug. It's caustic. And if that bile migrates upward into the stomach, which it does in a subset of patients, you get chemical gastritis. Not acid-driven, so proton pump inhibitors do almost nothing. Herman: And that's the first crack patients fall through. They go to a gastroenterologist, the gastroenterologist sees gastritis on the scope, prescribes a PPI, and it doesn't work. Because the problem isn't acid — it's bile. You need a bile acid sequestrant like cholestyramine, or in some cases ursodeoxycholic acid to change the composition of the bile itself. But if your GI doc isn't specifically thinking about post-cholecystectomy physiology, they may never go there. Corn: Meanwhile the nutritionist is working a completely separate problem: how do you eat enough calories when fat makes you bloat and fiber makes you bloat and you're losing weight you can't afford to lose? And the psychologist is dealing with the fact that you've now spent two years being told your scopes look normal, your bloodwork is fine, maybe it's stress. Herman: That survey from twenty twenty-three — only twelve percent of academic medical centers have a dedicated post-cholecystectomy clinic. The other eighty-eight percent of patients are bouncing between specialties that don't talk to each other. By the time someone gets a coherent plan, they've typically seen three to five different specialists. Corn: Which is exactly why the council idea has teeth. You're not asking one model to be a superdoctor who knows everything. You're asking four models to each be one specialist who knows their domain deeply, and then you're asking a fifth model to sit them all at the same table and hammer out a plan. Herman: That's the key difference from just asking Claude or GPT for medical advice. A single model gives you a single synthesis — it might be good, it might be shallow, you have no way to know. A council forces disagreement into the open. If the gastroenterologist model says low-fat diet and the nutritionist model says you need more healthy fats to maintain weight, the synthesis layer has to reconcile that. You see the tension instead of having it smoothed over. Corn: The geography-slanted version adds another dimension. Take ursodeoxycholic acid — it's prescribed far more routinely for bile reflux in Europe than in the US. If you only run your council through models trained predominantly on American medical literature, you might never hear about it. Run a European-slanted model in parallel, and suddenly it surfaces. Herman: Then you can take that finding to your actual doctor and say, here's a treatment that's standard in Germany, what do you think? That's a fundamentally different conversation than showing up and saying, the AI told me to try this thing I can't pronounce. Herman: Let's walk through a concrete example, because the architecture makes more sense when you see it in action. Take a patient — call it Daniel's exact situation — bloating after meals, gastritis that doesn't respond to PPIs, weight creeping up despite eating less. You spin up three role-based models. Corn: You're prompting each one with a specific clinical identity, not just "give me medical advice." You'd say something like: "You are a gastroenterologist specializing in bile acid malabsorption and post-cholecystectomy physiology. A patient presents with these symptoms, these lab results, this surgical history. Recommend diagnostic tests and treatment options with citations. Herman: The nutritionist model gets the same case but a different lens: "You are a clinical nutritionist specializing in post-surgical malabsorption. Design a meal plan that meets caloric needs while minimizing fat-induced bloating, accounting for possible FODMAP sensitivities." And the psychologist model gets: "You are a health psychologist specializing in chronic gastrointestinal conditions. Address the anxiety around eating, the identity shift from food-lover to food-avoider, and recommend evidence-based interventions. Corn: Each model stays in its lane. The gastroenterologist isn't trying to design a meal plan. The nutritionist isn't speculating about bile acid sequestrants. You're deliberately constraining each one to go deep rather than wide. Herman: Here's where the synthesis layer earns its keep. The gastroenterologist recommends cholestyramine — a bile acid sequestrant that binds bile in the intestine so it can't irritate anything. The nutritionist recommends a low-fat, low-FODMAP diet with small frequent meals. The psychologist recommends cognitive behavioral therapy focused on health anxiety and food-related avoidance behaviors. Corn: Three perfectly reasonable recommendations that, if you just handed them to a patient as a list, would miss a critical interaction. Cholestyramine doesn't just bind bile — it also binds fat-soluble vitamins. A, D, E, K. So if the patient follows the nutritionist's meal plan without supplementation, they could end up deficient. Herman: That's exactly the kind of catch a good synthesis model would flag. You feed all three outputs into a fourth model with a chairperson prompt: "You are synthesizing recommendations from three specialists. Identify contradictions, flag drug-nutrient interactions, note where plans reinforce or undermine each other, and produce a unified summary a patient can discuss with their primary physician. Corn: The output isn't "here's what to do." It's "here's what three perspectives surfaced, here's where they agree, here's where they conflict, and here are the questions to ask your doctor." That last part is crucial — the council doesn't prescribe, it prepares. Herman: The geography-slanted version works differently but the synthesis challenge is similar. You might prompt a US-trained gastroenterologist model and a Japan-trained gastroenterologist model with the same case. The US model might reach for a PPI plus a bile acid sequestrant. The Japanese model, where H. pylori eradication protocols are more aggressive and where post-gastrectomy syndromes have driven a lot of research into motility disorders, might recommend a prokinetic agent or a different class of mucosal protectant. Corn: The synthesis layer surfaces the disagreement explicitly: "Model A recommends a PPI. Model B notes that in bile reflux gastritis, reducing gastric acidity may actually worsen symptoms because acid helps neutralize bile's corrosive effects. Discuss this tension with your gastroenterologist." You're not getting a verdict — you're getting a map of the controversy. Herman: Which is arguably more valuable. A single model might confidently recommend the wrong thing. A council shows you where the field itself is uncertain, and that's information a patient can actually use. Herman: This is where we need to talk about the risks, because the elegance of the architecture can make you forget how many things can go wrong. A model hallucinates a drug interaction that doesn't exist — and the synthesis layer treats it as a real constraint to reconcile. Or the council misses a rare contraindication because none of the training data surfaced it. You could walk into your doctor's office with a beautifully formatted report that is confidently wrong in ways you can't detect. Corn: The bias problem is real in a way that's easy to hand-wave but hard to catch in practice. If you're a woman with post-cholecystectomy pain, the training data across multiple models may reflect the same blind spot — the well-documented tendency to underdiagnose or psychologize women's pain. A council of four models that all share the same bias isn't a council, it's an echo chamber with better branding. Herman: The other thing that worries me: no peer-reviewed study has tested this for chronic disease management. We have case reports, we have people experimenting on themselves on Reddit, but we don't have a controlled trial that says running an LLM council improves outcomes compared to standard care. That doesn't mean it doesn't work — it means we don't know. Corn: Which is why the framing matters so much. If you treat the council as a diagnostic tool, you're misusing it. If you treat it as a hypothesis generator — something that surfaces options and tensions you can then take to a human expert — that's a defensible use case. The council doesn't give you answers, it gives you better questions. Herman: The practical question is: how would someone actually set this up today? The good news is the tools exist. You can access Claude, GPT-4o, and DeepSeek all through free tiers or low-cost subscriptions. You don't need a server farm. You need a symptom log, your lab results, a list of current medications, and about an hour to run the prompts. Corn: The missing piece is the synthesis workflow. Nobody has built a turnkey tool that routes your case to four models, collects the outputs, and feeds them to a fifth. You're doing this manually — copying and pasting between chat windows. It's clunky, but it works. And honestly, the manual step might be a feature, not a bug. It forces you to read each output carefully instead of trusting a black box. Herman: I'd add one practical note: run the council monthly. Feed in updated symptom logs, new lab results, any medication changes. The synthesis model can track deltas — your bloating score dropped after starting cholestyramine, your energy improved after the nutritionist's meal plan adjustments. That longitudinal tracking is something even a good specialist visit often misses, because the doctor sees a snapshot, not a trend line. Corn: Compare this to Ada or Buoy — they give you a single-path decision tree. You answer questions, it narrows possibilities, you get an output. Useful for triage, useless for a condition as tangled as post-cholecystectomy syndrome where the real challenge isn't identifying the problem but coordinating responses across domains. Herman: Those tools are validated. An LLM council is not. That's the tradeoff. Flexibility versus proven safety. You're essentially running an unregulated clinical trial on yourself every time you spin one up. Corn: The rural Texas case makes this concrete in a way I think Daniel would appreciate. Patient with post-cholecystectomy syndrome, three hours from the nearest gastroenterologist. The US-slanted model recommends a low-fat diet and a PPI. The European-slanted model recommends ursodeoxycholic acid and a bile acid sequestrant. The synthesis model flags that PPIs may actually worsen bile reflux by reducing gastric acidity, and suggests discussing UDCA with their doctor. Herman: That's the moment the council earns its keep. Without the European-slanted model, the patient never hears about UDCA. Without the synthesis layer, they don't learn that the PPI their local doctor prescribed might be making things worse. They walk into their next appointment with a specific question and a specific paper to reference, not just a vague sense that something's wrong. Corn: The geographic slant is underrated as a strategy. Medical practice is path-dependent — treatments become standard in one country because of historical research priorities, regulatory decisions, or pharmaceutical market dynamics, not because of a global consensus on what works best. Running models trained on different regional literatures is a cheap way to surface those differences. Herman: If you're listening and thinking, I want to try this — here's where I'd start. Build a symptom log first. Not a diary, not "felt bad today." A structured log: what you ate, when you ate it, what symptoms appeared and when, severity on a one-to-five scale, what medications you took. Do this for two weeks before you even touch an AI. The council is only as good as the data you feed it. Corn: Include your current treatment list — every medication, every supplement, dosages, timing. If you're on a PPI and it's not working, the council needs to know that. If you tried cholestyramine and it gave you constipation, that's data too. The models can't ask follow-up questions the way a doctor can, so you have to front-load the context. Herman: Then set up your council. Free tiers of Claude, GPT-4o, and DeepSeek are plenty to start. Prompt three models as three different specialists — gastroenterologist, nutritionist, health psychologist. Give each one the same symptom log and treatment history, but a different clinical lens. Run the outputs through a fourth model with a synthesis prompt that explicitly asks it to flag disagreements, not smooth them over. Corn: The disagreements are the product. If all four models agree on everything, you've learned nothing except that your prompts were probably too vague. What you want is the nutritionist saying more fiber and the gastroenterologist saying less fiber for now, and the synthesis model explaining why both might be right at different stages. Herman: Bring the output to your doctor. Not as a demand — as a conversation starter. "I ran my case through a few AI models prompted as different specialists. Here's what surfaced. What do you make of this?" Some doctors will bristle. Others will be curious. Either way, you're not asking them to rubber-stamp an AI plan — you're asking them to help you evaluate hypotheses you couldn't have generated on your own. Corn: Run it monthly. Same prompts, updated symptom log. Track which recommendations your doctor agreed with, which ones helped, which ones didn't. Over three or four cycles, you'll have something no single specialist visit can give you: a longitudinal map of what moves your symptoms and what doesn't, cross-referenced against the evolving literature. Herman: The tool is for collaboration, not replacement. And honestly, if you try this, document it. Share what you learn. Because right now, the only people building LLM councils for chronic illness are patients experimenting on themselves in isolation. The more of those experiments get shared, the faster we figure out what works. Herman: The question that lingers for me is whether this stays a DIY thing for the technically inclined, or whether it eventually gets folded into standard care. I could see a world where chronic disease clinics run councils as a routine part of intake — here's your symptom log, we'll spin up four perspectives and synthesize them before your appointment. But I could also see it staying in the Reddit threads and self-experimentation forums, because the liability question is a nightmare. Who's responsible when the council misses something? Corn: That's probably the hinge. If a hospital deploys an LLM council and it hallucinates a contraindication that harms a patient, there's a lawsuit with a clear target. If a patient runs it at home and brings the output to their doctor, the liability sits nowhere — which is both the appeal and the risk. You get access to something powerful, but you're also the quality control department. Herman: The multimodal angle Daniel hinted at changes the equation though. Right now we're feeding these councils text — symptom logs, lab numbers. But the models coming down the pipeline can analyze images. You could photograph every meal, every stool sample, and feed that into the council alongside your symptom scores. The synthesis layer could correlate what you ate on Tuesday with how you felt on Wednesday with what showed up in the lab on Thursday. That's a closed loop that no human clinician has the bandwidth to maintain. Corn: It's the "quantified self" movement finally getting a brain. All that data people have been collecting with no idea what to do with it — suddenly there's something that can actually read the patterns. Herman: That's where I land on Daniel's question. Will this work for post-cholecystectomy syndrome? The honest answer is we don't know yet. But the architecture is sound enough that it's worth trying, and the cost of trying is basically zero. If you document what you learn and share it, you're not just helping yourself — you're building the evidence base that doesn't exist yet. Corn: If you try this approach — whether for post-cholecystectomy syndrome or any complex chronic condition — email us. We genuinely want to hear how it goes. What worked, what didn't, what surprised you. The address is show at my weird prompts dot com. Herman: Now: Hilbert's daily fun fact. Hilbert: During the 1980s, researchers on the Yamal Peninsula recorded a Nenets shamanic coronation ritual in which the new shaman had to stand inside a circle of reindeer-hide drums and sing a single note until the combined resonance of the drums produced an audible overtone — believed to be the voice of the spirit accepting the appointment. Corn:...I have so many questions. Herman: None of which we have time for. This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop. If you enjoyed this episode, leave us a review wherever you listen — it helps. We'll be back soon.