Create Jira Ticket
This idea cannot be sent to Jira.
DBbun Invisible Illness Burden Simulator
Solo entry by Uri Kartoun, PhD, founder of DBbun LLC, Cambridge, Massachusetts. DBbun won Phase 1 and is participating in Phase 2.
Proposed solution: the DBbun Invisible Illness Burden Simulator turns a published paper on Lyme disease or Long COVID into a live what-if simulator whose every value is labeled observed, assumed or projected, with federal context beside it. The public demo has 25 finished simulators and 13 cross-pollination analyses, each combining a Lyme disease paper and a Long COVID paper.
Solution Overview (MVP Demonstration):
User-Facing Application
Live demo, prototype, or working product
https://dbbun.github.io/dbbun-invisible-illness-burden-simulator/
Access instructions (if needed)
On the demo's Simulation Library, open simulator 3 (MyLymeData). Read Overview, switch to the Care delivery teams view, then open Explore, change one assumption, and see the Export tab. Press play on its audio overview and open the transcript. Open simulator 2 and read Evidence & Sources for the cost inputs and federal check. To see a cross-pollination analysis, go back to the Simulation Library and, under simulator 3, click "Cross-pollination analysis available: with 13" to open the analysis of simulators 3 and 13.
The public site is read-only; uploading a paper and the API run in the full tool on the author's computer.
Results for the American People
DBbun lets a patient advocate see which burdens a study's results do not cover, a care team see which assumption it must not treat as patient fact, and a policy analyst tell one study's figure from a population estimate before budgeting.
It turns a published paper into a simulator that labels each value observed, assumed or projected, shows federal context beside it, and lets anyone change an assumption.
I designed it by listening to people. I held eight interviews (six patient advocates, a Long COVID patient, a government deputy director) and posted 17 questions on the LymeX Healthathon forum; 15 received replies. After listening, I added a medical-disbelief category, a claims-undercounting check and an Excel export for budget analysts. For patients who may struggle with screens, pages fit phones and tablets, and every simulator has a spoken overview and transcript.
Across this non-systematically chosen library of 25 simulators, results are tagged to 1.4 of 16 burden categories on average, none to more than 3. Six categories have no tagged result for either condition, among them three that people described to me: care continuity, parent-to-child transmission and long-term opportunity costs.
Thirteen cross-pollination analyses pair a Lyme and a Long COVID paper and ask what one condition could learn from the other. Six Lyme simulators report a diagnostic-delay result and no Long COVID simulator does: could Long COVID's time to diagnosis be measured?
Data, Tech & Evidence Driven Approach
Four Claude Sonnet calls extract findings with source locators, build an interactive simulator, write named scenarios and forecasts, and generate Python source. Each value is labeled observed, assumed or projected, and ordinary code does the arithmetic.
A local machine-learning text classifier, using term frequency-inverse document frequency (TF-IDF) weighting, decides with no model call whether a document concerns Lyme disease or Long COVID. A retrieval scorer then selects up to ten units from a library of about 1,454, drawn from 30 federal dataset families across 12 agencies. A unit is a processed summary of one federal file, or of one part of a file. Federal data supplies context and checks; the paper supplies the inputs.
Code performs three federal operations: restating dollars with the BLS price index, comparing Lyme case counts with CDC surveillance, and comparing Long COVID prevalence with Household Pulse. Where a cited figure has no year, code declines to compare it rather than guess.
Example: Simulator 2's source reports $2,227 per Lyme episode in 2022 dollars; code restates it as about $2,450 in 2025 dollars, labeled general inflation, not medical-care inflation.
On the five simulators with forecasts, a Monte Carlo simulation of 500 random draws within each assumption's range shades the chart, labeled illustrative, not a confidence interval. Cross-pollination tables are computed in code; the model tags burdens and writes wording, which code checks against fixed lists.
Two guardrails: sex- and gender-specific findings are reported separately, and Long COVID figures sourced from claims or ICD-10 codes are flagged, because such codes capture under half of cases in the population studied by Tian et al. (JAMA Network Open, 2026).
Execution, Efficiency & Delivery
DBbun works as a public, read-only demonstration at https://dbbun.github.io/dbbun-invisible-illness-burden-simulator/. Anyone can open 25 simulators, change assumptions, read 13 saved cross-pollination analyses, hear spoken overviews, and download bundles. The use is evidence exploration: inspect a study's sources and limits, change an assumption, export the results.
Generating a new simulator from a paper, or up to five related documents, runs in the full tool on my computer, not on the public site. It takes about four to nine minutes and $0.40 to $0.96 in model fees per paper (average $0.69, 25 runs), excluding my review time, hosting and maintenance. A documented API, exercised locally, supports another system submitting papers and retrieving results. There is no public self-service upload yet.
After federal reviewers' feedback, I built a ranked list of federal sources, observed, assumed and projected labels, price-year and payer labels, and forecast bands. Five supporting documents are attached: feedback log, quality assurance, interoperability, policy brief, references.
AI critiques reviewed earlier versions of every simulator and cross-pollination analysis, and a separate AI code review (GPT-5 Codex) found ten issues: nine were fixed with regression tests, one is documented. All 25 simulators were then regenerated under the revised rules. 233 automated tests pass.
Scalability, Reuse & Societal Impact:
DBbun is infrastructure, not a one-disease tool: its extraction, evidence labels, arithmetic, API and FHIR export are not written for one disease; the curated federal evidence, prompts and burden categories are built for Lyme disease and Long COVID. To test reuse, I ran two collapsed-lung papers through it, changing only a condition label and scope statement. Each produced a labeled simulator, unvalidated. That shows initial reuse; a new condition needs curated evidence. Societal impact is the evidence gap: six of 16 burden categories have no tagged result for either condition, and the 2022 to 2025 National Health Interview Survey adult codebooks I reviewed have no Lyme question. Deployment: the demo is public and read-only; the full tool runs on one computer with a documented API. Next steps: a hosted version with self-service upload, CMS production access for Blue Button, and asking federal surveys to add items on caregiver hours and schooling disrupted. Standards: each model exports as HL7 FHIR R4 Observation resources with provenance to the source paper; all 26 files pass the official HL7 validator with no errors (warnings only; terminology checks off). Blue Button 2.0 was demonstrated against the CMS sandbox with synthetic beneficiaries, and the CMS Interoperability Framework's consent and transparency principles are echoed. The aim: a reusable evidence layer for any condition and its comorbidities, ideally across PubMed.
As a data scientist at Google who also happened to have gotten sick with ME/CFS, this kind of tooling really excites me! Making research maximally useful and accessible to the people who need it most makes a difference. I wish I would have had something like this when I first was sick.
Reply 1
Wow! This is highly innovative in making publications and the issues addressed or neglected in publications highly accessible.
Users tagged:
Reply 0
Uri, impressive work on DBbun — especially building 25 working simulators, 13 cross-pollination analyses, and a well-tested framework that can extend beyond a single condition.
I’m particularly interested in the interoperability side of DBbun. Looking ahead, do you envision the DBbun API being able to accept structured outputs from other AI or public-health systems as inputs to its burden, care-delivery, or cost simulations?
For example, Lyme AI-ert generates geographic risk signals, temporal trends, weekly alerts, and predicted annual Lyme cases. I could see an interesting connection where these upstream signals feed into DBbun to explore downstream burden, care-access, resource-planning, or cost scenarios.
The weekly alert data and predicted annual Lyme cases are already available through interactive maps in my Lyme AI-ert project, so you’re welcome to explore them anytime here:
https://invisibleillness.crowdicity.com/post/896784
I think connecting early-warning intelligence with downstream burden and planning could be an interesting direction for future collaboration.
Reply 1
Great to see your work, Uri. Coukd younput in a link to whether research trials have measurable results focused on.decreasing illnesses and disability. Most research is useless for real patients.
NIH has a $40B annual budget yet has had NO clinical trials to find amd treat the root causes and genomics. For exam0le, the ABCD, IACC, CHARGE and ReJARGE research trials did not help ch8ldren with autism because data collection was the end goal. 20 years and mbillios of dollars spent atNIH have been downstream, not on root causes by design. Drug compamies/for profits do the trials, for drugs and profit, not causes.
NIH gathered data, wasted decades, spent billions.
Tracj the NIH spending tp see if it focuses on real patients, not mice, on causes, on treaents that already work as in 2025 EVANHEA tial. mandate to reduce illness and disability. They seem to have forgotten.it. Suicides, mental illnesses and dementia are results as no one benefitted who hav the 3Bs borrelia/Lyme, Bartonella, babesia and PANS.
Track if any priority to most disabl8ng, costly, misdiagnosed diseases and if good use of taxayer money too., Those are mandates too for NIH research. Yet for 40 years, no clinical trials on 3Bs, B12/MTHFR, etc root causes of neuropsychiatric disorders, the #1 need. Downstream amyloid got billions, is a correlate triggered by indections like Lyme. Known to cause plaques since 1980s monkey and 2011 human (Miklossy's research). Neuroimmune.org, AlzPi.org, PsychiatryRedefined.org, LymeDisease.org researchers have education tet NIH, CDC used none of it in to educate doctors in their AMA toolbox is outdated dogma.
Reply 1
Users tagged:
Reply 1
Guau- qué genial 😍
Reply 1