Watch

Transcript

1. 0:00 Undiagnosed No More looks for patients whose invisible illness, from Long COVID and ME/CFS to lupus and depression, is missing from the health record. Every patient shown is synthetic.
2. 0:07 U.S. government data trains the UNOM AI models, which serve four users: clinicians, patients, policymakers, and developers. Each part of the picture opens its page.
3. 0:17 The patient app opens on Maria's synthetic answers. Each page is one part of a validated questionnaire.
4. 0:24 Her answers to the NIH RECOVER symptom questions and the DePaul questions on crashes after activity are already filled in.
5. 0:31 Next steps on through daily life and the condition self-checks, where she chose the Long COVID check.
6. 0:39 Her summary: 16 points on the RECOVER symptom score, which reaches the research cutoff of 11, a positive DePaul screen, and daily-life grade 3. Each card says what to discuss.
7. 0:48 Patterns in my record runs the same nine models on her synthetic record, on the phone, and says in plain words what it finds. Patients never see a probability.
8. 0:59 Print makes her visit summary. Download saves her answers as a FHIR file that clinic software can read.
9. 1:06 The patient app is the only part of the site in Spanish. Español in the bar switches it, and English switches it back.
10. 1:14 Send to my clinician (demo) opens the clinician view with her answers attached.
11. 1:19 The clinician view opens on Maria's chart. Needs review marks Long COVID High: 15.4 times the average with her answers, 6.1 times from the chart alone.
12. 1:27 From the patient: Review responses opens her answers from the patient app, with her RECOVER score, her screen result, her daily-life grade, and her own words.
13. 1:34 The Long COVID tab shows each finding with an arrow, an odds ratio, and a bar on one fixed scale. The first row is her RECOVER answer, which multiplies the odds by 2.83.
14. 1:43 Hovering a bar shows its effect: an estimate of 15.9% with her answers against 6.2% from the chart alone.
15. 1:49 Hovering estimated 15.9% shows what is estimated: the chance of a first Long COVID code 91 to 455 days after COVID-19.
16. 1:56 The timeline shows every dated visit, diagnosis, test, and prescription, with the COVID-19 index, day 90, and the scoring date.
17. 2:04 Dana: the Lyme disease model marks her chart Elevated, 1.5 times the average, from an April visit for an insect or tick bite and a rash, treated with doxycycline.
18. 2:11 Beside it: knee pain and facial nerve palsy from her chart, and her county's reported incidence, 72.44 cases per 100,000 people a year.
19. 2:18 Sofia: six models flag her chart. Fibromyalgia is High, 24.6 times the average with her answers and 9.6 times from the chart alone.
20. 2:25 Her answers enter through her fibromyalgia self-check, which matches the pattern of the 2016 criteria.
21. 2:33 The Atlas shows one condition at a time. For Long COVID, 1 in 9 adults who have ever had it carry its code in their health record.
22. 2:40 State map (CDC Household Pulse Survey): 5.3% of US adults currently have Long COVID, 2.6% in Illinois to 10.4% in Mississippi.
23. 2:47 Who goes uncoded, by sex, age group, and race and ethnicity: 9.4% of women's Long COVID is coded, against 12.0% of men's.
24. 2:54 Fibromyalgia by state: the share of people who report it who have a code, from 64% in Alabama to 79% in Pennsylvania.
25. 3:01 Who goes uncoded: in every age, sex, and race and ethnicity group, a smaller share is coded than for hypertension and type 2 diabetes.
26. 3:09 What else is in the record: in people who report fibromyalgia, Ehlers-Danlos syndrome or hypermobility was coded 5.0 times as often as in other survey respondents.
27. 3:15 Back to Long COVID for the cost calculator: with the default assumptions, a two-step screen costs $116 in clinician time for each later-coded patient it finds.
28. 3:23 Every assumption can be changed. With 10 minutes of clinician review per patient instead of 5, the two-step screen costs $233 for each patient found.
29. 3:34 The developer documentation starts with how the pieces fit: the model files, one scorer, and the places it runs.
30. 3:41 Try the scorer runs it in the browser on a synthetic test patient. Sofia with the fibromyalgia model scores High, 9.60 times the average, the same as the reference output.
31. 3:49 The CDS Hooks service scores a chart sent from an EHR and answers with a card for each flagged condition. For Maria, one card: Long COVID, High, 6.1 times the average.
32. 3:56 The model registry lists each versioned model file with its model card and expected outputs for 18 test patients.
33. 4:04 The About page shows the data sources, the conditions, the models, the four papers, and the team.
34. 4:16 Undiagnosed No More: the patient app, the clinician view, the Atlas, and the developer documentation, at undiagnosednomore.org.