Demo video

Chapters and transcript

1. 0:00 The problem

   I'm Amir Asiaee from Vanderbilt University Medical Center, and my teammate is Dr. Kaveh Aryan. This is Undiagnosed No More. Invisible illnesses, like Long COVID, chronic fatigue syndrome, fibromyalgia, and Lyme disease, are hard to see and slow to diagnose. And even after a doctor names the illness, and the patient is in care or on treatment for it, the diagnosis often never reaches the record. In the NIH All of Us Research Program, only about 1 in 9 adults who ever had Long COVID carry its code. Of people in care for chronic fatigue, only 24 percent have the code, against 85 percent for hypertension. Without the code, the patient is invisible to referrals, trials, and research.

2. 0:45 How it works

   We built nine open AI models from All of Us and public CDC, Census, and CMS data, with two innovations. First, models trained on individual records inside the All of Us enclave can't be taken out of it, which keeps them out of the clinic. We fit ours outside the enclave, from the aggregate counts it releases, so they can run anywhere. Second, a model trained on diagnosis codes only learns who gets coded. Eight of ours learn from patients' own reports of a doctor's diagnosis, and they rank the uncoded patients higher than code-trained models do. The same models serve four users: patients, clinicians, policymakers, and developers.

3. 1:26 For patients

   Here is Maria, a synthetic patient with symptoms since COVID. These standard questionnaires are usually given in a clinic, if at all, or are hard to find. Here she fills them out on her own, without waiting for a clinician: NIH RECOVER, the DePaul screen for crashes after activity, the PHQ-9, and the GAD-7, in English or Spanish. She doesn't need an account, and her answers stay on her phone.

   Her summary explains what each result means, and what to discuss. It doesn't diagnose; it gives her a voice. She can print it, download it as a standard file that clinic software reads, or send it to her clinician.

4. 2:05 For clinicians

   Her clinician sees Needs review: Long COVID is high, 15.4 times the average with Maria's answers, and 6.1 times from her chart alone.

   Her answers multiply the odds by 2.83, through published likelihood ratios, so the patient's own voice changes the clinical picture. Each dated chart entry shows how much it moves the estimate. Next to them are the published criteria, here the 2024 National Academies definition, and a concrete next step: a post-COVID clinic referral. The same view covers all nine illnesses.

5. 2:40 Inside a real EHR launch

   This isn't a mock-up. Here the same app launches from the public SMART sandbox, the way an EHR opens it, and reads Maria's chart over FHIR, the standard interface certified EHRs provide. A CDS Hooks service can also push these cards into an EHR automatically.

6. 2:57 For policymakers

   For policymakers, the Atlas turns our research into numbers they can act on. For each illness, we link what people tell the All of Us survey, that a doctor diagnosed them, to what their own health records show, and map who goes uncoded by state, sex, age, and race and ethnicity. For Long COVID, we compare the codes in All of Us with how common it is in the CDC and Census Household Pulse Survey: only about 1 in 9 who ever had it carry the code, and women's is coded less often than men's.

   Our research also prices the delay. Before their Long COVID was coded, patients had about seven extra visits compared with matched patients, about $900 per patient at Medicare prices, and our model flagged them a median of 55 days before the code. A questionnaire and a five-minute chart review cost about $116 per patient found, and with the default assumptions, the savings from earlier recognition cover that cost. Every assumption can be changed.

7. 4:00 For developers

   For developers, everything is free and open: nine predictive models as versioned files with model cards, and one scorer that takes a patient's record and returns the estimated chance of each diagnosis, with the record entries behind it. You can try it right here on 18 synthetic test patients.

8. 4:18 The science

   All of this rests on our own research. Learning from patients' reports ranks the uncoded patients higher than a code-trained model on all eight illnesses, from 0.61 against 0.53 for endometriosis to 0.82 against 0.65 for lupus.

   Both innovations, and every number you've seen, are documented in four papers, all public on medRxiv.

9. 4:42 Try it

   Undiagnosed No More is open, standards-based, and deployable today. Try it at undiagnosednomore.org. Thank you.