n=2,840 patients. Multi-site. Naturalistic. Not the same as an RCT — but the next-best thing for understanding how a tool actually performs in community-clinic populations rather than highly-selected academic samples.
Three things matter when you run a multi-site naturalistic registry. The boring lessons that mattered most:
1. Consent flow has to be friction-free for the family or your data is biased toward the most-engaged 10%. We split clinical consent (already obtained at intake) from research consent (separate form, freely revocable, plain English at 8th-grade reading level). Opt-in rate: 64% — much higher than published-trial norms.
2. Motion artifact rejection must be aggressive but tunable. A 14-year-old wearing Mendi at home generates wildly different artifacts than a 35-year-old in a clinic chair. We tuned MAR per-context (FPR 0.8% in clinic, 1.2% at home) and reported both in the sidecar.
3. Sites need a working IRB packet on day one. We published the sample at /downloads — replace the bracketed names, walk through your site's IRB, you'll typically get exempt-or-expedited status in 4–6 weeks for this kind of registry.
Three things that surprised us:
(a) Adolescents stuck with the protocol better than adults. Mendi's gamified UX won on engagement; the clinic check-in won on accountability. Combined, adherence was 4× our prior protocol baselines.
(b) Sleep efficiency from Oura predicted next-session ΔHbO better than the prior session's reward score. We had to add the cross-session pattern detector partway through the registry to catch that signal — clinicians had been intuiting it for years.
(c) Insurance reimbursement for CPT 90901 (biofeedback without psychotherapy) was approved by 11 of 14 payers we tested. The number cited as 'won't pay' (Aetna, Cigna for adolescent ADHD specifically) was payer-specific medical-policy disagreement, not a code-level rejection.