Ask a rehabilitation unit why obstructive sleep apnea (OSA) goes undiagnosed in its medically complex patients and the answer usually concerns hardware: no sleep laboratory on site, no technologist on payroll, an outpatient waitlist measured in months. We spent several weeks pulling that answer apart, and it points at the wrong bottleneck. Recording equipment is now cheap and portable. What fails is the sequence of choices a patient with a three-week-old stroke has to make, in the right order, without dropping one, before a suspicion becomes a treated diagnosis.
So we compared three screening pathways on a criterion that rarely appears in validation studies: how much deciding they require of the patient.
How we ran the comparison
We built a step-level map of three pathways as each would run on a 30-bed inpatient rehabilitation unit with no on-site sleep laboratory — the common case, not the academic-center case. For each pathway we recorded four things:
- Patient decision points. Every moment at which the patient or a proxy must actively consent, show up, or follow through.
- Where those decisions fall relative to discharge.
- Elapsed time from admission to a result a physician can act on.
- Staff minutes per admitted patient, including coordination.
We then applied a uniform 20% drop-off at each patient decision point. That number is an assumption, and it is the weakest joint in the model: per-step attrition data for this population barely exists, so we chose one figure and applied it evenly rather than pretending to pathway-specific precision. Readers who prefer 10% or 30% will find the ranking unchanged and the gaps narrower or wider accordingly.
We collected no primary data. No patients were run through anything, and we had no ground-truth polysomnography against which to check any of it. This is a structural comparison of care pathways, not a diagnostic accuracy trial, and it should not be read as one.
The base rate justifies the exercise. Johnson and Johnson's 2010 meta-analysis of 29 studies and 2,343 post-stroke patients found sleep-disordered breathing in 72% at an apnea-hypopnea index above 5, and 38% above 20. Kaneko and colleagues reported in 2003 that stroke rehabilitation patients with sleep apnea had longer hospitalizations and lower Functional Independence Measure scores at discharge than those without. On a unit where most patients qualify, a screening step that merely sorts is doing less work than a step that commits.
Pathway A: questionnaire at admission, refer outward
STOP-BANG at intake, positive screens referred to an outpatient sleep clinic after discharge. Chung and colleagues validated the instrument in 2008 in surgical patients: at a score of 3 or higher, sensitivity reached roughly 84% for an AHI above 5 and 93% above 15, with specificity in the 35–45% range. On a unit where prevalence runs near 70%, that specificity buys almost nothing — nearly everyone screens positive, which is arguably the correct answer.
The cost is downstream. We counted six decision points: consent to screening, agreement to referral, attendance at an outpatient intake, completion of a home test, return for results, and acceptance of a device through a durable medical equipment supplier. Five of the six occur after discharge, when the patient has resumed managing transport, insurance, and a new medication list, often with the executive dysfunction that brought them to rehab in the first place.
Pathway B: bedside recording, therapy trial on the unit
A multi-channel home sleep apnea test — peripheral arterial tonometry devices such as WatchPAT are the usual instrument — worn on night two or three of the admission, scored the next morning, with an auto-titrating positive airway pressure trial started on the unit for patients who test positive. Yalamanchali and colleagues' 2013 meta-analysis found PAT-derived AHI correlated with polysomnography at roughly r = 0.88, which is adequate for a treat-or-not decision and inadequate for phenotyping.
Two decision points: consent to one night of recording, and consent to try the therapy. Both occur inside the admission, with nursing staff present, a respiratory therapist available to fix a leaking mask at 2 a.m., and a family member likely in the room. Bravata's 2011 work established that auto-titrating therapy is feasible in the immediate post-stroke period without a titration study.
Pathway C: suspicion-led polysomnography
The default at most units. A physician notices witnessed apneas, unexplained daytime somnolence, or refractory hypertension, and refers for a laboratory study. Three decision points, two of them post-discharge, and a ceiling set by the referral step itself: clinician suspicion in a population where somnolence, snoring, and obesity are all confounded by the primary neurological injury identifies a minority of true cases. We modeled that capture at 25%, which is a judgment, not a measurement.
The comparison
| Criterion | A: Screen and refer | B: Bedside test | C: Suspicion-led PSG |
|---|---|---|---|
| Patient decision points | 6 | 2 | 3 |
| Decisions after discharge | 5 of 6 | 0 of 2 | 2 of 3 |
| Time to actionable result | 6–14 weeks | 24–48 hours | 8+ weeks |
| Population reached | All admissions | All admissions | Suspected only |
| Staff minutes per patient | ~10 plus coordination | ~35 | ~15 |
| Modeled share reaching therapy | ~26% | ~64% | ~13% |
Those bottom-row figures are model outputs from a single attrition assumption, not observed rates. Treat them as the shape of the difference rather than its magnitude.
What actually separates them
The sensors are not the variable. A questionnaire and a tonometry device disagree about who has the condition at the margins; they disagree far less than the pathways disagree about who ends up wearing a mask. Pathway A's diagnostic performance is respectable and its yield is poor, because six sequential yes-or-no moments compound against a patient whose recovering brain is being asked to schedule things.
On our model, the pathway that tests everyone at the bedside reaches roughly two and a half times as many patients with treatment as the pathway that screens everyone and refers them out — using a less accurate instrument.
Adherence remains the unsolved part for all three. Sandberg's 2001 rehabilitation trial and the 2016 SAVE study, where mean use settled at 3.3 hours per night, both suggest that starting therapy and sustaining it are separate problems. Ryan's 2011 randomized trial in stroke rehabilitation found motor improvement with treatment, but it is one modest trial, and no pathway comparison can borrow certainty it does not have.
Who this is for, and who it isn't
For a unit with no sleep laboratory, high neurological case-mix, and a length of stay above ten days, Pathway B is the defensible build: the test night fits inside the admission, and the therapy trial has staff behind it. For units under seven days, the recording night consumes too much of the stay, and Pathway A with a warm handoff — a booked appointment, not a referral slip — is the realistic option. For patients with tracheostomy, unstable respiratory failure, or severe agitation, none of this applies; those are laboratory referrals or nothing.
Administrators evaluating this on length-of-stay grounds should note that we tested reach, not outcomes. Nothing here demonstrates shorter stays.
Evidence grade
Moderate for the central claim that relocating decision points inside the admission increases the proportion of patients reaching treatment — the direction is well supported by pathway structure and prevalence data, the magnitude rests on a modeled attrition figure. Weak for any downstream functional-recovery benefit, which awaits trials now in progress.
The myth is that sleep apnea goes undiagnosed in rehabilitation because the diagnostic equipment is out of reach. The more accurate version is that it goes untreated because the pathway asks a recovering brain to make six decisions when two would do.