Open questions for the clinician console

Decisions the clinician console needs before it is built. Answer the ones in your area and skip the rest. The option marked as now is how the current design works, so choosing it means no change. Recommended marks our suggestion, not a default.

Answers save in this browser as you go. When you're done, press Submit answers at the bottom. You can submit again later if you change your mind. “What we found” under each question has the research behind it.

Fix first

Problems in the current console. Neither one depends on the new design.

0. A “yes” on PHQ-9 item 9 reaches no onebug

When a patient answers item 9 (thoughts of self-harm) above 0, the app shows them the 988 card. But it uploads only the total score, so the clinician is never told. Fixing the app is a one-line change. What's missing is a follow-up plan.

Who follows up on a positive item 9?
How fast?
What should the platform store?
What we found
  • The app sets safetyFlag but completeSurvey sends only total, maximum and interpretation. The server already stores whatever extra key the app sends, so the fix is one key plus a server rule that raises the flag.
  • Standard practice for a positive item 9 is a structured risk assessment (e.g. the C-SSRS) before the encounter ends [9].
  • Item 9 alone is sensitive but has a low positive predictive value, 28.6% [10]. That argues for a quick human follow-up, not an automated alarm.

0b. The AI report calls a 1-point mood change “clinically worse”bug

The report treats any check-in move of 4% as a clinical change. On the PHQ-9 that is about 1 point out of 27, and on the GAD-7 less than 1 point. The published thresholds are 5 points (PHQ-9) and 4 points (GAD-7).

How should check-in changes be graded?
What we found
  • Check-ins are stored as score ÷ max × 100, and reporting.py labels a 4-point move on that scale “CLINICALLY BETTER/WORSE”.
  • PHQ-9: 5 points is the long-standing figure [5]; newer work puts it nearer 3.7, depending on baseline severity [6]. GAD-7: 4 points [7].
  • Movement tests were fixed the same way earlier (the real-change rule). Check-ins never were.

For the medical team

Each patient's summary opens with a cover: one word per clinical area (e.g. “Balance: Stable”), then the recommended next steps in three tiers, Top priority, Next step and Keep watching. Flags and test detail sit below it. These questions decide what goes on the cover and in what order.

1. Which tier does each signal belong in?

The current design:

SignalTier
Fall during a testTop priority
Clinical cutoff crossedTop priority
PHQ-9 item 9 above 0Top priority
Confirmed decline (2 tests in a row)Next step
One large move, not yet confirmedNext step: re-test
Early sign of improvementKeep watching
App likely deleted / phone signed outTop priority or Next step (coordinator)
A patient stays past a cutoff for several tests. Their tier:
Should a confirmed decline ever be top priority?
The rest of the table:
What we found
  • Falls are settled. The 2026 AHA/ASA rehabilitation guideline makes a falls-prevention program a Class 1 recommendation for every stroke patient; 40–58% fall in the first year [1][2].
  • Today the code flags a cutoff whenever the latest test is past it, and there is no way to acknowledge one.
  • No published source says how urgent crossing one of the app's cutoffs is. That's your call.
  • Whatever the tiers, every flag becomes a step, so nothing is hidden by putting flags below the steps.

2. How should “needs a clinical evaluation” be shown?

Some steps need an in-person exam or a referral, not just a note to the patient or a re-test.

Show it as:
Why a tag
  • The three tiers answer when. “Needs an exam” answers how. Making it a tier would put one step in two places.

3. Which measures make up each domain?

Each measure is placed on its own, not each test, because one test can measure two things. Our proposal, one line of reasoning each. A ? marks the ones we're least sure of.

DomainMeasuresWhy
Balance & gaitSway: all 15Standing still
Balance & gaitTimed Up & Go: total time, turns, transitions, sitting back downWalking and turning
Balance & gait ?Timed Up & Go: sit-to-stand power and durationRising from a chair. Could be Motor (leg strength)
Balance & gait ?Timed Up & Go: delay before the first stepCould be Cognition (planning the step)
Balance & gait10 m walk and everyday walking: allWalking
MotorFinger tapping: allHand speed and rhythm
MotorLine & circle, spiral: allDrawing accuracy and tremor
MotorReach: allArm reach
Motor ?Arm Raise: allArm, like reach; shoulder, not hand. Listed with reaction time. Could sit elsewhere
MotorReaction time: simple and choice timeA phone tap is mostly hand speed after stroke [12]
Motor ?Reaction time: variabilityCould be Cognition (attention)
CognitionReaction time: choice cost, error rateChoice minus simple takes hand speed out. A wrong side is a decision error
CognitionSky Search (game)Far-space neglect
SpeechAll speech measuresMotor speech from a read passage. “Words recognised” is clarity, not word-finding
Reaction time feeds two domains. Is that right?
The rest of the table:

Two placements are left:

Where does the Barthel (daily-living independence) check-in go?
Where does adherence (“This week”) go?
What we found
  • Domains in the current design: Balance & gait, Motor, Speech, Cognition, Mood, Daily living. Fugl-Meyer would join Motor once it has a change rule.
  • Reaction time: simple reaction time measures attention [11], but on a motor-heavy task it mostly reflects hand speed after stroke [12]. A phone tap is motor-heavy.
  • Sky Search: the old version measured “range of motion”; the current one measures far-space neglect. The old label is still in three places in the code and will be fixed.

4. A domain improved in one test and declined in another. What one word does it get?

The domain's word:
What we found
  • The game scoring already works this way: each domain is graded once, at its worst marker.
  • Unconfirmed moves only appear in the small line under the word. A flagged domain always shows a red ■ Flag.

5. Mood

The current design: one Mood word on every view, PHQ-9 detail one click down.

Who owns a mood step?
Which scales feed the Mood word?
Flag mood when the PHQ-9 reaches 10 (“moderate”)?
The word for “no real change”:
What we found
  • The app already gives both PHQ-9 and GAD-7.
  • The PHQ-9 severity bands are 5/10/15/20 [8]. For real-change thresholds, see 0b.
  • About a third of stroke survivors get depression, which is linked to worse recovery and higher mortality [13]. That's why mood shows on every view.

6. Who owns which steps?

The current design: each step has one owner (Physician, Therapist or Coordinator). Your own steps are marked Yours and listed first within a tier.

Owners:
The “Yours” marking:
What we found
  • Balance and falls → PT [1]. Speech, language and cognition → SLP [14]. App and phone problems → coordinator.
  • The guidelines give mood and safety no single owner.

7. Which view presets can merge?

Six presets in three families. Medical: stroke physician, PM&R, primary care. Rehab: PT/OT, SLP. Coordination: stroke coordinator. A preset only changes order and emphasis; it never hides anything.

Presets:
The test
  • Ask each role for the three things they'd look at first in a 3–5 minute review. Any two roles with the same list can share a preset.
  • The 2026 guideline publishes separate take-home messages for each discipline [1][14], so presets by discipline match how it divides the work.

8. Medications and the care plan

Clinicians may want to record medications and treatment plans, and to see next steps with medical reasoning (e.g. “DAPT for 90 days”). The current design only reminds them of their own plan (“Antiplatelet course ends around Dec 1”). It never proposes a drug.

What should the system do?
How is it entered?
Who can edit it?
What we found
  • The DAPT example matches the AAN's 2022 advisory: clopidogrel plus aspirin for up to 90 days in symptomatic 70–99% intracranial stenosis [15].
  • FDA's January 2026 guidance relaxed the rules for software that recommends a drug, but not when the recommendation comes from sensor signals [3]. Our biomarkers are sensor signals, so suggestions could only come from history a clinician entered.
  • Typed text collects names and dates, which breaks de-identification (see 14).

9. Steps suggested by the AI

The current design: AI suggestions can appear on the cover, with a dashed “Suggested” label and evidence. They can add a step but never remove or change one.

Where can AI suggestions appear?
What we found
  • Research on automation bias finds clinicians both follow wrong advice and skip actions they weren't prompted for [16].
  • Keeping suggestions out of Top priority means an empty Top priority always means nothing urgent, not that the AI missed something.

10. Who reads the report?

The current design: the lead clinician's 3–5 minute read before a visit or handoff, and the copy that goes in the chart.

The report is read by:

11. Is the desktop quiet enough?

The current design: patient tiles on the desktop show no test values (“Mood worse · two check-ins in a row”, not “PHQ-9 7 → 12”), and patients who are on track show as a count rather than as tiles. Both can be switched back in settings.

The desktop is:

The patient list

12. Identicons or dots?

The identicon is the small pattern beside each token. It is made from the token alone, so it tells you nothing about the person. Its job is to make similar tokens (G3GT-Z6DK / G3GT-Z6KD) easy to tell apart.

Is the identicon a problem?
Beside each token, show:
What we found
  • HIPAA allows a random code that isn't derived from the person. The token is random, and the identicon comes only from the token [4].
  • The photo rule covers images of the person, not a pattern made from a random code.

13. What makes the patient list feel crowded?

Mostly:

For a privacy lawyer

14. Do the tokens keep us out of HIPAA?

Our reading: not for Deployed clinics. The platform becomes a business associate and needs a signed BAA. And the data isn't de-identified under Safe Harbor, because every test has a date and the phone's push token is stored.

Deployed clinics need a BAA:
For Research clinics, which route keeps the test dates?
What we found
  • HHS's app guidance: a developer hired by a provider, whose app the provider tells patients to use, is a business associate [17].
  • Safe Harbor removes every date element except the year, and device identifiers [4]. Dates of service are the most common reason data fails it [18].
  • A limited data set may keep dates, but only for research, public health or operations, under a data use agreement. It still excludes device identifiers, so the push token must be hashed, and the recipient may not contact patients. That fits Research clinics, not Deployed ones [4].
  • Outside HIPAA, the FTC Health Breach Notification Rule covers health apps (since July 2024) [19].

15. Sharing with doctors outside the clinic

E.g. the patient's primary care doctor at another organisation. HIPAA allows sharing for treatment without the patient's authorization [20].

How should they get access?

Anything else?

Sources
  1. AHA/ASA 2026 adult stroke rehabilitation guideline, summaries: NeurologyLive, HealthDay, PT take-home messages. Class/level checked against summaries, not the full text (Stroke, doi 10.1161/STR.0000000000000536).
  2. Falls after stroke: Cochrane review; community cohorts.
  3. FDA revised CDS guidance (Jan 2026): Covington summary, FDA guidance PDF.
  4. 45 CFR 164.514(b)(2), (c) and (e): Cornell LII.
  5. PHQ-9 5-point change (Löwe et al. 2004): UKE record.
  6. Bauer-Staeb et al. 2021, baseline-dependent MCID: PMC8485844.
  7. Toussaint et al. 2020, GAD-7 MCID 4 points: UKE record.
  8. Kroenke, Spitzer & Williams 2001, PHQ-9 severity cutpoints: UW summary.
  9. Positive item 9 → structured risk assessment: Care New England, Ann Fam Med 2021.
  10. Item 9 vs eC-SSRS (PPV 28.6%): PubMed 29477096.
  11. Attention assessment after stroke, scoping review: PeerJ.
  12. Motor demands confound speed measures after stroke: Front Neurol 2017.
  13. AHA/ASA scientific statement on post-stroke depression (Towfighi 2017): PHQ-9 validity in PSD, statement doi 10.1161/STR.0000000000000113.
  14. SLP role, 2026 guideline: AHA take-home messages for SLPs.
  15. AAN 2022 practice advisory on intracranial atherosclerosis: HCPLive, AAN press release.
  16. Goddard, Roudsari & Wyatt, automation bias systematic review (JAMIA 2012): PMC3240751.
  17. HHS OCR, Health App Use Scenarios & HIPAA (Feb 2016): PDF.
  18. Expert determination and dates of service: J. D. Cook; HHS guidance on de-identification.
  19. FTC Health Breach Notification Rule, 2024 final rule: Covington, MoFo.
  20. 45 CFR 164.506(c), treatment disclosures: HHS fact sheet.

Not legal or medical advice: a starting point for the people who sign off.