(HOPES Project > HTML Text Understanding)

Turning the “so-what?” into

actionable thresholds

Below is a quick‐and‐dirty cheat-sheet that teams building digital-phenotyping dashboards often start from.  Think of every number as a first draft that should be stress-tested on your own cohort (different phones, cultures, meds, etc.).  Where the literature gives a concrete effect size, I quote it; where it doesn’t, I show a pragmatic “%-change-from-baseline” rule that research groups and commercial apps are converging on.

Signal (24 h aggregate unless noted) Threshold that tends to matter What it may flag Key source(s)
Sleep duration (Fitbit / Oura) < 6 h in a night or ≥ 30 %↓ from 2-week personal median Incipient mania, relapse risk Reduced need for sleep is reported by 69-99 % of manic patients
HRV (night-time RMSSD) ≥ 30–50 %↓ from euthymic baseline (often ≈ 15–20 ms absolute drop) Autonomic arousal, impending mania, high stress Mania shows within-subject RMSSD drop vs. euthymia
Steps / GPS distance ≥ 30 %↓ daily steps or travel radius < 7 mi (healthy controls ≈ 12 mi) Social withdrawal, negative symptoms, bipolar depression Schizophrenia sample 7 mi vs. 12 mi in controls
Night-time screen minutes (00:00-06:00) > 120 min and ≥ 50 % above baseline Sleep disruption → mood lability Meta-review on passive sensing & depression finds screen-time spikes predictive of low mood
Call / text bursts (08:00-22:00) > 50 %↑ outgoing + incoming frequency for ≥ 2 days Hypomania / mania, irritability Multiple DP studies on bipolar communication surges (summarised in your table)
Keyboard backspace ratio > 0.26 (high), 0.18 (medium), < 0.11 (low) High group associated with mania, low with controls Three-class mixture model on 2 948 day-records
Median inter-key delay (IKD) Start with personal baseline.  Flag if ≥ 25 % faster (psychomotor agitation) or ≥ 25 % slower (psychomotor retardation) for ≥ 8 h rolling window.  Population means: IKD ≈ 0.69 s (SD = 0.36) Faster → possible mania / anxiety; slower → depression / fatigue No fixed cut-off yet; baseline deviation performs best across studies

Why the literature is thin on “tapping speed” cut-offs

Keystroke dynamics are continuous, device- and age-dependent.  Absolute millisecond cut-offs fail quickly in the wild.  Most research therefore treats typing metrics exactly as you suggested for sleep: detect a significant deviation from the person’s own baseline.

Practical recipe that works in pilots:

  1. Collect 14 days of “stable” data (no admissions, no YMRS spikes).
  2. Compute the median IKD per day.
  3. Z-score today’s value against that rolling window.
  4. Trigger an event if |z| ≥ 1.5 (≈ 25 % change for IKD) and another corroborating stream changes (e.g., backspace-rate or sleep).

This approach gave AUC = 0.80 for manic prediction in the BiAffect dataset when combined with HRV and sleep.

Building a simple rule engine

graph LR
  RawSensors -->|14-d rolling median| BaselineCalc
  RawSensors --> Today
  BaselineCalc --> Δ
  Today --> Δ
  Δ -->|threshold table| Flag
  Flag --> CompositeScore
  HRV & Sleep & Typing & Mobility --> CompositeScore
  CompositeScore --> Alert(Clinician / EMA prompt)