Below is a categorized list of core constructs—latent physiological, behavioral, cognitive, and environmental processes—that researchers map onto digital proxies when studying serious mental illnesses (SMI). Each of these constructs captures a facet of daily functioning or physiology that’s often disrupted in schizophrenia, bipolar disorder, and severe depression.

Approach: I am starting from clinical constructs to explain their relevance, and then finding associations to digital phenotyping studies. Next, I want to determine appropriate thresholds—what constitutes "high" or "low" values for these measures.

The Big So what question after this exercise?

1. Sleep, Mobility, Phone usage, HRV seems to be used most, repeatably across clinical constructs.
2. Most papers flag “relative drop or spike” as a meaningful early warning signals rather than a raw/fixed numbers.
3. Better to combine streams of signals; e.g Short Sleep hours + Late night screen time + increased burst of text messages / tapping predict ?mania?
4. DP are proxy, direct or indirect to clinical constructs.
5. Digital engagement = cognitive/social footprint. Steps and GPS map the physical footprint; screen time, app logs, unlocks, and typing dynamics map the cognitive and social one — both lenses are needed.
6. Thresholds should be individualised, not absolute. There is no universal “15 app switches/hour = anxiety”. Use the patient as their own baseline (rolling 14–30 day window) and flag deviations on the residual after removing trend + seasonality (Z ≥ 2–3σ, KL divergence shift, or anomaly score >0.6 from an Isolation Forest).
7. Statistical anomaly ≠ clinical anomaly. A holiday, a new mobile game, or a work crisis can spike app-switching. Always cross-reference at least two passive streams (e.g. high app-switching plus low location entropy → anxiety spiral; high app-switching plus high mobility → just a busy day).

Digital Engagement — A Closer Look (Learnt 11 Jun 2026)

While Sleep, Mobility, and HRV map the physical footprint, digital engagement (screen time, app logs, unlocks, typing dynamics) maps the cognitive and social footprint. Yesterday's reading surfaced four sub-constructs worth tracking, plus the math underneath:

1. Attention Fragmentation — high app-switching frequency paired with very short sessions (open, close within 5s, jump to the next). Proxy for anxiety, mania/hypomania, and executive dysfunction.
2. Digital Cocooning — total screen time may stay constant or even rise, but the mix shifts from communication/productivity apps toward passive consumption (streaming, scrolling). Proxy for depressive withdrawal and social anxiety — and a useful complement to the GPS "stays home" signal.
3. Circadian Disruption (Diurnal Patterns) — nighttime unlocks and sustained screen activity between roughly 1–5 AM. Direct proxy for insomnia, sleep fragmentation, and (in bipolar) a transitioning manic phase.
4. HCI Micro-Mechanics — keystroke latency, typing speed, backspace frequency, scroll velocity. Slowed and error-prone interaction maps to psychomotor retardation (depression). Erratic, hyper-fast typing and rapid scrolling map to impulsivity / racing thoughts (mania, anxiety). Pioneered by Mindstrong; correlates with neuropsych tests for executive function and processing speed.

How the math actually works (under the hood):

• Dynamic baseline: Decompose the behaviour time series as $Y(t) = T(t) + S(t) + e(t)$ — trend + seasonality + residual. Digital phenotyping lives in the residual.