Multi-signal predictive monitoring for early psychiatric relapse warning
Predictive relapse-alert systems for conditions like major depressive disorder, bipolar disorder, and schizophrenia rest on a body of "digital phenotyping" research (a term coined by Jukka-Pekka Onnela at Harvard) showing that passively and actively collected behavioral data — mood ratings, sleep patterns, physical activity, sometimes smartphone usage patterns and voice/text sentiment — carry detectable signal about impending symptom changes, often before the patient or clinician would otherwise notice.
Relapse-prediction systems typically combine several complementary data streams, each capturing a different facet of the prodromal (pre-relapse) symptom trajectory well documented in longitudinal psychiatric research:
• Self-reported mood ratings: brief daily or twice-daily 1–10 mood/energy scales, the most direct symptom proxy but subject to reporting fatigue and potential denial or minimization during early symptom onset • Sleep duration and quality: via wearable actigraphy or phone-based sleep tracking; sleep disruption is one of the most robustly documented prodromal signals across mood disorders — reduced sleep need is a core DSM-5 criterion for manic episodes, and sleep disturbance frequently precedes depressive episode onset by days to weeks • Physical activity: step count and movement patterns via phone accelerometer or wearable; both marked activity increases (agitation, mania) and decreases (psychomotor retardation, depression) carry signal, making activity a genuinely bidirectional indicator depending on the specific disorder trajectory • Journal or free-text sentiment: natural language processing applied to patient-authored journal entries or symptom-tracking free text, extracting sentiment polarity and linguistic markers (increased first-person singular pronoun use, absolutist language) that have been independently associated with depressive symptom severity in computational linguistics research • Smartphone usage metadata (in more advanced systems): call/text frequency and social contact patterns, screen time, geolocation-derived mobility radius — capturing social withdrawal, a well-documented relapse prodrome across multiple psychiatric conditions
A defining methodological requirement for relapse-prediction models — distinguishing them from most other clinical prediction tasks — is that population-average thresholds perform poorly, because inter-individual variability in normal mood, sleep, and activity patterns is enormous relative to the within-individual signal that actually predicts relapse.
A patient with a lifelong pattern of sleeping 5.5 hours per night and a patient who typically sleeps 8.5 hours present very differently in absolute terms, but a drop to 4 hours means something quite different for each — for the first patient it may represent a meaningful ~30% reduction signaling concern, while for the second it may represent an even larger absolute and relative change warranting more urgent attention. A population-average threshold ("alert if sleep <5 hours") would systematically under-flag the second patient and over-flag nothing useful for the first, missing the actual predictive signal, which lies in deviation from that specific individual's own established pattern, not in the absolute value.
This has led relapse-prediction system design toward "idiographic" (individual-specific, sometimes called "n-of-1") statistical modeling rather than purely "nomothetic" (population-level) approaches: each patient's model is calibrated primarily against their own historical data — typically requiring an initial 2–4 week baseline-establishment period of stable, non-relapse data collection before the system can reliably distinguish meaningful deviation from ordinary day-to-day noise in that individual's signal.
Most production systems use a hybrid: a population-level model provides reasonable initial predictions and cold-start behavior for new users before sufficient individual data accumulates, then progressively shifts weight toward the individual-specific baseline as more of that patient's own data becomes available — a form of empirical Bayes shrinkage common in personalized medicine modeling more broadly, balancing the statistical stability of population data against the higher relevance of individual-specific patterns.
The core clinical value proposition of predictive relapse-alert systems is the prodromal window: published research shows that subtle behavioral changes in sleep, activity, and mood frequently precede clinically obvious symptom escalation by several days to weeks, creating a genuine intervention opportunity that routine, infrequent clinical visits structurally cannot capture.
Longitudinal psychiatric research on relapse prodromes — including foundational work on bipolar disorder relapse signatures (e.g., studies using the Bipolar Prodrome Symptom Scale) and more recent digital-phenotyping cohort studies — consistently finds that no single behavioral signal reliably predicts relapse across all patients or all episode types; instead, the combined pattern across multiple signal streams carries substantially more predictive information than any one stream alone. A patient's sleep might fragment before mood visibly declines; another patient's activity level might shift while sleep remains superficially stable; still another might show declining journal-entry sentiment days before either sleep or activity changes appreciably.
This motivates the near-universal architectural choice in production systems to combine multiple weighted signal streams into a composite risk index, rather than alerting off any single threshold crossing — typically using either a rule-based weighted-sum approach (transparent and clinician-interpretable, but less adaptive) or a supervised machine-learning model trained on historical relapse-labeled data for that population (potentially more accurate but requiring larger labeled training datasets, which are harder to obtain given how comparatively rare, and how variably defined, "relapse" events are in real-world clinical data).
A structural motivation for these systems is the mismatch between typical outpatient follow-up frequency (commonly monthly or less for stable patients) and the days-to-weeks timescale over which a prodrome unfolds — a routine visit occurring after the prodromal window has already progressed to overt relapse cannot deliver an intervention during the window when it would be most effective, which is precisely the structural gap continuous digital monitoring is designed to close.
Setting the alert threshold in a relapse-prediction system involves a three-way tradeoff rarely present in this explicit form in other clinical prediction contexts: sensitivity (catching true relapses), specificity (avoiding false alarms and alert fatigue), and lead time (how far in advance of full relapse the alert fires) — and these three properties are not independent, improving one typically costs another.
Lowering the alert threshold to catch deviation earlier in the prodromal window increases achievable lead time (more days of advance warning) but also increases the false-alarm rate, since early-stage deviation is inherently noisier and harder to distinguish from ordinary day-to-day variability than fully-developed prodromal signal closer to actual relapse onset. Raising the threshold improves specificity and reduces false alarms but correspondingly compresses the lead time, arriving closer to — or in the worst case, after — the point where early intervention would have been most valuable.
Unlike many clinical prediction contexts where the optimal operating point can be chosen from a standard ROC-curve analysis in isolation, relapse-alert threshold-setting must also explicitly account for care-team response capacity, echoing the same alert-fatigue dynamic documented in DTx adherence-monitoring and other digital-health alerting systems: a threshold generating more alerts than the clinical team can meaningfully triage and respond to degrades the entire system's real-world effectiveness regardless of the underlying statistical sensitivity, since overwhelmed responders begin deprioritizing or delaying response to individual alerts.
Published relapse-prediction system evaluations (spanning bipolar disorder, schizophrenia relapse, and depression recurrence prediction studies) generally report sensitivity in the 70–90% range at operating thresholds chosen to keep false-alarm burden clinically manageable, with achievable lead time varying considerably by disorder (manic-episode prodromes tend to be somewhat more behaviorally distinctive and thus detectable with longer lead time than depressive-episode prodromes) and by the richness of available signal streams for that specific patient.
An accurate, well-timed alert has no clinical value without an effective response pathway behind it. The final and arguably most consequential design element of a relapse-prevention system is the clinician-in-the-loop response protocol: what actually happens, and how quickly, once an alert fires.
Well-designed systems specify, in advance, a graded response protocol tied to alert severity — a lower-severity deviation alert might trigger an automated check-in message or a brief self-report symptom check, while a higher-severity, high-confidence alert triggers a same-day or next-day clinician phone contact and potential medication or safety-plan review. This graduated design mirrors the severity-tiered response principles seen in other digital-health alerting systems (DTx dropout-risk outreach, mental-health chatbot triage escalation), reflecting a broader pattern across digital psychiatric tools: automated detection is paired with human clinical judgment for the actual intervention decision, rather than the system attempting closed-loop automated intervention.
The honest state of the randomized-trial evidence for relapse-alert systems specifically (as opposed to the underlying digital-phenotyping detection accuracy, which has a more mature evidence base) remains comparatively early and mixed — several published pilot and feasibility RCTs in bipolar disorder and schizophrenia relapse prevention show promising signal for reduced hospitalization or reduced full-relapse rates among alert-monitored patients, but effect sizes vary considerably and appear substantially dependent on response-protocol quality and speed rather than being a straightforward function of detection-algorithm accuracy alone — reinforcing that, as with DTx outreach effectiveness more broadly, the human response infrastructure surrounding a predictive alert is at least as important to real-world outcomes as the underlying model's statistical performance.
A recurring finding across published digital-phenotyping relapse-prevention pilots is that clinician trust and workflow integration — whether care teams actually review and act on alerts promptly within their existing workflow, rather than treating them as a secondary or ignorable notification stream — is frequently the single largest determinant of whether a statistically accurate detection system translates into measurable reductions in real-world relapse or hospitalization rates.