RESEARCH · METHODOLOGY

PsychFlo Signal Methodology

A full technical account of the theoretical frameworks, signal design, ensemble model architecture, and validation approach behind PsychFlo's burnout and workforce risk scoring system.

Published May 2026 · Version 1.2 · Internal working paper

1. Theoretical Foundations

PsychFlo's scoring model is built on five peer-reviewed organisational behaviour frameworks, selected for their empirical validation, predictive relevance to workplace burnout, and operationalisability through passive behavioural signals.

Job Demands-Resources Model (JD-R)

Demerouti, Bakker, Nachreiner & Schaufeli (2001); Bakker & Demerouti (2017)

Foundation framework. JD-R dimensions map directly to PsychFlo's 9 signal categories. Job demands (workload, complexity, time pressure) and job resources (autonomy, feedback, social support) are operationalised through passive behavioural proxies rather than self-report.

Maslach Burnout Inventory (MBI) Dimensions

Maslach & Leiter (1997); Maslach, Schaufeli & Leiter (2001); Leiter & Maslach (2016 update)

MBI's three dimensions — emotional exhaustion, depersonalisation, and reduced personal accomplishment — are operationalised through communication cadence, participation asymmetry, and output quality signals respectively. No direct self-report is used.

Psychological Safety Scale

Edmondson (1999); Edmondson & Lei (2014); Edmondson (2019)

Participation asymmetry across meeting contributions, response latency differentials, and feedback frequency patterns serve as validated proxies for psychological safety climate. These signals are measured continuously rather than via periodic survey.

Effort-Recovery Theory

Meijman & Mulder (1998); Sonnentag & Fritz (2015)

Supports the 6–10 week predictive window design. Effort-recovery theory predicts that insufficient recovery from cumulative load produces measurable behavioural signal changes before clinical symptom onset, validating the use of behavioural proxies as early indicators.

Conservation of Resources Theory (COR)

Hobfoll (1989, 2001, 2011)

Resource depletion patterns — particularly resource loss spirals — are captured through trend analysis of composite signal scores over rolling 4-week windows. COR predicts accelerating decline once depletion passes threshold, which informs PsychFlo's alert escalation logic.

2. Signal Architecture — 9 Scoring Engines

PsychFlo collects passive behavioural metadata across nine signal categories. No message content, document content, or individual surveillance outputs are produced at any stage. All signals are aggregated to team level (minimum 5 members) before scoring.

01 · Work Pattern Load

After-hours activity drift, context-switch frequency, recovery time between high-density days

Job demands — time pressure, workload

02 · Meeting Density

Meeting-to-focus-time ratio, back-to-back meeting frequency, meeting attendance trends

Job demands — cognitive load, autonomy erosion

03 · Communication Cadence

Response latency trends, message volume drift, participation distribution shift

JD-R resources — social support; MBI depersonalisation proxy

04 · Output Quality Signals

Task completion rate trends, deadline proximity stress patterns, complexity-to-completion ratio

MBI reduced accomplishment proxy

05 · Participation Asymmetry

Meeting contribution distribution, async feedback frequency, initiative-taking patterns

Psychological safety proxy (Edmondson 1999)

06 · Collaboration Network

Interaction graph changes, cross-team contact frequency, manager interaction cadence

JD-R resources — social support, feedback availability

07 · Focus Block Integrity

Uninterrupted deep work availability, fragmentation index, calendar density patterns

Job demands — cognitive load; resource: autonomy

08 · Recovery Signal

Rolling 4-week composite trend, weekend disconnect patterns, escalation velocity

Effort-Recovery Theory (Meijman & Mulder); COR spirals

09 · Policy & Compliance Flags

HR policy exposure risk score, absence pattern anomalies, tribunal risk indicators

Legal and organisational health risk layer

3. Ensemble Model Architecture

Risk scores are produced by an ensemble of three model types, each operating on different signal subsets and time windows. Scores are combined using a weighted averaging approach tuned against retrospective organisational outcome data.

Gradient Boosted Trees (XGBoost)

Primary burnout risk classifier. Trained on time-series aggregated signal features across 4-week rolling windows. Produces probability scores (0–1) for team-level burnout risk state.

scikit-learn / XGBoost · Python ML pipeline · Supabase Postgres time-series store

Anomaly Detection Layer (Isolation Forest)

Identifies sudden signal departures from established team baselines. Operates independently of the classifier to surface acute risk events that may not yet register in the trend-based model.

scikit-learn IsolationForest · per-team baseline modelling · rolling z-score normalisation

Longitudinal Trend Model (ARIMA-derived)

Produces the 6–10 week forward signal trajectory. Derived from effort-recovery literature on signal latency between load accumulation and measurable behavioural change. Not a clinical prediction — a risk trajectory indicator.

statsmodels ARIMA · Vercel Edge Functions (score delivery) · Redis scoring cache

4. Predictive Window — The 6–10 Week Claim

Basis and limitations of the 6–10 week predictive horizon

The 6–10 week predictive window is derived from a synthesis of signal latency findings across 14 organisational behaviour frameworks, principally effort-recovery theory (Meijman & Mulder, 1998; Sonnentag & Fritz, 2015) and COR resource depletion research (Hobfoll, 2001). These frameworks characterise the typical lag between the onset of measurable behavioural change and clinically observable burnout — not a specific validated dataset from PsychFlo's own pilot.

PsychFlo's own validation against this window is ongoing. The claim should be understood as a theoretical framing derived from published research, with PsychFlo's prospective validation in progress via controlled pilot design (treatment vs. control, 90-day window, primary endpoint: voluntary turnover differential and sick day rate).

5. Privacy-Preserving Design

✓No message, document, or file content is ever accessed or processed — only metadata (timing, volume, frequency)
✓All signals are aggregated at team level. Minimum team size: 5 members. No individual-level outputs are technically possible.
✓Signal data is anonymised at the collection layer before entering the scoring pipeline. Re-identification is architecturally prevented.
✓Employees have full access to view and delete their own signal contribution data via the employee portal.
✓Opt-in for all employee-facing features. Opt-out does not affect employment record or trigger any management flag.
✓Data Processing Agreement (GDPR Article 28) available for all customers. Regional data residency by default.

6. Validation Status and Limitations

Internal signal review

Complete

Signal architecture reviewed against published OB literature. Each signal category mapped to at least one peer-reviewed framework with validated construct validity.

Pilot deployment data

In progress

Three pilot organisations (28–160 employees) active. Prospective data collection underway. Full validation analysis to be published when sample size is sufficient for significance testing.

Third-party academic validation

Planned

Academic partnership with OB research group in progress. Peer-reviewed publication of signal methodology and validation results targeted.

Prospective controlled trial

Planned

Randomised design: treatment group (PsychFlo active) vs. control (standard HR). 90-day window. Primary endpoints: voluntary turnover and sick day rate differential.

7. Key References

Bakker, A. B., & Demerouti, E. (2017). Job demands–resources theory: Taking stock and looking forward. Journal of Occupational Health Psychology, 22(3), 273–285.

Demerouti, E., Bakker, A. B., Nachreiner, F., & Schaufeli, W. B. (2001). The job demands-resources model of burnout. Journal of Applied Psychology, 86(3), 499–512.

Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.

Hobfoll, S. E. (1989). Conservation of resources: A new attempt at conceptualizing stress. American Psychologist, 44(3), 513–524.

Hobfoll, S. E. (2001). The influence of culture, community, and the nested-self in the stress process. Applied Psychology, 50(3), 337–421.

Leiter, M. P., & Maslach, C. (2016). Burnout. In G. Fink (Ed.), Stress: Concepts, Cognition, Emotion, and Behavior (pp. 351–357). Academic Press.

Maslach, C., & Leiter, M. P. (1997). The truth about burnout. San Francisco: Jossey-Bass.

Maslach, C., Schaufeli, W. B., & Leiter, M. P. (2001). Job burnout. Annual Review of Psychology, 52, 397–422.

Meijman, T. F., & Mulder, G. (1998). Psychological aspects of workload. In P. J. D. Drenth & H. Thierry (Eds.), Handbook of Work and Organizational Psychology (Vol. 2, pp. 5–33). Psychology Press.

Sonnentag, S., & Fritz, C. (2015). Recovery from job stress: The stressor-detachment model as an integrative framework. Journal of Organizational Behavior, 36(S1), S72–S103.

Questions about the methodology?

We welcome scrutiny. Contact us for the full technical appendix or to discuss validation partnership.

Request technical appendix →