Machine learning for workplace mental health is still mostly a lab exercise: researchers prioritize algorithmic novelty over domain-driven, causal, or deployment studies. Without interdisciplinary co‑design and randomized or quasi‑experimental evaluation linking predictions to wages, absenteeism and productivity, ML’s economic value for workplaces will remain unproven.

Main Finding

The use of machine learning (ML) in occupational mental health is still nascent and largely centered on methodological and computer-science development. To realize ML’s substantive potential for improving occupational health research and policy, greater interdisciplinary collaboration—bringing together ML specialists, occupational health researchers, economists, clinicians, and policymakers—is required.

Key Points

• Current work emphasizes algorithmic and methodological innovation rather than applied, domain-driven questions in occupational mental health.
• There is limited integration of domain expertise (psychology, occupational health, economics) into model design, validation, and interpretation.
• Practical deployment, external validity, causal identification, and evaluation of interventions remain underexplored.
• Ethical, privacy, and fairness concerns (sensitive health data, workplace power dynamics) receive insufficient attention in many methodological studies.
• Interdisciplinary collaboration is necessary to move from proof-of-concept models to tools that can inform workplace policy and practice.

Data & Methods

• Typical data sources used or needed:
  • Self-report surveys and standardized mental-health instruments.
  • Administrative/HR records (absences, turnover, performance metrics).
  • Passive digital traces and sensors (smartphone usage, wearables, email/text metadata) — high potential but privacy-sensitive.
  • Clinical records where available (with strict privacy controls).
  • Longitudinal cohort studies to assess dynamics and causal effects.
• Common ML methods observed or relevant:
  • Supervised learning for prediction (classification/regression of mental-health outcomes).
  • Natural language processing (NLP) for text-based signals (emails, chat, free-text survey responses).
  • Unsupervised learning for clustering and phenotyping.
  • Representation learning and transfer learning to leverage heterogeneous data sources.
  • Emerging integration with causal-inference approaches (causal forests, targeted learning) is limited but necessary.
• Methodological gaps:
  • Few studies combine predictive ML with causal identification or randomized evaluation.
  • Model validation on out-of-sample, cross-population, and real-world deployment settings is rare.
  • Insufficient attention to measurement error, confounding, and sample selection typical in workplace data.
  • Data governance, de-identification, and privacy-preserving techniques (federated learning, differential privacy) are underutilized.

Implications for AI Economics

• Measurement improvements: ML can produce higher-frequency, granular measures of worker mental health and stress, improving labor-market models of productivity, absenteeism, job exit, and human capital accumulation.
• Policy targeting and evaluation: Predictive models could enable earlier, better-targeted interventions (EAPs, workload adjustments), but must be paired with causal evaluation to assess impact on economic outcomes.
• Labor-market dynamics: Better identification of mental-health-related productivity shocks can refine estimates of wage penalties, turnover costs, and the returns to workplace mental-health programs.
• Cost–benefit and adoption: Economists can quantify the welfare and firm-level returns to ML-enabled interventions; adoption will hinge on demonstrated effectiveness, legal compliance, and trust.
• Risks and distributional effects: ML deployment in workplaces risks bias and adverse selection; economic analysis should assess heterogeneity across occupations, firms, and worker types to avoid exacerbating inequalities.
• Research agenda for AI economics:
  • Foster co-designed studies: pair ML teams with occupational-health researchers and economists from project inception.
  • Prioritize datasets that link mental-health measures to economic outcomes (wages, productivity, employment spells).
  • Develop standards for external validation, causal inference, and transparent reporting of algorithmic performance and harms.
  • Study governance, privacy, and incentive structures influencing firm and worker uptake.
  • Evaluate scalable interventions with randomized or quasi-experimental designs to estimate welfare impacts.

In short, realizing the economic and policy value of ML in occupational mental health requires shifting from isolated methodological work toward interdisciplinary, evaluation-focused research that links predictive tools to causal evidence and economic outcomes.