A 2026 observational study led by the Centre for Addiction and Mental Health (CAMH), published in npj Mental Health Research, trained machine learning models on electronic health records from over 17,000 psychiatric inpatients to predict aggressive incidents. The analysis revealed significantly higher false-positive rates for Black and Middle Eastern patients, men, individuals admitted by police, and those with unstable housing. This retrospective cohort study, while methodologically rigorous in its fairness auditing across intersecting demographic and social factors, relies on subjective clinician-documented labels known to embed systemic biases. No conflicts of interest were reported, yet its single-center Canadian sample limits generalizability.

The MedicalXpress coverage summarizes these results and quotes lead researchers Dr. Marta Maslej and Dr. Laura Sikstrom effectively. However, it underplays the self-fulfilling prophecy dynamic: biased AI flags trigger increased surveillance, coercive interventions, and loss of autonomy, which can provoke the very aggression the model predicts, eroding patient trust and worsening outcomes. This mechanism was largely missed.

The findings fit a broader, troubling pattern. Obermeyer et al.'s 2019 observational study (n≈200,000, published in Science) exposed how a commercial US health algorithm underestimated Black patients' needs by using costs as a proxy for illness severity, a proxy distorted by unequal access. Likewise, the 2016 ProPublica investigation of the COMPAS recidivism tool demonstrated racial disparities with false-positive rates for Black defendants nearly twice that of white defendants. In psychiatry, these issues are magnified because 'aggression' labels derive from subjective observations historically tainted by bias, such as the mid-20th-century over-diagnosis of schizophrenia among Black Americans.

What current coverage often gets wrong is framing fairness as a secondary checkbox rather than a foundational requirement. As hospitals in the Netherlands, Switzerland, China, the US, and Canada accelerate AI deployment for de-escalation and resource allocation, the absence of mandatory equity audits, prospective RCTs measuring real-world clinical harm, or community co-design leaves marginalized groups vulnerable. Without shifting from binary individual risk scores to systemic bias detection, as Sikstrom advocates, these tools will automate and scale historical inequities.

Genuine solutions require technical fixes like reweighting training data or using counterfactual fairness metrics, paired with policy: regulatory oversight akin to FDA AI device guidance, transparent reporting of subgroup performance, and rejection of deployment when bias thresholds are breached. The CAMH team's computational-ethnographic approach in the Predictive Care Lab represents progress, yet broader adoption without safeguards risks deepening mistrust, increasing coercive care, and ultimately harming the very patients psychiatry aims to protect. This is not inevitable, but correcting course demands treating equity as non-negotiable engineering and ethical priority.