A groundbreaking study spearheaded by researchers at the Centre for Addiction and Mental Health (CAMH) has unveiled a critical flaw in the deployment of artificial intelligence (AI) within psychiatric care: AI systems trained to predict aggressive behavior in acute psychiatric settings can inadvertently intensify existing social and structural inequities. This pioneering research, recently published in npj Mental Health Research, reveals that machine learning models, when applied without comprehensive fairness evaluation, tend to disproportionately overestimate the risk of aggression among marginalized groups. As AI rapidly integrates into healthcare, ensuring equitable algorithmic outcomes emerges as an urgent necessity to prevent exacerbating disparities within mental health treatment.
The utilization of AI in clinical psychiatry pivots on predictive models designed to anticipate aggressive incidents, aiming to enable preemptive intervention and improve patient and staff safety. However, the foundational data these models learn from are often subjective behavioral assessments conducted by clinicians. These assessments, shaped by historical and systemic biases, introduce a latent form of prejudice that can propagate through the AI’s decision-making processes. Dr. Marta Maslej, a Staff Scientist at CAMH’s Krembil Centre for Neuroinformatics and co-author of the study, stresses that the subjective nature of psychiatric assessments forms a critical source of bias. Without embedding fairness considerations into AI development, these tools risk fostering mistrust and potentially inciting aggressive episodes that could have otherwise been avoided.
The impetus for this research stems from increasing adoption of AI in psychiatry across diverse healthcare environments globally, including countries like the Netherlands, Switzerland, China, the United States, and Canada. These AI systems aim to forecast violent or aggressive behavior to allow for timely intervention. Yet, previous investigations scarcely address whether their predictive accuracy is uniform across demographic and social strata, particularly in psychiatric contexts where patient experiences are deeply intertwined with social determinants. CAMH’s research team addressed this knowledge gap through a rigorous machine learning analysis utilizing electronic health records from over 17,000 inpatient admissions.
Their analyses illuminated stark disparities in model performance: false positive rates were markedly elevated for Black and Middle Eastern patients, males, individuals admitted by police, and those with unstable or communal living situations. These patterns suggest the AI intends to mitigate risk but ends up disproportionately flagging individuals already subject to structural surveillance and marginalization. This “algorithmic amplification” of bias not only misguides clinical decision-making but may entrench stereotypes and unevenly allocate resources or interventions, compounding systemic inequities inherent in psychiatric care.
This seminal work redefines the conceptualization of fairness in health AI as a foundational requirement rather than an ancillary feature. Integrating fairness analyses in model design and evaluation becomes essential to foster trust and uphold ethical standards in psychiatric AI applications. CAMH’s commitment to responsible AI underscores an ethical framework emphasizing transparency, patient-centeredness, and equity. The responsible application of AI in psychiatry must navigate the delicate balance between technological innovation and social justice, ensuring vulnerable populations are not inadvertently targetted or overlooked.
To foster these goals, the Krembil Centre’s Predictive Care Lab, co-led by Drs. Laura Sikstrom and Marta Maslej, applies innovative computational ethnography methods to scrutinize AI’s real-world impacts in mental health. This interdisciplinary approach combines machine learning with anthropological insights to detect and counteract biases, contributing novel perspectives toward equitable AI integration. Their team recently secured funding from Canadian Institutes of Health Research to develop FARE+ — an advanced AI framework engineered to identify key drivers of algorithmic bias and formulate mitigation strategies, facilitating more equitable and clinically relevant risk assessments.
Beyond predicting individual risk, Dr. Laura Sikstrom elucidates a transformative paradigm shift empowered by this research. The future of psychiatric AI lies in systemic bias detection rather than binary risk determination. By redirecting AI’s focus from isolated patient characteristics to broader patterns of inequity, this approach could help dismantle entrenched disparities in mental healthcare delivery. Such patient-centered tools embody health equity values, promoting safer environments for both patients and clinical staff and fostering therapeutic alliances unhampered by prejudicial algorithmic judgments.
The collaborative nature of the study, led by medical student and former research trainee Yifan Wang alongside senior researchers at KCNI, exemplifies the fusion of clinical expertise and data science critical for advancing just AI. Supported by a SSHRC Insight Development Grant and a Google Award for Inclusion Research, the work highlights the imperatives of interdisciplinary funding and cooperation to address complex socio-technical challenges. By dissecting AI’s domain-specific failures and intricacies, CAMH sets a precedent for responsible innovation that other mental health centers worldwide can emulate.
This research not only calls into question the readiness of current AI tools for widespread deployment in sensitive psychiatric environments but also spurs reflection on the ethical ramifications of automated decision-making in mental health. Patients’ trust, clinical effectiveness, and societal equity hinge upon the ability to design AI systems that are both accurate and just. As AI gains traction globally, CAMH’s study cautions policymakers and practitioners to prioritize fairness analyses alongside technological advancement to prevent unintended harm and facilitate equitable mental health outcomes.
In summary, this landmark investigation demonstrates that AI’s promise in predicting aggression in psychiatric care—while substantial—is intrinsically linked to addressing embedded biases within training data and clinical practice. It eloquently advocates that fairness in AI is not a peripheral concern but a core requirement to safeguard ethical, equitable, and effective mental health interventions. The push towards patient-centric, bias-aware AI systems ushers in a new era of compassionate technology that can enhance psychiatric care without perpetuating historical injustices.
As AI continues to reshape mental healthcare landscapes, the insights from CAMH’s research urge a paradigm where technological potential aligns with social responsibility. The development and deployment of AI predictive models must rigorously account for structural inequities, ensuring AI acts as a tool for inclusion rather than exclusion. In doing so, AI can evolve from a double-edged sword into a powerful catalyst for transforming mental health systems into fairer, more trustworthy, and clinically impactful spaces.
Subject of Research: Not applicable
Article Title: Fairness analysis of machine learning predictions of aggression in acute psychiatric care
News Publication Date: 2-Mar-2026
Web References:
https://www.nature.com/articles/s44184-026-00194-6
Keywords: Artificial intelligence, Clinical psychology, Machine learning, Fairness, Mental health, Psychiatric care, Bias mitigation, Algorithmic equity, Predictive modeling, Structural inequities
