The Invisible Patients: A Systematic Review of Quiet Harms and Intersectional Bias in Clinical Large Language Models

Large Language Models (LLMs) hold significant promise for transforming healthcare delivery, yet their implementation risks reinforcing biases that may exacerbate existing health inequities. Prior work has established that bias exists in clinical LLMs, but a critical gap remains regarding how these biases distribute across sociodemographic groups, which high-risk clinical applications are most affected, and how multiple forms of disadvantage intersect – particularly for women in under-resourced settings, who face a compounding “Double Effect” of bias. Following PRISMA guidelines we systematically searched PubMed, IEEE Xplore, Scopus, ScienceDirect, and SpringerLink for studies published between 2022 and 2026, supplemented by manual search. Of 1,695 records identified, a four-tier screening process (identification, automated filtering, title/abstract screening, and full-text eligibility review) yielded 31 empirical studies for data synthesis, extracted across model, clinical domain, bias type, evaluation method, and key findings. Gender bias was the most frequently reported harm, appearing in 20 of 31 studies (64.5%): clinical-vignette studies found models under-triage female cardiac patients and attribute physiological pain to psychological causes, a “performative empathy gap” in which models offer authoritative guidance to male patients but simplistic responses to female ones, and synthetic patient generators that produced male patients 92.8% of the time. Racial and ethnic bias was second most common (16 of 31 studies, 51.6%), including reduced imaging recommendations for Black patients, specialized biomedical models inheriting bias from historically inequitable electronic health records, a Western cultural default in multilingual models that disadvantages Global South users, and age bias that collapses patient profiles toward a middle-aged default. Clinical Decision Support emerged as the highest-risk domain for ethical and performance inconsistency, while Patient Empowerment and Medical Education concentrated age, gender, and representational “erasure” biases, and Administrative Efficiency reflected socioeconomic and religious bias. These biases compound for minority women through a “Double Effect” of intersecting racial, gender, linguistic, and geographic disadvantage, producing what the literature terms the “invisible patient”: measurable collapses in representation, clinical safety, linguistic access, and resource allocation relative to a privileged, English-speaking baseline. Reviewed mitigation strategies span technical interventions – curated fine-tuning, demographic-token masking, counterfactual prompting – and governance measures such as mandatory bias-score disclosure and clinician AI-literacy training, though none has been validated at scale or in real-world deployment. Across the reviewed literature, bias in clinical LLMs is not merely a technical artifact but a structural reproduction of the sociodemographic inequities embedded in training data, and increasingly fluent models mask rather than eliminate it – an Automation Bias Paradox in which greater linguistic polish lulls clinicians into misplaced trust. We conclude that current LLMs are not yet safe for equitable routine clinical use, and that closing this gap requires mandatory ethical oversight, including standardized transparency labels and intersectional audits, rather than technical scaling alone.