Heidi Smith
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Heidi Smith has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Hajira Dambha-Miller
Primary Care Research Centre, University of Southampton, Southampton, UK
3 shared publications - Lysanne Veerle Michels
University of Southampton, UK
3 shared publications - Andrew A. S. Soltan
Department of Engineering Science, University of Oxford, Oxford, UK
1 shared publication Hajira Dambha-Miller
University of Southampton, Primary Care Research Centre, Aldermoor Health Centre, Aldermoor Close, Southampton, UK
1 shared publicationLucy Smith
University of Southampton, Primary Care Research Centre, Aldermoor Health Centre, Aldermoor Close, Southampton, UK
1 shared publication- Lucy Smith
Primary Care Research Centre, University of Southampton, Southampton, UK
1 shared publication
Latest Publications
- Open Access
- 18 September 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Are Large Language Models amplifying bias in medical science? (Preprint)
UNSTRUCTURED Large Language Models (LLMs) offer unprecedented efficiency across medical research, yet current evaluation focuses almost exclusively on output accuracy, neglecting critical issues of fairness, bias, and global representation. Because LLMs retrieve from and synthesize an evidence base already skewed toward high-income populations, their widespread adoption creates a feedback loop of recursive amplification that repeatedly surfaces dominant data while filtering out novel hypotheses and evidence from under-researched groups. Even if individual outputs remain factually accurate, this process risks creating an artificial scientific consensus that fails to reflect global health needs. Because standard disclosure rules cannot track how AI alters the broader literature landscape, evaluation paradigms must shift from checking study-level accuracy to measuring field-level representation. Establishing longitudinal literature-surveillance frameworks to monitor distributional skew, source diversity, and demographic representation, alongside strict reporting standards for model usage, is essential to prevent LLMs from exacerbating structural health inequities and undermining scientific validity.
- 18 September 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessAre Large Language Models amplifying bias in medical science? (Preprint)
UNSTRUCTURED Large Language Models (LLMs) offer unprecedented efficiency across medical research, yet current evaluation focuses almost exclusively on output accuracy, neglecting critical issues of fairness, bias, and global representation. Because LLMs retrieve from and synthesize an evidence base already skewed toward high-income populations, their widespread adoption creates a feedback loop of recursive amplification that repeatedly surfaces dominant data while filtering out novel hypotheses and evidence from under-researched groups. Even if individual outputs remain factually accurate, this process risks creating an artificial scientific consensus that fails to reflect global health needs. Because standard disclosure rules cannot track how AI alters the broader literature landscape, evaluation paradigms must shift from checking study-level accuracy to measuring field-level representation. Establishing longitudinal literature-surveillance frameworks to monitor distributional skew, source diversity, and demographic representation, alongside strict reporting standards for model usage, is essential to prevent LLMs from exacerbating structural health inequities and undermining scientific validity.
- Open Access
- 28 May 2026
- Published by Elsevier in SSRN Electronic Journal
Equity Considerations in AI-Enabled Conversational Systems: A Cross-Sectional Comparative Study of Clinical Responses
Background: Artificial intelligence (AI)-enabled conversational tools are increasingly used for health information and clinical decision support. Concerns persist regarding systematic bias, limited transparency, and the potential to exacerbate health inequity in real-world use. Empirical evaluation of how these tools operationalise equity-relevant factors in clinical responses remains limited. We aimed to compare the inclusion of equity considerations across widely used AI tools in response to standardised health queries.Methods: In this cross-sectional comparative study, we developed clinical and health information queries reflecting patient and clinician perspectives across common scenarios. Responses were generated using multiple widely used AI tools (ChatGPT, Claude, Copilot, DeepSeek, Doubao, Gemini, and Grok). Queries were standardised and geolocation controlled to minimise localisation effects. Responses were assessed using a structured framework with inclusion of equity-relevant factors (age, sex or gender, ethnicity, socioeconomic context, and access to care), clarity of clinical guidance, and specification of non-diagnostic outputs.Findings: Substantial heterogeneity was observed. Grok demonstrated the highest inclusion of equity-relevant factors, followed by Claude and DeepSeek, whereas Doubao showed minimal inclusion. ChatGPT, Copilot, and Gemini showed intermediate performance. Age was referenced in over half of responses, whereas sex or gender and ethnicity were included in fewer than half. Socioeconomic context and access to care were infrequently represented. Although all AI tools provided clinical guidance, fewer than half explicitly specified outputs as non-diagnostic. Patterns of equity inclusion varied by geographic context across tools.Interpretation: AI tools demonstrate inconsistent and limited operationalisation of equity-relevant factors, with implications for clinical safety and differential performance across populations. Strengthening transparency requirements, embedding standardised equity evaluation frameworks, and aligning regulatory oversight with real-world deployment are essential to ensure these technologies mitigate rather than exacerbate health inequity.
- 28 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEquity Considerations in AI-Enabled Conversational Systems: A Cross-Sectional Comparative Study of Clinical Responses
Background: Artificial intelligence (AI)-enabled conversational tools are increasingly used for health information and clinical decision support. Concerns persist regarding systematic bias, limited transparency, and the potential to exacerbate health inequity in real-world use. Empirical evaluation of how these tools operationalise equity-relevant factors in clinical responses remains limited. We aimed to compare the inclusion of equity considerations across widely used AI tools in response to standardised health queries.Methods: In this cross-sectional comparative study, we developed clinical and health information queries reflecting patient and clinician perspectives across common scenarios. Responses were generated using multiple widely used AI tools (ChatGPT, Claude, Copilot, DeepSeek, Doubao, Gemini, and Grok). Queries were standardised and geolocation controlled to minimise localisation effects. Responses were assessed using a structured framework with inclusion of equity-relevant factors (age, sex or gender, ethnicity, socioeconomic context, and access to care), clarity of clinical guidance, and specification of non-diagnostic outputs.Findings: Substantial heterogeneity was observed. Grok demonstrated the highest inclusion of equity-relevant factors, followed by Claude and DeepSeek, whereas Doubao showed minimal inclusion. ChatGPT, Copilot, and Gemini showed intermediate performance. Age was referenced in over half of responses, whereas sex or gender and ethnicity were included in fewer than half. Socioeconomic context and access to care were infrequently represented. Although all AI tools provided clinical guidance, fewer than half explicitly specified outputs as non-diagnostic. Patterns of equity inclusion varied by geographic context across tools.Interpretation: AI tools demonstrate inconsistent and limited operationalisation of equity-relevant factors, with implications for clinical safety and differential performance across populations. Strengthening transparency requirements, embedding standardised equity evaluation frameworks, and aligning regulatory oversight with real-world deployment are essential to ensure these technologies mitigate rather than exacerbate health inequity.
- Open Access
- 29 April 2026
- Published by MDPI AG in Preprints.org
Environmental Scanning of Climate Adaptation Tools in Healthcare
Introduction: Extreme weather events are increasing in frequency and intensity due to climate change, contributing to substantial morbidity and mortality globally, particularly among vulnerable populations. In the UK, climate adaptation within health systems remains insufficiently developed. However, there is limited understanding of the tools currently available for the identification and management of populations at risk during extreme weather. This study aims to systematically characterise UK-based climate adaptation tools used in healthcare settings. Methods: Environmental scanning was conducted, because no centralised database exists for climate adaptation tools in healthcare, and many relevant resources are not captured in traditional academic or grey literature repositories. Structured Google searching by two independent reviewers enabled identification of publicly available and practice-oriented tools accessed in real-world settings. Eligible resources included UK-based tools designed for healthcare professionals, local authorities, or patients that incorporate meteorological data to mitigate climate-related health risks. Results: Nine tools met inclusion criteria, comprising e-learning platforms, online dashboards, integrated clinical software, and structured workflows. Most targeted healthcare professionals, with few targeting local authorities and none targeting patient self-management. Data inputs and outcomes measures were heterogeneous, spanning risks related to heat, cold, flooding, and air pollution. Reporting was inconsistent, as nearly half the tools were not publicly accessible and all demonstrated limited transparency. Conclusion: The current landscape of UK climate adaptation tools in healthcare is fragmented, with variability in accessibility, evidence, and scope. Clearer reporting and greater coordination in how such tools are catalogued may support more consistent and equitable responses to climate-related health risks.
- 29 April 2026
- Published by MDPI AG in Preprints.org
Open AccessEnvironmental Scanning of Climate Adaptation Tools in Healthcare
Introduction: Extreme weather events are increasing in frequency and intensity due to climate change, contributing to substantial morbidity and mortality globally, particularly among vulnerable populations. In the UK, climate adaptation within health systems remains insufficiently developed. However, there is limited understanding of the tools currently available for the identification and management of populations at risk during extreme weather. This study aims to systematically characterise UK-based climate adaptation tools used in healthcare settings. Methods: Environmental scanning was conducted, because no centralised database exists for climate adaptation tools in healthcare, and many relevant resources are not captured in traditional academic or grey literature repositories. Structured Google searching by two independent reviewers enabled identification of publicly available and practice-oriented tools accessed in real-world settings. Eligible resources included UK-based tools designed for healthcare professionals, local authorities, or patients that incorporate meteorological data to mitigate climate-related health risks. Results: Nine tools met inclusion criteria, comprising e-learning platforms, online dashboards, integrated clinical software, and structured workflows. Most targeted healthcare professionals, with few targeting local authorities and none targeting patient self-management. Data inputs and outcomes measures were heterogeneous, spanning risks related to heat, cold, flooding, and air pollution. Reporting was inconsistent, as nearly half the tools were not publicly accessible and all demonstrated limited transparency. Conclusion: The current landscape of UK climate adaptation tools in healthcare is fragmented, with variability in accessibility, evidence, and scope. Clearer reporting and greater coordination in how such tools are catalogued may support more consistent and equitable responses to climate-related health risks.