Hey, I'm Ariel
Guerra-Adames

News

Research Interests

My main research interest is in how we can use language models to quantify human biases in emergency medical decision-making. Essentially, fairness research but for humans rather than AI systems. I develop counterfactual methods using fine-tuned LLMs to detect and measure cognitive biases in clinical settings, with implications for healthcare equity across different populations and healthcare systems.

I have worked on several projects related to the analysis of long-unstructured text for different tasks ranging from sentiment analysis to topic modeling, and more recently, text classification on data from social media and the health domain. My work has been validated on data from both French (Bordeaux University Hospital) and American (MIMIC-IV) emergency departments.

I have also explored the application of Brain-Computer Interfaces with non-invasive electrodes during my Engineering thesis. This work consisted in the online processing of non-stationary EEG signals for the control of a device.

There's more to it than that however, if you have any questions regarding my past work, don't hesitate to ask!

Publications

* = Equal contribution

• A Counterfactual LLM Framework for Detecting Human Biases: A Case Study of Sex/Gender in Emergency Triage
  Ariel Guerra-Adames, Marta Avalos-Fernandez, Océane Dorémus, Leo Anthony Celi, Cédric Gil-Jardiné and Emmanuel Lagarde
  arXiv preprint arXiv:2511.17124. 2025
  
  Abstract PDF BibTeX
  We present a novel, domain-agnostic counterfactual approach that uses Large Language Models (LLMs) to quantify gender disparities in human clinical decision-making. The method trains an LLM to emulate observed decisions, then evaluates counterfactual pairs in which only gender is flipped, estimating directional disparities while holding all other clinical factors constant. We study emergency triage, validating the approach on more than 150,000 admissions to the Bordeaux University Hospital (France) and replicating results on a subset of MIMIC-IV across a different language, population, and healthcare system. In the Bordeaux cohort, otherwise identical presentations were approximately 2.1% more likely to receive a lower-severity triage score when presented as female rather than male; scaled to national emergency volumes in France, this corresponds to more than 200,000 lower-severity assignments per year. Modality-specific analyses indicate that both explicit tabular gender indicators and implicit textual gender cues contribute to the disparity. Beyond emergency care, the approach supports bias audits in other settings (e.g., hiring, academic, and justice decisions), providing a scalable tool to detect and address inequities in real-world decision-making.
  @misc{guerraadames2025counterfactualllmframeworkdetecting, title={A Counterfactual LLM Framework for Detecting Human Biases: A Case Study of Sex/Gender in Emergency Triage}, author={Ariel Guerra-Adames and Marta Avalos-Fernandez and Océane Dorémus and Leo Anthony Celi and Cédric Gil-Jardiné and Emmanuel Lagarde}, year={2025}, eprint={2511.17124}, archivePrefix={arXiv}, primaryClass={cs.CY}, url={https://arxiv.org/abs/2511.17124} }
  
  
• Beyond Ethics: How Inclusive Innovation Drives Economic Returns in Medical AI
  Balagopal Unnikrishnan*, Ariel Guerra-Adames*, Amin Adibi*, Sameer Peesapati, Rafal Kocielnik, Shira Fischer, Hillary Clinton Kasimbazi, Rodrigo Gameiro, Alina Peluso, Chrystinne Oliveira Fernandes, Maximin Lange, Lovedeep Gondara, Leo Anthony Celi
  arXiv preprint arXiv:2510.10338. 2025
  
  Abstract PDF BibTeX
  While ethical arguments for fairness in healthcare AI are well-established, the economic and strategic value of inclusive design remains underexplored. This perspective introduces the "inclusive innovation dividend" -- the counterintuitive principle that solutions engineered for diverse, constrained use cases generate superior economic returns in broader markets. Drawing from assistive technologies that evolved into billion-dollar mainstream industries, we demonstrate how inclusive healthcare AI development creates business value beyond compliance requirements. We identify four mechanisms through which inclusive innovation drives returns: (1) market expansion via geographic scalability and trust acceleration; (2) risk mitigation through reduced remediation costs and litigation exposure; (3) performance dividends from superior generalization and reduced technical debt, and (4) competitive advantages in talent acquisition and clinical adoption. We present the Healthcare AI Inclusive Innovation Framework (HAIIF), a practical scoring system that enables organizations to evaluate AI investments based on their potential to capture these benefits. HAIIF provides structured guidance for resource allocation, transforming fairness and inclusivity from regulatory checkboxes into sources of strategic differentiation. Our findings suggest that organizations investing incrementally in inclusive design can achieve expanded market reach and sustained competitive advantages, while those treating these considerations as overhead face compounding disadvantages as network effects and data advantages accrue to early movers.
  @article{unnikrishnan2025beyond, title={Beyond Ethics: How Inclusive Innovation Drives Economic Returns in Medical AI}, author={Unnikrishnan, Balagopal and Guerra-Adames, Ariel and Adibi, Amin and Peesapati, Sameer and Kocielnik, Rafal and Fischer, Shira and Kasimbazi, Hillary Clinton and Gameiro, Rodrigo and Peluso, Alina and Fernandes, Chrystinne Oliveira and Lange, Maximin and Gondara, Lovedeep and Celi, Leo Anthony}, journal={arXiv preprint arXiv:2510.10338}, year={2025} }
  
  
• Synthetic medical data generation: state of the art and application to trauma mechanism classification
  Océane Dorémus*, Ariel Guerra-Adames*, Marta Avalos-Fernandez, Valentin Jouhet, Cédric Gil-Jardiné and Emmanuel Lagarde
  Proceedings of CIBB 2025. 2025
  
  Abstract PDF BibTeX
  This paper provides a comprehensive review of synthetic medical data generation techniques, with a focus on their application to trauma mechanism classification. We explore various generative models and their effectiveness in creating realistic medical datasets while addressing privacy and data scarcity challenges.
  @InProceedings{doremus2025synthetic, title={Synthetic medical data generation: state of the art and application to trauma mechanism classification}, author={Dor{\'e}mus, Oc{\'e}ane and Guerra-Adames, Ariel and Avalos-Fernandez, Marta and Jouhet, Valentin and Gil-Jardin{\'e}, C{\'e}dric and Lagarde, Emmanuel}, booktitle={Proceedings of CIBB 2025}, year={2025} }
  
  
• Uncovering Judgment Biases in Emergency Triage: A Public Health Approach Based on Large Language Models 🏆 Best Paper Award
  Ariel Guerra-Adames, Marta Avalos-Fernandez, Océane Doremus, Cédric Gil-Jardiné and Emmanuel Lagarde
  Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:420-439. 2025
  
  Abstract PDF BibTeX
  Judgment biases in emergency triage can adversely affect patient outcomes. This study examines sex/gender biases using four advanced language models fine-tuned on real-world emergency department data. We introduce a novel approach based on the testing method, commonly used in hiring bias detection, by automatically altering triage notes to change patient sex references. Results indicate a significant bias: female patients are assigned lower severity ratings than male patients with identical clinical conditions. This bias is more pronounced with female nurses or when patients report higher pain levels but diminishes with increased nurse experience. Identifying these biases can inform interventions such as enhanced training, protocol updates, and machine learning tools to support clinical decision-making.
  @InProceedings{pmlr-v259-guerra-adames25a, title = {Uncovering Judgment Biases in Emergency Triage: A Public Health Approach Based on Large Language Models}, author = {Guerra-Adames, Ariel and Avalos-Fernandez, Marta and Doremus, Oc{\'{e}}ane and Gil-Jardin{\'{e}}, C{\'{e}}dric and Lagarde, Emmanuel}, booktitle = {Proceedings of the 4th Machine Learning for Health Symposium}, pages = {420--439}, year = {2025}, editor = {Hegselmann, Stefan and Zhou, Helen and Healey, Elizabeth and Chang, Trenton and Ellington, Caleb and Mhasawade, Vishwali and Tonekaboni, Sana and Argaw, Peniel and Zhang, Haoran}, volume = {259}, series = {Proceedings of Machine Learning Research}, month = {15--16 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v259/main/assets/guerra-adames25a/guerra-adames25a.pdf}, url = {https://proceedings.mlr.press/v259/guerra-adames25a.html}, abstract = {Judgment biases in emergency triage can adversely affect patient outcomes. This study examines sex/gender biases using four advanced language models fine-tuned on real-world emergency department data. We introduce a novel approach based on the testing method, commonly used in hiring bias detection, by automatically altering triage notes to change patient sex references. Results indicate a significant bias: female patients are assigned lower severity ratings than male patients with identical clinical conditions. This bias is more pronounced with female nurses or when patients report higher pain levels but diminishes with increased nurse experience. Identifying these biases can inform interventions such as enhanced training, protocol updates, and machine learning tools to support clinical decision-making.} }
  
  
• AI-Driven Emergency Patient Flow Optimization is Both an Unmissable Opportunity and a Risk of Systematizing Health Disparities
  Dylan Russon, Marta Avalos-Fernandez, Ariel Guerra-Adames, Cédric Gil-Jardiné and Emmanuel Lagarde
  The International FLAIRS Conference Proceedings, 37(1). 2024
  
  Abstract PDF BibTeX
  There is a burgeoning interest in harnessing artificial intelligence (AI) to enhance patient flow within emergency departments (EDs). However, this advancement is accompanied by a significant risk: by relying on historical healthcare data, these AI tools may perpetuate existing systemic biases associated with gender, age, ethnicity, and socioeconomic status. This paper surveys studies identifying biases in ED data, offering context for concern about these biases. These insights are valuable for researchers developing AI to optimize ED workflows while accounting for ethical considerations.
  @inproceedings{russon2024ai, title={AI-Driven Emergency Patient Flow Optimization is Both an Unmissable Opportunity and a Risk of Systematizing Health Disparities: Human Bias in ED: Implications for AI-Driven Patient Flow Optimization}, author={Russon, Dylan and Avalos, Marta and Guerra-Adames, Ariel and Gil-Jardin{\'e}, C{\'e}dric and Lagarde, Emmanuel}, booktitle={The International FLAIRS Conference Proceedings}, volume={37}, year={2024} }
  
  
• Harnessing Moderate-Sized Language Models for Reliable Patient Data De-identification in Emergency Department Records: An Evaluation of Strategies and Performance
  Océane Dorémus, Dylan Russon, Benjamin Contrand, Ariel Guerra-Adames, Marta Avalos-Fernandez, Cédric Gil-Jardiné et Emmanuel Lagarde
  JMIR AI, 2024.
  
  Abstract PDF BibTeX
  The digitization of healthcare, facilitated by the adoption of electronic health record (EHR) systems, has revolutionized data-driven medical research and patient care. While this digital transformation offers substantial benefits in healthcare efficiency and accessibility, it concurrently raises significant concerns over privacy and data security. Initially, the journey towards protecting patient data de-identification saw the transition from rule-based systems to more mixed approaches including machine learning for de-identifying patient data. Subsequently, the emergence of Large Language Models (LLMs) has represented a further opportunity in this domain, offering unparalleled potential for enhancing the accuracy of context-sensitive de-identification. However, despite LLMs offering significant potential, the deployment of the most advanced models in hospital environments is frequently hindered by data security issues and the extensive hardware resources required.
  @article{doremus2024harnessing, title={Harnessing Moderate-Sized Language Models for Reliable Patient Data De-identification in Emergency Department Records: An Evaluation of Strategies and Performance}, author={Dorémus, Océane and Russon, Dylan and Contrand, Benjamin and Guerra-Adames, Ariel and Avalos-Fernandez, Marta and Gil-Jardiné, Cédric and Lagarde, Emmanuel}, journal={JMIR AI}, year={2024}, }
  
  
• Analyzing Spanish-Language Public Sentiment in the Context of a Pandemic and Social Unrest: The Panama Case
  Fernando Arias*, Ariel Guerra-Adames*, Maytee Zambrano, Efraín Quintero-Guerra and Nathalia Tejedor-Flores.
  International Journal of Environmental Research and Public Health. 2022.
  
  Abstract PDF BibTeX
  Over the past decade, an increase in global connectivity and social media users has changed the way in which opinions and sentiments are shared. Platforms such as Twitter can act as public forums for expressing opinions on non-personal matters, but often also as an outlet for individuals to share their feelings and personal thoughts. This becomes especially evident during times of crisis, such as a massive civil disorder or a pandemic. This study proposes the estimation and analysis of sentiments expressed by Twitter users of the Republic of Panama during the years 2019 and 2020. The proposed workflow is comprised of the extraction, quantification, processing and analysis of Spanish-language Twitter data based on Sentiment Analysis. This case of study highlights the importance of developing natural language processing resources explicitly devised for supporting opinion mining applications in Latin American countries, where language regionalisms can drastically change the lexicon on each country. A comparative analysis performed between popular machine learning algorithms demonstrated that a version of a distributed gradient boosting algorithm could infer sentiment polarity contained in Spanish text in an accurate and time-effective manner. This algorithm is the tool used to analyze over 20 million tweets produced between the years of 2019 and 2020 by residents of the Republic of Panama, accurately displaying strong sentiment responses to events occurred in the country over the two years that the analysis performed spanned. The obtained results highlight the potential that methodologies such as the one proposed in this study could have for transparent government monitoring of responses to public policies on a population scale.
  @article{arias2022analyzing, title={Analyzing Spanish-Language Public Sentiment in the Context of a Pandemic and Social Unrest: The Panama Case}, author={Arias, Fernando and Guerra-Adames, Ariel and Zambrano, Maytee and Quintero-Guerra, Efra{\'\i}n and Tejedor-Flores, Nathalia}, journal={International Journal of Environmental Research and Public Health}, volume={19}, number={16}, pages={10328}, year={2022}, publisher={MDPI} }
  
  
• Sentiment Analysis of Public Social Media as a Tool for Health-Related Topics
  Fernando Arias, Maytee Zambrano, Ariel Guerra-Adames, Nathalia Tejedor-Flores and Miguel Vargas-Lombardo.
  IEEE Access. June 2022.
  
  Abstract PDF BibTeX
  For decades, researchers have experimented with the possibility that machines can equal human linguistic capabilities. Recently, advances in the field of natural language processing (NLP) as well as a substantial increase in available naturally occurring linguistic data on social media platforms have made more advanced methodologies such as sentiment analysis (SA) gain substantial momentum on contemporary applications. This document compiles what the authors consider to be some of the most important concepts related to SA, as well as techniques and processes necessary for the various stages of its implementation. Furthermore, specific applications related to the extraction and classification of social media data using novel SA techniques are presented and quantified, with an emphasis on those used for the identification of mental health degradation during the COVID-19 pandemic. Finally, the authors present several conclusions highlighting the most prominent benefits and drawbacks of the methods discussed, followed by a brief discussion of possible future applications of certain methods of interest.
  @article{arias2022sentiment, title={Sentiment analysis of public social media as a tool for health-related topics}, author={Arias, Fernando and Nunez, Maytee Zambrano and Guerra-Adames, Ariel and Tejedor-Flores, Nathalia and Vargas-Lombardo, Miguel}, journal={IEEE Access}, volume={10}, pages={74850--74872}, year={2022}, publisher={IEEE} }
  
  
• Control of a Simulated Terrestrial Vehicle Through the Real-Time Processing and Classification of Electroencephalogram Data
  Ariel Guerra-Adames, Danilo Cáceres and Fernando Merchan.
  Global Medical Engineering Physics Exchanges/Pan American Health Care Exchanges (GMEPE/PAHCE). March 2022.
  
  Abstract PDF BibTeX
  In recent years and in parallel with advances in commercially available computing power, the brain computer interface (BCI) has emerged as a new use for the information obtained from existing electroencephalogram (EEG) acquisition technologies. This system works as a communication pathway between an individual's central nervous system and a computing unit. The BCI described in this study has as primary directive the control of a simulated vehicle, for which it utilizes both neurological activity associated to an individual's state of focus in the frequency domain, as well as artifacts produced by intentional blinks, found in the time domain. In order to successfully translate the user's intention into commands, the acquired signals are spectrally and temporally pre-processed, features of interest are extracted by means of the wavelet packet decomposition (WPD), and are finally fed into a convolutional neural network (CNN) model, which outputs recursive predictions about the user's mental state. The set of these general steps is capable of emitting navigational commands to the vehicle, and at the same time, return a user's focus state related to the task in hand in the form of a visual feedback loop. Tests performed with the complete system, demonstrate that a system as the one proposed is satisfactorily functional, being capable of carrying the user's intent to the simulated vehicle with a latency and margin of error within what's expected for BCI applications based on EEG readings.
  @inproceedings{guerra2022control, title={Control of a Simulated Terrestrial Vehicle Through the Real-Time Processing and Classification of Electroencephalogram Data}, author={Guerra-Adames, A and C{\'a}ceres, D and Merch{\'a}n, F}, booktitle={2022 Global Medical Engineering Physics Exchanges/Pan American Health Care Exchanges (GMEPE/PAHCE)}, pages={1--1}, year={2022}, organization={IEEE} }
  
  
• Real Time Operator Focus Monitoring System Based on a Brain-Computer Interface
  Ariel Guerra-Adames, Danilo Cáceres, Fernando Merchan and Kang-Hyun Jo.
  IEEE 29th International Symposium on Industrial Electronics (ISIE). June 2020.
  
  Abstract PDF BibTeX
  Recent advances in applications of Brain-Computer Interfaces (BCIs) for control and communication in more complex systems have demonstrated their potential for the augmentation of human cognition. This paper aims to develop a BrainComputer Interface application capable of monitoring the focus level of a user in real time and produce a binarized output signal based on such electrophysiological activity to show that a real-time, low-cost BCI can be implemented in industrialized settings. This can be achieved by applying a signal processing stage consisting of a DC offset subtraction layer, a bandpass filter bank and a number of thresholds tuned for each individual user. Experiments performed showed that while it is possible to monitor the desired neurological features, a number of undesired artifacts present at all times in brain signals make online analysis especially challenging.
  @inproceedings{guerra2020real, title={Real Time Operator Focus Monitoring System Based on a Brain-Computer Interface}, author={Guerra, Ariel and Caceres, Danilo and Merchan, Fernando and Jo, Kang-Hyun}, booktitle={2020 IEEE 29th International Symposium on Industrial Electronics (ISIE)}, pages={1415--1420}, year={2020}, organization={IEEE} }
  
  
• Bioacoustic classification of Antillean manatee vocalization spectrograms using deep convolutional neural networks
  Fernando Merchan, Ariel Guerra-Adames, Héctor Poveda, Héctor M. Guzman and Javier E. Sanchez-Galan.
  Applied Sciences. May 2020.
  
  Abstract PDF BibTeX
  We evaluated the potential of using convolutional neural networks in classifying spectrograms of Antillean manatee (Trichechus manatus manatus) vocalizations. Spectrograms using binary, linear and logarithmic amplitude formats were considered. Two deep convolutional neural networks (DCNN) architectures were tested: linear (fixed filter size) and pyramidal (incremental filter size). Six experiments were devised for testing the accuracy obtained for each spectrogram representation and architecture combination. Results show that binary spectrograms with both linear and pyramidal architectures with dropout provide a classification rate of 94–99% on the training and 92–98% on the testing set, respectively. The pyramidal network presents a shorter training and inference time. Results from the convolutional neural networks (CNN) are substantially better when compared with a signal processing fast Fourier transform (FFT)-based harmonic search approach in terms of accuracy and F1 Score. Taken together, these results prove the validity of using spectrograms and using DCNNs for manatee vocalization classification. These results can be used to improve future software and hardware implementations for the estimation of the manatee population in Panama.
  @article{merchan2020bioacoustic, title={Bioacoustic classification of Antillean manatee vocalization spectrograms using deep convolutional neural networks}, author={Merchan, Fernando and Guerra, Ariel and Poveda, H{\'e}ctor and Guzm{\'a}n, H{\'e}ctor M and Sanchez-Galan, Javier E}, journal={Applied Sciences}, volume={10}, number={9}, pages={3286}, year={2020}, publisher={MDPI} }

Teaching

I teach 64 hours per year at the Master's and PhD level at the University of Bordeaux, covering core topics in machine learning and its applications to healthcare:

Talks & Presentations

• Nov 2025 "Comment l'IA Peut Débusquer les Biais de Genre" at Intelligence Artificielle & Néphrologie 2025, Paris, France


• Apr 2025 "Synthetic Medical Data Generation : State of the art and application to trauma mechanism clssification" at Dataquitaine 2025, Bordeaux, France


• Dec 2024 🏆 Best Paper Award 🏆 "Uncovering Judgment Biases in Emergency Triage: A Public Health Approach Based on Large Language Models" at ML4H Symposium, Vancouver, Canada


• Jul 2024 "Réflexions pour la conception d'un protocole expérimental de détection des biais dans le triage d'urgence hospitalière à l'aide de modèles de langage" at EvalLLM2024, Toulouse, France


• Mar 2024 "Optimisation des services des urgences hospitalières à l'aide de l'IA : une opportunité incontournable et un risque de systématisation des disparités de santé" at Dataquitaine 2024, Bordeaux, France


• Jan 2024 🏆 Best Pitch Award🏆 "Learning PDEs from Data: Can We Solve the Non-registration Issue?" at Manutech SLEIGHT Science Event #11, Saint-Étienne, France


• Jun 2021 "Control de un Vehículo Mediante el Procesamiento e Interpretación en Tiempo Real de Señales de Electroencefalografía" at APANAC 2021, Panama City, Panama

In the Media

• 🎙️ Radio "Créer une IA médicale sans biais" — France Culture, Avec Sciences

  Interview discussing the development of inclusive medical AI and how LLMs can help detect cognitive biases in healthcare.

• 📰 Press "Biais cognitifs dans le soin : comment l'IA générative pourrait aider à améliorer la prise en charge" — INSERM Press Release