Counterfactual Explanations for Patient and Student Outcomes: Surfacing 'What If?' Scenarios in Explainable NLP

While Arnaud et al. (2023) used LIME to interpret emergency department admission predictions from triage notes, their approach focuses on highlighting input features driving the model’s decision. But what if we could show clinicians not just why a prediction was made, but how changing certain elements of the input could flip that decision? Counterfactual explanations, popular in finance and legal XAI (see Adewumi et al., 2025), answer “what if?” questions: “What if this symptom wasn’t present?” or “What if the student had included a different argument in their essay?” This approach hasn’t been systematically explored in healthcare or education NLP. By generating actionable, human-centered counterfactuals, this research would empower practitioners to guide patient care plans or instructional interventions. The novelty lies in adapting counterfactual techniques from other domains for text-based, domain-specific data, addressing both transparency and actionable feedback—something current explainable NLP models seldom provide in these sensitive applications.

References:

1. Explainable NLP Model for Predicting Patient Admissions at Emergency Department Using Triage Notes. Émilien Arnaud, Mahmoud Elbattah, Pedro A. Moreno-Sánchez, Gilles Dequen, Daniel Aiham Ghazali (2023). BigData Congress [Services Society].
2. Democratizing public-impact algorithms: Advancing equitable and explainable AI systems for decision-making in U.S. health, finance, and education sectors. Farouk G. Adewumi, Chibuzor Njoku, Uchechukwu Okafor (2025). International Journal of Management & Entrepreneurship Research.