Patient-Centred Explainability in IVF Outcome Prediction

Patients need explanations beyond technical model features: The study found that standard explainable AI approaches focusing on which input variables influence predictions are insufficient for IVF patients. Users require contextual information about how the model works, its limitations, and how their specific situation relates to the data it was trained on.

Data shifts and model exclusions significantly impact patient trust: Patient feedback revealed that concerns about whether the model applies to their particular circumstances—such as changes in medical practices over time or exclusion of certain patient groups from training data—were major drivers of distrust, yet these issues are often overlooked in typical AI explainability research.

Patients develop complex mental models and want personalized dialogue: Rather than one-size-fits-all explanations, the research showed that patients gathering extensive information about their fertility treatment benefit from interactive, dialogue-based interfaces that can address their individual questions and concerns about prediction reliability.

Current explainable AI practices fall short in high-stakes healthcare: The paper highlights a gap between how explainability is typically designed in academic AI research and what patients actually need to make informed decisions about their medical care, suggesting that healthcare contexts require fundamentally different approaches than other domains.

Patient-centered design requires moving beyond model-centric thinking: The findings call for explainability research to expand beyond explaining model mechanics to include broader questions about data representativeness, model scope, and real-world applicability that matter most to users making consequential life decisions.