Patient preference predictors revisited: technically feasible, ethically desirable, yet must be clinically relevant

Critical Care Ferrario et al. Critical Care (2025) 29:437 https://doi.org/10.1186/s13054-025-05637-8 Open Access PERSPECTIVE Patient preference predictors revisited: technically feasible, ethically desirable, yet must be clinically relevant Andrea Ferrario1,2,3, Beatrix Göcking1,4, Giovanna Brandi4,5, Emanuela Keller4,6 and Nikola Biller-Andorno1* Abstract Although goal-concordant care is central to patient-centered medicine, determining treatment preferences for incapacitated patients remains a challenge. Nearly two decades ago, algorithms were proposed to estimate the most likely treatment preferences in the absence of advance directives, aiming to support surrogate decisionmaking. This idea has evolved into a race toward increasingly complex models, driven by the assumption that expanding data collection and refining predictive methods will yield more accurate approximations of patients’ unknown treatment preferences. Despite extensive debate on the epistemic, ethical, and clinical challenges of these algorithms, none have been successfully implemented in clinical practice. We contend that this failure does not stem from any of these challenges but, rather, from conceptualizing these models simply as technically sophisticated replicas of advance directives, abstracting a few high-level treatment preferences across all clinical contexts while ignoring setting-specific, temporal, and relational factors. The barrier to the implementation of these models is fundamentally a technology design problem that requires a novel design perspective to ensure their clinical relevance. We discuss this perspective using neuro-intensive care as a case study and examine how algorithmic models could support time-sensitive decision-making for patients with severe acute brain injury. The success of patient preference predictions depends not only on their being technically feasible and ethically promising but on ensuring clinical relevance. Keywords Goal-concordant care, Decision-making, Goal of care preferences, Patient preference, Artificial intelligence, Machine learning, Brain injuries, Neurological intensive care unit, Critical care *Correspondence: Nikola Biller-Andorno 1 Institute of Biomedical Ethics and History of Medicine, University of Zurich, Zurich, Switzerland 2 ETH Zurich, Zurich, Switzerland 3 University of Applied Sciences and Arts of Southern Switzerland (SUPSI), Dalle Molle Institute for Artificial Intelligence (IDSIA), Lugano, Switzerland 4 Institute of Intensive Care Medicine, University Hospital Zurich, Zurich, Switzerland 5 University of Zurich, Zurich, Switzerland 6 Department of Neurosurgery, University Hospital Zurich, Zurich, Switzerland Background Goal-concordant care—ensuring that medical treatments align with a patient’s values and preferences—is a key principle of patient-centered medicine, particularly in high-stakes, preference-sensitive decisions [1]. However, determining preferences for incapacitated patients remains a persistent challenge, involving notable epistemic and normative challenges. The absence of advance directives (ADs)—here meant as a document used to record living wills—and the limitations of surrogate decision-making have led different authors to propose the use of algorithms—namely, a set of rules run by computer © The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creati vecommons.org/licenses/by-nc-nd/4.0/. Ferrario et al. Critical Care (2025) 29:437 systems to predict an output of interest given input data—to estimate patient’s most likely treatment preferences in case of their incapacitation [2, 3]. Since their proposal, these algorithms, originally called patient preference predictor (PPP or P3) have risen to a broader technological competition—a ‘race for Px’— where Px denotes increasingly personalized predictors, ranging from the original P3 to the more recent P3.5 and P4, each reflecting a higher degree of predictive personalization [4, 5]. In fact, the underlying hypothesis driving the race is that expanding the granularity of data and implementing methods from artificial intelligence (AI) will yield more accurate approximations of patient’s true, albeit often unknown, treatment preferences. As a result, while early algorithms suggested population-based rules to predict category-level treatment preferences, more advanced approaches suggest considering machine learning incorporating large-scale data integration and natural language processing to compute individualized preferences [4, 5]. Our deliberate use of the verb ‘suggest’ underscores the theoretical nature of this endeavor. Despite extensive discourse on the epistemic and ethical challenges posed by Px, which span a broad and growing body of literature [2, 5–12], after 18 years since its original proposal, no Px has been adopted in real-world clinical environments. The original promise of the PPP—to enhance goal-concordant care by aiding clinicians and surrogate decision-makers in making more informed treatment decisions—remains unfulfilled. In this paper, we argue that this implementation gap descends from a conceptual challenge that we name ‘substitutive design fallacy’ affecting the Px proposition since its inception. Under this fallacy, Px models are simply conceptualized as technically complex replicas of ADs that compute few, high-level treatment preferences and make them available across clinical scenarios, regardless of care setting, decision-making timelines, or relational dynamics among human participants. While this conceptualization may have initially served as a useful rhetorical instrument to promote the idea of algorithmically enhanced preference elicitation [2, 3], it has condemned Px to replicate the well-known limitations of ADs (e.g., context insensitivity), neglecting the procedural and social dimensions required to meaningfully integrate Px into clinical decision-making that are key to the design of this technology. Thus, we contend that the barrier to the clinical implementation of Px models is fundamentally a (...truncated)