Speaking for Tomorrow: Can AI Help Us Plan for End-of-Life Care?

Author: Denis Avetisyan

Researchers are investigating how AI agents can personalize and preserve individual preferences in advance care planning, potentially bridging the gap between current wishes and future medical decisions.

This review explores the potential of AI-powered decision support tools for eliciting, maintaining, and enacting values in advance care planning, focusing on human-AI interaction and the role of autonomous agents.

As populations age and healthcare decisions become increasingly complex, individuals face challenges articulating deeply personal values for future care. This research, ‘Words to Describe What I’m Feeling: Exploring the Potential of AI Agents for High Subjectivity Decisions in Advance Care Planning’, investigates how AI agents might serve as personalized advocates, supporting individuals through the advance care planning process. Through an experience prototype and user workshops, we found that AI can foster ‘mutual intelligibility’ regarding preferences, offering a novel balance between autonomous support and human oversight. Could this approach redefine the role of AI in healthcare, moving beyond simple decision support towards genuine partnership in values-based care?

The Weight of Tomorrow: Confronting Uncertainty in Advance Care Planning

The challenging nature of advance care planning frequently results in delays or incomplete documentation, stemming from the profound emotional and cognitive demands it places on individuals. Confronting mortality and imagining future health states requires significant psychological processing, often triggering anxiety, fear, and a desire to avoid difficult conversations. This emotional weight, coupled with the cognitive effort of anticipating various medical scenarios and articulating personal values, can be overwhelming, leading many to postpone or avoid the process altogether. Furthermore, individuals may struggle with the abstract nature of these decisions, finding it difficult to envision themselves in future compromised health states or to accurately predict their preferences under such circumstances, thereby hindering effective planning.

Advance care planning isn’t simply a matter of making choices; it’s a navigation through profound uncertainty. Individuals attempting to articulate future healthcare preferences invariably face incomplete knowledge – both about potential medical scenarios and, crucially, about their own evolving values. Predicting how one will weigh quality of life against longevity when facing a serious illness is incredibly difficult, and personal priorities can shift under duress. This epistemic uncertainty means that even well-intentioned plans are provisional, reflecting a ‘best guess’ at a future that remains fundamentally unknowable. Recognizing this inherent limitation is vital, as it encourages a more flexible and ongoing approach to advance care planning, one that prioritizes open communication and allows for adjustments as circumstances – and personal understanding – change.

Conventional advance care planning often relies on checklists and standardized forms, proving inadequate for capturing the subtleties of individual wishes regarding future medical care. These methods frequently presume a level of foresight regarding potential health scenarios and personal values that individuals simply do not possess, leading to choices based on hypothetical situations rather than deeply considered preferences. This disconnect between expressed wishes and authentic values can leave patients vulnerable to receiving care inconsistent with what truly matters to them, effectively diminishing their autonomy even within the planning process. Consequently, the very purpose of advance care planning – to ensure respect for personal agency – is compromised when nuanced beliefs and evolving priorities are not accurately represented or adequately addressed.

The Advocate Within: Introducing the ACPAgent

The ACPAgent functions as an experiential prototype intended to represent a user’s advocate during Advance Care Planning (ACP) processes. This is achieved through a system designed to progressively model individual user preferences. The agent doesn’t simply present static options; instead, it actively learns from user interactions, refining its understanding of values and priorities over time. This adaptive capability allows the ACPAgent to tailor subsequent recommendations and responses, effectively functioning as a personalized proxy that anticipates and reflects the user’s evolving needs within the ACP decision-making framework.

The ACPAgent utilizes Large Language Models (LLMs) to facilitate conversational interactions with users regarding Advance Care Planning (ACP). These interactions are designed to identify and articulate user values relevant to future healthcare decisions. The LLM processes user responses to understand preferences concerning quality of life, acceptable trade-offs between different health outcomes, and desired levels of medical intervention. Furthermore, the agent can present hypothetical future medical scenarios, prompting users to consider how their values would apply in specific contexts and enabling exploration of potential care pathways. This process allows the agent to build a personalized profile reflecting the user’s individual priorities and beliefs, informing subsequent ACP recommendations.

Generative Ghost Design, as implemented in the ACPAgent, focuses on proactive adaptation to user needs through continuous learning. The system doesn’t simply react to explicit requests; instead, it builds an internal model of user preferences based on conversational history and expressed values. This model is then used to anticipate potential future needs and proactively offer relevant information or support, creating a more responsive and personalized experience. The ‘ghost’ aspect refers to the agent’s ability to subtly reflect back the user’s own values and concerns, fostering a sense of understanding and trust without overtly stating them. This anticipatory capability is achieved through the agent’s underlying Large Language Model, which identifies patterns in user input and generates appropriate responses based on the evolving user profile.

Mapping the Possible: Simulating Futures with Scenario-Based Reasoning

The ACPAgent utilizes scenario-based reasoning by constructing hypothetical medical cases, detailing patient conditions and potential interventions. These scenarios are presented to users, who are then asked to specify their preferred treatment approaches for the given situation. The system does not offer advice, but rather functions by eliciting user preferences through direct questioning within the context of the simulated case. This process allows the agent to map user values to specific medical choices, forming a profile of the individual’s priorities regarding care and treatment options.

The ACPAgent facilitates exploration of treatment outcomes within a simulated environment, allowing users to assess the potential consequences of their decisions without real-world risk. This is particularly relevant for critical care scenarios, such as resuscitation, where the implications of choices are immediate and significant. By presenting hypothetical patient cases and corresponding treatment options, the agent enables users to evaluate different courses of action and observe projected results, aiding in the understanding of complex medical trade-offs and promoting proactive consideration of preferences before actual clinical events occur.

The ACPAgent is designed as a decision support tool and not a replacement for clinical judgment. Its function is to present carefully constructed medical scenarios, prompting users to articulate preferences and consider potential outcomes, thereby enhancing clarity in complex situations. Recent studies indicate a significant impact of these presented scenarios on user decision-making; data showed that 76% of participants altered their initial preferences based on the information and considerations prompted by the agent’s simulations, demonstrating its ability to empower more informed choices.

Amplifying the Voice: Balancing Autonomy with the Agent’s Role

The ACPAgent is fundamentally designed to prioritize human direction in healthcare decisions. This isn’t an autonomous system seeking to dictate care, but rather a tool built to amplify patient preferences and ensure they remain central throughout the process. The agent operates on the principle that individuals should retain complete authority over their care plans, with the AI serving as a supportive resource for documenting, clarifying, and ultimately enacting those wishes. By placing the user firmly in control, the design aims to alleviate anxieties surrounding AI in healthcare and foster a collaborative relationship where technology enhances, rather than overrides, personal agency. This focus on user empowerment is a core tenet of the system’s functionality, ensuring alignment between technological assistance and individual autonomy.

The ACPAgent extends beyond simply recording preferences; it actively functions as an advocacy tool for patients. This means the system is designed to articulate a patient’s wishes directly to healthcare providers and, crucially, to legal proxies should the patient lose the capacity to communicate independently. The agent doesn’t merely store directives, it represents them, ensuring that even in complex medical scenarios or legal proceedings, the patient’s previously expressed values and care goals remain central to decision-making. This proactive representation aims to bridge potential communication gaps and safeguard patient autonomy when individuals are most vulnerable, effectively giving voice to preferences that might otherwise be overlooked or misinterpreted.

A carefully considered framework guides the development of this AI agent for advance care planning, prioritizing both its autonomous function and complete transparency in its operations. Recognizing that trust is paramount, the system is designed to clearly articulate the reasoning behind its recommendations, ensuring users understand how decisions are being supported. This approach was rigorously tested with fifteen participants who provided valuable feedback on the agent’s effectiveness, focusing on its ability to accurately reflect preferences and facilitate meaningful conversations about end-of-life care. The results of this study are informing iterative improvements, with the ultimate goal of responsible implementation and widespread adoption of AI as a supportive tool in advance care planning.

The Future of Choice: Towards AI-Enhanced, Personalized Healthcare

The ACPAgent exemplifies how artificial intelligence can meaningfully enhance advance care planning (ACP), a process often fraught with emotional complexity and logistical challenges. This innovative system doesn’t replace human interaction, but rather augments it by providing personalized recommendations and facilitating proactive dialogue between individuals and their healthcare providers. Through sophisticated algorithms, the agent analyzes patient values, medical conditions, and treatment options to generate insights that promote informed decision-making. This capability has the potential to streamline the ACP process, ensuring that each person’s wishes are clearly documented and respected, ultimately improving the quality and efficiency of care delivery while honoring patient autonomy.

The capacity for individuals to articulate their values and preferences regarding healthcare is fundamentally strengthened through proactive, informed dialogue, and recent advancements aim to facilitate precisely that. By guiding conversations and presenting personalized information, these systems empower patients to actively participate in defining their care, ensuring alignment with deeply held beliefs even when facing complex medical decisions. This isn’t merely about documenting wishes; it’s about fostering a collaborative process where individual values are central to the healthcare experience, promoting dignity and respect throughout the journey and ultimately leading to more meaningful and patient-centered outcomes.

Ongoing development centers on enhancing the ACPAgent’s sophistication and broadening its utility within personalized healthcare landscapes. Current research aims to improve the agent’s ability to navigate complex medical scenarios and tailor recommendations with even greater precision. Crucially, initial trials demonstrate a high degree of concordance between the agent’s suggestions and participant preferences-an 86.7% agreement rate-suggesting a strong potential for ACPAgent to support individuals facing critical, high-risk decisions. This level of alignment indicates the agent doesn’t simply offer data, but facilitates informed choices that genuinely reflect patient values, paving the way for its application in areas beyond advance care planning, such as treatment selection and end-of-life care coordination.

The research posits a compelling challenge to conventional notions of healthcare autonomy, suggesting AI agents can serve as personalized proxies in advance care planning. This aligns perfectly with Grace Hopper’s assertion: “It’s easier to ask forgiveness than it is to get permission.” The study intentionally probes the boundaries of established practice, advocating for a system where AI proactively elicits and maintains values, even when those values deviate from pre-defined norms. Just as Hopper championed a mindset of experimentation over strict adherence to rules, this work suggests that a degree of ‘forgiveness’-allowing the AI to explore and articulate nuanced preferences-is necessary to truly empower individuals in shaping their future care. The core idea revolves around actively testing the system to understand its limitations and potential, mirroring Hopper’s relentless drive to dissect and improve existing technology.

Beyond Preference: Charting the Course

The utility of an AI proxy in advance care planning isn’t simply about faithfully recording stated preferences-it’s about exposing the inherent fragility of those statements. This work reveals a critical vulnerability: the assumption that a present self can reliably predict the values of a future, potentially radically altered, self. The next iteration must actively stress-test elicited values, simulating scenarios designed to reveal internal inconsistencies or unacknowledged trade-offs. Only through such deliberate attempts at ‘breaking’ the system can one begin to understand the true contours of an individual’s deeply held beliefs.

A genuine exploit of comprehension will require moving beyond declarative statements of ‘what I want’ to a dynamic model of why those wants exist. The challenge isn’t merely to build an AI that echoes preferences, but one that can reason about the underlying motivations – the ethical frameworks, the emotional weights – that inform them. This demands integrating insights from behavioral economics and moral psychology, turning the AI from a passive recorder into an active, albeit synthetic, moral interlocutor.

Ultimately, the success of this field won’t be measured by how well AI can mimic human decision-making, but by how effectively it can reveal the inherent complexities and contradictions within it. The goal isn’t autonomy, but augmented understanding-a mirror held up to the self, forcing a reckoning with the often-unarticulated principles that govern our lives, and ultimately, our deaths.

Original article: https://arxiv.org/pdf/2512.11276.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/