A Personalized Patient Preference Predictor for Substituted Judgments in Healthcare: Technically Feasible and Ethically Desirable

Brian D. Earp, Sebastian Porsdam Mann, Jemima Allen, Sabine Salloch, Vynn Suren, Karin Jongsma, Matthias Braun, Dominic Wilkinson, Walter Sinnott‐Armstrong, Annette Rid, David Wendler, Julian SavulescuThe American Journal of Bioethics 2024 · DOI

TL;DR

This paper proposes a "Personalized Patient Preference Predictor" (P4) that uses fine-tuned Large Language Models (LLMs) to infer the medical treatment preferences of incapacitated individuals. Unlike previous approaches relying on population data, the P4 leverages person-specific text data to potentially improve accuracy and better respect patient autonomy, addressing ethical concerns about substituted judgment.

What Problem Does This Paper Address?

When individuals lose the capacity to make their own medical decisions, their loved ones or healthcare providers must make "substituted judgments" about their care. This is a challenging and often distressing task, as surrogates may struggle to accurately guess what the patient would have wanted. While advance directives can help, they are often incomplete or unavailable.

A previous proposal, the "Patient Preference Predictor" (PPP), suggested using algorithms trained on population-level demographic data to predict individual preferences. However, critics argued that this approach fails to respect patient autonomy because it relies on generalized data that may not reflect an individual's unique values or reasons for their preferences. This paper tackles this ethical and practical gap by proposing a more individualized approach.

Methods

The core innovation is the Personalized Patient Preference Predictor (P4). Unlike the PPP, which uses broad demographic correlations, the P4 is designed to be fine-tuned on person-specific data. This fine-tuning process, leveraging recent advances in machine learning and LLMs, allows the model to learn from an individual's own writings, communications, and other digital footprints.

The P4 aims to predict not only first-order treatment preferences but also second-order "process" preferences (e.g., how decisions should be made) and the confidence level associated with these predictions. The paper outlines several potential versions of the P4 based on different types of input data, ranging from personal emails and social media posts to more structured value-eliciting experiments.

Datasets

The paper does not describe specific datasets used for training or evaluation of a P4 prototype. Instead, it discusses the types of person-specific data that could be used for fine-tuning an LLM to create a P4. These include:

  • Personal communications (emails, blog posts, social media)
  • Electronic health records (EHRs)
  • Responses to medical questionnaires or value-eliciting choice experiments
  • Speech-to-text transcriptions of conversations
  • Purchasing and browsing histories

The paper emphasizes that the development of such P4s is still theoretical, and empirical work is needed to establish the effectiveness of these data sources.

Key Insights and Ideas

  • Personalization is Key for Autonomy: The central argument is that using individual-specific data, rather than population averages, is crucial for respecting a patient's autonomy in substituted judgment.
  • LLMs as "Digital Psychological Twins": The P4 concept envisions LLMs fine-tuned on personal data as creating a digital representation that can "mimic" the patient's likely preferences.
  • Beyond First-Order Preferences: The P4 aims to capture not just what a patient would want, but also how they would want decisions to be made about them, addressing "process preferences."
  • Addressing Criticisms of PPP: The P4 is presented as a direct response to autonomy-based objections against the PPP, particularly the concern that generalized data fails to respect individual values.
  • Technical Feasibility: The paper highlights recent AI advancements, particularly in LLM fine-tuning, suggesting that a P4 is technically achievable.

Key Results

As the P4 is a theoretical proposal, no empirical results or quantitative accuracy metrics are presented. The paper focuses on:

  • Argument for Technical Feasibility: Citing existing LLM capabilities in inferring non-medical preferences and adapting to individual styles, the authors argue that inferring medical preferences is plausible.
  • Ethical Argumentation: The paper argues that the P4, by using person-specific data, is ethically superior to the PPP in respecting patient autonomy.
  • Comparison to Human Surrogates: The authors suggest that a P4 could potentially be more accurate than human surrogates, who are often only slightly better than chance in predicting patient preferences.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Theoretical Nature: The P4 is a hypothetical concept, and empirical studies are needed to build and test prototypes, comparing their accuracy to PPPs and human surrogates.
  • Data Privacy and Consent: Obtaining consent for the use of personal data for fine-tuning is a significant ethical and practical challenge. The paper suggests solutions like locally stored LLMs and privacy-preserving services.
  • Explainability: While LLMs can infer preferences, making their predictions explainable (i.e., citing the reasons behind them) is crucial for trust and scrutiny. This is an active area of AI research.
  • Confidence Intervals: LLMs can present confident-sounding outputs regardless of actual certainty. Developing methods for LLMs to express their confidence levels is necessary to prevent over-reliance.
  • Integration into Practice: How the P4 would be integrated into existing advance care planning and shared decision-making processes needs further consideration.
  • Language and Cultural Specificity: The current capabilities of LLMs are often language-dependent, and research is needed to ensure P4s work across different languages and cultural contexts.
  • Uptake and Implementation Barriers: Similar to advance directives, practical barriers to uptake and adherence by healthcare professionals are anticipated.

Summary updated

Abstract

When making substituted judgments for incapacitated patients, surrogates often struggle to guess what the patient would want if they had capacity. Surrogates may also agonize over having the (sole) responsibility of making such a determination. To address such concerns, a Patient Preference Predictor (PPP) has been proposed that would use an algorithm to infer the treatment preferences of individual patients from population-level data about the known preferences of people with similar demographic characteristics. However, critics have suggested that even if such a PPP were more accurate, on average, than human surrogates in identifying patient preferences, the proposed algorithm would nevertheless fail to respect the patient's (former) autonomy since it draws on the 'wrong' kind of data: namely, data that are not specific to the individual patient and which therefore may not reflect their actual values, or their reasons for having the preferences they do. Taking such criticisms on board, we here propose a new approach: the <i>Personalized</i> Patient Preference Predictor (P4). The P4 is based on recent advances in machine learning, which allow technologies including large language models to be more cheaply and efficiently 'fine-tuned' on person-specific data. The P4, unlike the PPP, would be able to infer an individual patient's preferences from material (e.g., prior treatment decisions) that is in fact specific to them. Thus, we argue, in addition to being potentially more accurate at the individual level than the previously proposed PPP, the predictions of a P4 would also more directly reflect each patient's own reasons and values. In this article, we review recent discoveries in artificial intelligence research that suggest a P4 is technically feasible, and argue that, if it is developed and appropriately deployed, it should assuage some of the main autonomy-based concerns of critics of the original PPP. We then consider various objections to our proposal and offer some tentative replies.
BibTeX
@article{earp2024personalized,
  author = {Brian D. Earp and Sebastian Porsdam Mann and Jemima Allen and Sabine Salloch and Vynn Suren and Karin Jongsma and Matthias Braun and Dominic Wilkinson and Walter Sinnott‐Armstrong and Annette Rid and David Wendler and Julian Savulescu},
  title = {{A Personalized Patient Preference Predictor for Substituted Judgments in Healthcare: Technically Feasible and Ethically Desirable}},
  journal = {The American Journal of Bioethics},
  year = {2024},
  doi = {10.1080/15265161.2023.2296402},
}
Figure