مُتنبئ مخصص لتفضيلات المرضى للحكم البديل في الرعاية الصحية: قابل للتنفيذ تقنيًا ومرغوب أخلاقيًا
A Personalized Patient Preference Predictor for Substituted Judgments in Healthcare: Technically Feasible and Ethically Desirable

المجلة: The American Journal of Bioethics، المجلد: 24، العدد: 7
DOI: https://doi.org/10.1080/15265161.2023.2296402
PMID: https://pubmed.ncbi.nlm.nih.gov/38226965
تاريخ النشر: 2024-01-16
المؤلف: Brian D. Earp وآخرون
الموضوع الرئيسي: التواصل بين المرضى ومقدمي الرعاية الصحية

نظرة عامة

في هذا القسم، يقترح المؤلفون نهجًا جديدًا للتنبؤ بتفضيلات علاج المرضى للأفراد غير القادرين، يسمى مُتنبئ تفضيلات المرضى المخصص (P4). يهدف هذا النظام إلى تعزيز دقة التنبؤات بالتفضيلات من خلال استخدام خوارزميات التعلم الآلي، وخاصة النماذج اللغوية الكبيرة (LLMs) المعدلة بشكل دقيق، المدربة على بيانات محددة لكل فرد بدلاً من بيانات السكان العامة. يجادل المؤلفون بأن الـ P4 لا يتمتع فقط بإمكانية تقديم تنبؤات أكثر دقة من مُتنبئ تفضيلات المرضى المقترح سابقًا (PPP) ولكنه أيضًا يحترم استقلالية المرضى من خلال عكس قيمهم الفريدة وأسباب تفضيلاتهم.

يحدد المؤلفون عدة خطوات حاسمة ضرورية لتطوير وتأكيد صحة الـ P4، بما في ذلك إنشاء نماذج أولية، وتقييم أنواع مختلفة من المعلومات للتعديل الدقيق، واستطلاعات الرأي العامة بشأن استخدام هذه التكنولوجيا في اتخاذ القرارات. يؤكدون على أهمية دمج الـ P4 في تخطيط الرعاية المتقدمة الحالي وأطر اتخاذ القرار بالنيابة، لضمان دعمه بدلاً من تقويض استقلالية المرضى. علاوة على ذلك، يسلطون الضوء على الحاجة إلى أن يوفر الـ P4 تنبؤات قابلة للتفسير، مما يسمح بالتدقيق المستقل من قبل المهنيين الصحيين وأفراد الأسرة. تتناول المناقشة أيضًا الآثار الأوسع للذكاء الاصطناعي في اتخاذ القرارات الأخلاقية، مقترحة أن الذكاء الاصطناعي يمكن أن يلعب دورًا كبيرًا في توضيح وتحدي القيم الفردية خارج السياقات الطبية.

مقدمة

تقدم مقدمة هذه الورقة البحثية دراسة حالة لامرأة تبلغ من العمر 26 عامًا، تُعرف باسم S، التي تعاني من سكتة قلبية أدت إلى إصابة دماغية شديدة وتشخيص بحالة نباتية مستمرة. مع عدم وجود توجيه مسبق، تواجه عائلتها معضلة بشأن استمرار العلاج الذي يحافظ على الحياة، مما يبرز التحديات التي يواجهها صناع القرار بالنيابة عند تحديد تفضيلات المريض في غياب توجيهات واضحة. تناقش الورقة الاعتماد على اتخاذ القرار بالنيابة في الممارسة السريرية، مع التأكيد على معيار “الحكم البديل”، الذي يهدف إلى احترام استقلالية المريض من خلال اتخاذ قرارات كما كان سيفعل المريض لو كان قادرًا. ومع ذلك، تشير إلى أن الوكلاء غالبًا ما يواجهون صعوبة في التنبؤ بدقة برغبات المريض ويعانون من ضغوط عاطفية كبيرة عند اتخاذ هذه القرارات الحاسمة.

لمعالجة هذه التحديات، يقترح المؤلفون تطوير “مُتنبئ تفضيلات المرضى” (PPP)، وهو خوارزمية قائمة على الكمبيوتر مصممة لاستنتاج تفضيلات المرضى بناءً على بيانات ديموغرافية وتفضيلات تم جمعها من استطلاع سكاني تمثيلي. يستند هذا النهج إلى أدلة تجريبية تشير إلى أن العوامل الديموغرافية يمكن أن تتنبأ بالخيارات الصحية. تقدم الورقة أيضًا تكيفًا جديدًا يسمى “مُتنبئ تفضيلات المرضى المخصص” (P4)، الذي يستخدم التعلم الآلي لإنشاء نموذج مخصص لتفضيلات المريض بناءً على بياناتهم الفردية. يهدف هذا النموذج إلى تعزيز دقة التنبؤ واحترام استقلالية المريض من خلال توفير رؤى في الوقت الحقيقي حول تفضيلات العلاج خلال الأزمات الصحية. يدعو المؤلفون إلى تطوير عاجل لمثل هذه المتنبئات، مؤكدين على الحاجة إلى التعاون بين الأخلاقيين والمهنيين الصحيين وخبراء الذكاء الاصطناعي والجمهور لمعالجة الاعتبارات الأخلاقية وتحديات التنفيذ.

مناقشة

يمثل نموذج الـ P4 المقترح تكرارًا مخصصًا لمُتنبئ تفضيلات المرضى (PPP)، مصممًا لاستخدام بيانات محددة لكل فرد لتعزيز دقة التنبؤ بتفضيلات العلاج للمرضى غير القادرين. على عكس الـ PPP، الذي يعتمد على بيانات ديموغرافية ومتوسطات جماعية، يهدف الـ P4 إلى دمج مجموعة متنوعة من المعلومات الشخصية—مثل الاتصالات المكتوبة، والسجلات الصحية الإلكترونية، والردود على سيناريوهات افتراضية—لتنبؤ كل من تفضيلات العلاج من الدرجة الأولى وتفضيلات الدرجة الثانية بشأن عمليات اتخاذ القرار. يُتصور أن يكون هذا النموذج نموذجًا لغويًا كبيرًا (LLM) مُعدلًا يتكيف مع الخصائص الفريدة للفرد، مما قد يحسن دقة التنبؤات مقارنة بالطرق الحالية.

يثير تنفيذ الـ P4 اعتبارات أخلاقية كبيرة، خاصة فيما يتعلق بالخصوصية والموافقة. من الضروري أن يعمل النموذج ضمن الأطر القانونية المعمول بها، مما يضمن استخدام بيانات المرضى بشكل أخلاقي ومع الموافقة المناسبة. يمكن أن يكون الـ P4 أداة مكملة لصناع القرار من العائلة، مما يخفف العبء عنهم مع احترام رغبات المرضى في المشاركة في رعايتهم. علاوة على ذلك، قد تعالج قدرة الـ P4 على التنبؤ بمجموعة أوسع من التفضيلات، بناءً على مجموعة بيانات أغنى، الاعتراضات القائمة على الاستقلالية التي تم طرحها ضد الـ PPP، حيث إنه مصمم لعكس قيم الفرد بدقة أكبر. بشكل عام، يحمل الـ P4 وعدًا لتعزيز اتخاذ القرار في البيئات الصحية حيث لا يستطيع المرضى التعبير عن تفضيلاتهم مباشرة.

القيود

تناقش قسم القيود التحديات العملية والمعرفية المرتبطة بنموذج الـ P4 لتنبؤ تفضيلات العلاج للمرضى غير القادرين. إحدى القضايا الرئيسية هي الصعوبة الجوهرية في تحديد ما يفضله المريض غير القادر بدقة، حيث لا يمكنهم التعبير عن تفضيلاتهم. بينما يمكن أن توفر السيناريوهات الافتراضية بعض الرؤى، إلا أنها قد لا تعكس بدقة القرارات الحقيقية بسبب التحيزات مثل الرغبة الاجتماعية. تمتد هذه القيود إلى جميع الطرق التنبؤية، بما في ذلك الوكلاء والتوجيهات المسبقة، مما يشير إلى أن اتخاذ القرار بالنيابة قد يكون قد وصل إلى حدوده التنبؤية. ومع ذلك، لا يحتاج الـ P4 إلى تحقيق يقين مطلق؛ بل يحتاج فقط إلى التفوق على القدرات التنبؤية الحالية للوكلاء البشريين.

بالإضافة إلى ذلك، يسلط القسم الضوء على القيود المحتملة لاستخدام النماذج اللغوية الكبيرة (LLMs) في فهم تفضيلات المرضى، خاصة فيما يتعلق بالوصول الاستبطاني إلى الحالات العاطفية. بينما يمكن أن تحلل LLMs البيانات النصية لاستنتاج التفضيلات، إلا أنها تفتقر إلى الفهم الدقيق الذي يمتلكه أفراد الأسرة من خلال التجارب المشتركة. قد يواجه تنفيذ الـ P4 أيضًا تحديات مشابهة لتلك التي تم مواجهتها مع التوجيهات المسبقة، مثل الوصول والامتثال من قبل المهنيين الصحيين. علاوة على ذلك، هناك خطر من الاعتماد المفرط على مخرجات الـ P4 بسبب العرض الواثق للمعلومات من قبل LLMs، والتي قد لا تعكس بدقة عدم اليقين الكامن في التنبؤ بتفضيلات الأفراد. إن معالجة هذه القضايا، بما في ذلك تطوير طرق للتعبير عن مستويات الثقة في التنبؤات، أمر بالغ الأهمية للتطبيق السريري للـ P4s.

Journal: The American Journal of Bioethics, Volume: 24, Issue: 7
DOI: https://doi.org/10.1080/15265161.2023.2296402
PMID: https://pubmed.ncbi.nlm.nih.gov/38226965
Publication Date: 2024-01-16
Author(s): Brian D. Earp et al.
Primary Topic: Patient-Provider Communication in Healthcare

Overview

In this section, the authors propose a novel approach to predicting patient treatment preferences for incapacitated individuals, termed the Personalized Patient Preference Predictor (P4). This system aims to enhance the accuracy of preference predictions by utilizing machine learning algorithms, specifically fine-tuned large language models (LLMs), trained on individual-specific data rather than general population data. The authors argue that the P4 not only has the potential to yield more accurate predictions than the previously proposed Patient Preference Predictor (PPP) but also respects the autonomy of patients by reflecting their unique values and reasons for preferences.

The authors outline several critical steps necessary for the development and validation of the P4, including the creation of prototypes, the evaluation of different types of information for fine-tuning, and public opinion surveys regarding the use of such technology in decision-making. They emphasize the importance of integrating the P4 into existing advance care planning and surrogate decision-making frameworks, ensuring that it supports rather than undermines patient autonomy. Furthermore, they highlight the need for the P4 to provide explainable predictions, allowing for independent scrutiny by healthcare professionals and family members. The discussion also touches on broader implications of AI in moral decision-making, suggesting that AI could play a significant role in elucidating and challenging individual values beyond medical contexts.

Introduction

The introduction of this research paper presents a case study of a 26-year-old woman, referred to as S, who suffers a cardiac arrest leading to severe brain injury and a diagnosis of persistent vegetative state. With no advance directive in place, her family faces a dilemma regarding the continuation of life-sustaining treatment, highlighting the challenges surrogate decision-makers encounter when determining a patient’s preferences in the absence of clear guidance. The paper discusses the reliance on surrogate decision-making in clinical practice, emphasizing the “substituted judgment” standard, which aims to respect the patient’s autonomy by making decisions as the patient would have if competent. However, it notes that surrogates often struggle to accurately predict the patient’s wishes and experience significant emotional distress in making these critical decisions.

To address these challenges, the authors propose the development of a “Patient Preference Predictor” (PPP), a computer-based algorithm designed to infer patient preferences based on demographic data and preferences collected from a representative population survey. This approach is grounded in empirical evidence that demographic factors can predict healthcare choices. The paper further introduces a novel adaptation called the “Personalized Patient Preference Predictor” (P4), which utilizes machine learning to create a personalized model of a patient’s preferences based on their individual data. This model aims to enhance predictive accuracy and respect for patient autonomy by providing real-time insights into treatment preferences during healthcare crises. The authors advocate for the urgent development of such preference predictors, emphasizing the need for collaboration among ethicists, healthcare professionals, AI experts, and the public to address ethical considerations and implementation challenges.

Discussion

The proposed P4 model represents a personalized iteration of the Patient Preference Predictor (PPP), designed to utilize individual-specific data to enhance the accuracy of predicting treatment preferences for incapacitated patients. Unlike the PPP, which relies on demographic data and group averages, the P4 aims to incorporate a variety of personal information—such as written communications, electronic health records, and responses to hypothetical scenarios—to predict both first-order treatment preferences and second-order preferences regarding decision-making processes. This model is envisioned as a fine-tuned large language model (LLM) that adapts to the unique characteristics of an individual, thereby potentially improving the accuracy of predictions compared to existing methods.

The implementation of the P4 raises significant ethical considerations, particularly concerning privacy and consent. It is crucial that the model operates within established legal frameworks, ensuring that patient data is used ethically and with appropriate consent. The P4 could serve as a supplementary tool to family decision-makers, alleviating the burden on them while respecting patients’ wishes for involvement in their care. Furthermore, the P4’s ability to predict a broader range of preferences, based on a richer dataset, may address autonomy-based objections that have been raised against the PPP, as it is designed to reflect the individual’s values more accurately. Overall, the P4 holds promise for enhancing decision-making in healthcare settings where patients are unable to express their preferences directly.

Limitations

The section on limitations discusses both practical and epistemic challenges associated with the P4 model for predicting the treatment preferences of incapacitated patients. A primary concern is the inherent difficulty in accurately determining what an incapacitated patient would prefer, as they cannot express their preferences. While hypothetical scenarios can provide some insight, they may not accurately reflect real-life decisions due to biases such as social desirability. This limitation extends to all predictive methods, including surrogates and advance directives, suggesting that surrogate decision-making may have reached its predictive limits. However, the P4 does not need to achieve absolute certainty; it merely needs to outperform the current predictive capabilities of human surrogates.

Additionally, the section highlights the potential limitations of using large language models (LLMs) in understanding patient preferences, particularly regarding introspective access to affective states. While LLMs can analyze textual data to infer preferences, they lack the nuanced understanding that family members possess through shared experiences. The implementation of P4s may also face challenges similar to those encountered with advance directives, such as accessibility and adherence by healthcare professionals. Furthermore, there is a risk of over-reliance on P4 outputs due to the confident presentation of information by LLMs, which may not accurately reflect the uncertainty inherent in predicting individual preferences. Addressing these issues, including the development of methods to express confidence levels in predictions, is crucial for the clinical application of P4s.