text
stringlengths 1
1.75k
|
|---|
A patient-reported outcome measure (PROM) is a standardized tool for capturing PROs: a survey, instrument, scale, or single-item measure used to assess particular PROs, such as symptoms, behaviors, or functional abilities, as perceived by the individual, obtained by directly asking the individual to self-report. PROMs serve as a means to assess patient-centered outcomes, guide clinical decision making, evaluate treatment effectiveness, and inform healthcare policy. |
PROMs vary in their specificity and adaptability. Types of PROM include generic instruments that can be applied across different health conditions or populations, and condition-specific measures designed to assess a particular aspect of a disease or condition. |
Examples of commonly used PROMs include PROMIS (Patient-Reported Outcomes Measurement Information System, a set of tools for measuring physical, mental, and social health, the generic EQ-5D and SF-36 surveys (used to measure general physical and mental wellbeing along with health-related quality of life (HRQoL)), or the condition-specific PRO-CTCAE (Patient-Reported Outcomes Version of the Common Terminology Criteria for Adverse Events, used in Cancer treatment clinical trials to monitor treatment-related symptoms). |
A specific category of PROM, the patient-experience measure (PREM), records patient perspectives on the quality of care, service, and treatments they receive, often in the form of a patient satisfaction survey. Together, PROMs and PREMs contribute to an understanding of the value of healthcare received by a patient over time. |
PROMs can be found as paper surveys, web-based tools, standalone apps on a phone or tablet, or as part of a hospital’s patient portal. The integration of PROMs with Electronic Health Records (EHRs) has become increasingly common, enabling healthcare providers to access and analyze PROM data alongside other clinical information, facilitating a more holistic view of patients' health status. |
What is artificial intelligence? |
Artificial Intelligence (AI) is a widely used term to describe any tool or system that simulates, augments, or automates the way people make sense of the world. In the context of healthcare, AI encompasses a wide range of techniques, from image classification to predictive modeling, chat agents, decision support, and text analysis. |
In contrast with traditional statistical methods which begin by applying a priori models to understand the relationship between selected variables, researchers use AI to infer patterns from data and to give shape to models that represent both explicit and implicit features in the data. |
Natural Language Processing (NLP) is a set of techniques that automate understanding of what humans mean when they use language to express themselves to one another. Patients’ own descriptive language—how they express qualities of their experience to other people—is a source of ground truth for PROs. Because of this, NLP is at the center of many AI methods used with PROMs. While NLP makes use of both AI and non-AI tools, the focus of this review is on AI-based NLP approaches, where learning is automated, and understanding proceeds directly from raw text rather than a traditional statistical approach, where understanding is derived using rule-based systems. |
Why are PROMs important? |
There is a limit to what objective measurements in healthcare can do. PROMs, as subjective, unmediated reports shared by patients, fill in the gaps that other methods cannot assess, such as “pain levels, patient experience, motivation, human factors, [...] and health priorities." |
* PROMs are patient-centered by design: they aim to measure the outcomes that are most meaningful to patients, and to make sure that patient experiences, symptoms, and quality of life are taken seriously in treatment decisions. |
* PROMs provide opportunities for open dialogue between healthcare providers and patients, promoting shared decision-making and better care coordination. |
* When implemented successfully, PROMs invite patients to become active participants in their healthcare journey rather than passive recipients of treatment. This can help build trust between patients and providers, especially among populations who have experienced mistrust or neglect from the healthcare system in the past. |
* Using PROMs can reduce healthcare costs by identifying early signs of disease progression or treatment failure, allowing for timely adjustments and early interventions. |
The literature covered in this review is not limited to studies using specifically identified PROM instruments. The scope is purposefully expansive in its inclusion of other forms of patient-generated health data (PGHD), such as unstructured data sourced from electronic health records, apps, fitness trackers and wearable devices, social media, and other settings. We use data both actively and passively to express ourselves, and to represent our experience. Research has shown that patient-reported outcomes can be inferred from unstructured clinical notes and social media using AI techniques such as natural language processing (NLP). It has also been suggested that passively collecting PGHD from apps and devices provides meaningful context and validation for PROMs while reducing the burden on patients that is often associated with actively completing PROM questionnaires. Moreover, the introduction of AI analytical methods has been indicated as opening the way for integrating diverse and unstructured PGHD with validated PROMs. |
PROMs, as media, are changing, interdependent objects. The instruments will change, and so will the tools and frameworks for validating them. This review seeks to identify the conditions of a possible future in which researchers and clinicians continually learn how best to listen to patients’ experiences. |
Methods: |
Search terms |
Four databases were searched for literature published in English in the past five years (2020/1/1—2024/4/10) for which the terms “patient-reported” and/or “patient-generated” appeared in the same context with “artificial intelligence” and/or “machine learning” and/or “natural language” in the title, abstract, and/or keywords. |
Database: |
Query: |
Web of science |
AB=("patient-reported" OR "patient-generated") AND ("artificial intelligence" OR "machine learning" OR "natural language") |
n=403 |
Scopus |
TITLE-ABS-KEY ( ( "patient-reported" OR "patient-generated" ) AND ( "artificial intelligence" OR "machine learning" OR “natural language”)) |
n=857 |
PubMed |
((patient-reported[Title/Abstract]) OR (patient-generated[Title/Abstract])) AND ((artificial intelligence[Title/Abstract]) OR (machine learning[Title/Abstract]) OR (natural language[Title/Abstract])) |
n=510 |
ArXiv |
patient-reported outcomes (all fields) |
AND artificial intelligence (all fields) |
n=24 |
Total search results: |
n=1795 |
Remove duplicates |
n=934 |
Filter out references that do not have the intersection of patient-generated health data and artificial intelligence as an objective focus or outcome (hand-review of abstract, title, and keywords) |
n=308 |
Remove references with no full-text paper available: |
n=15 |
Total references included |
n=537 |
Full-text analysis |
The widespread availability of large language models (LLM), with capacity for summarization and topic modeling that far exceeds earlier statistical approaches, has prompted interest in the (partial) automation of systematic literature reviews. It has been shown that automation using LLMs works best when applied to full-text analysis, a task well-served by a tendency to produce generalized variations and transformations of source data. For the Results section of this review, AI analysis was used to synthesize a summary of the objectives, questions, ongoing concerns and imperatives related to the design of AI-PROMs, according to the available literature. |
Nomic AI’s GPT4All library was selected because of its stable and well-maintained open source codebase, its integration of multiple state-of-the-art large language models (LLMs) from different developers, its active community of contributors, and its straightforward integration with retrieval-augmented generation (RAG) techniques. |
Retrieval-augmented generation (RAG) is an AI technique that combines the natural language understanding and generative capacity of pre-trained large language models (LLMs) with the fidelity of a database query. Given a selection of material—texts, images, instructions, etc—a RAG tool learns embeddings, or relational representations of material fragments, which it then stores as a body of knowledge. For the purposes of this review, embeddings were created from the total search results using the open source software tool Nomic Embed v1.5. |
Colloquially described as ‘chatting with the documents,’ Interacting with a RAG tool consists of selecting which documents to use as context, submitting prompts and receiving responses. While RAG tools respond reliably when questioned about the subject matter contained within the documents, the technique performs poorly when asked questions about the documents themselves—so-called metadata questions, such as comparing one paper to another, or characterizing patterns or distributions across the set of documents. Prompts like “select publications from the provided documents that investigate the integration of artificial intelligence and patient-generated health data” return a mixture of actual citations and plausible but non-existent (hallucinated) paper titles, authors, summaries, and digital object identifiers (DOIs). On the other hand, prompts like “provide a summary of the consequences of integrating PROMs with AI, in terms of patient experience, and in terms of sociotechnical effects” return a clear outline without errors or hallucinations, complete with a list of traceable citations from the provided context. At times, a RAG model will lose the scope of the whole context, and may focus on a small subset of papers. Hyperparameters, factors that adjust the model’s scope and variability with regard to its own internal parameters, can have an outsized effect on the quality, fidelity, and creativity of responses. In short, the experience of using a RAG tool for insight into a body of knowledge is less a factual exercise than a tactical one: iteration, synthesis, views of the landscape. Value is added, but actual reading for comprehension and comparison is not replaced. |
For the full-text analysis, two pre-trained large language models (LLMs) were used: “Nous Hermes 2 Mistral DPO,” a 7 billion parameter model trained by Mistral Al and fine tuned by Nous Research on the OpenHermes-2.5 dataset, and “Llama 3 Instruct,” an 8 billion parameter model trained by Meta. The text generated by each model was fact-checked, compared, and synthesized by hand. The Results section below draws in part from the generated results. |
Discussion |
Methodology |
A narrative review is a non-systematic methodology intended to address “topics that require a meaningful synthesis of research evidence that may be complex or broad and that require detailed, nuanced description and interpretation.” To understand scientific literature as literature, a body of research may be productively examined with subjective, interpretive tools rooted in humanities and social science traditions such as constructivism—a research approach that seeks to understand the social, experiential, and interpersonal foundations of knowledge. |
Due to the emergent nature of the topic at hand, and the special case of AI as a sociotechnical assemblage, the perspective of this review aligns with theory-driven research traditions that seek to understand ongoing and messy interrelationships: between context and mechanism, between material artifacts, language, affect, and social factors.These traditions include material semiotics (or material-discursive analysis) and diffractive analysis, a method characterized by Karen Barad as one of “affirmative engagement,” rather than the “disclosure, exposure and demystification” of critique. |
To understand how AI and PROMs fit together—in the collision of generalized, automated pattern-matching with first-person testimony, mostly offered under duress—situated perspectives are required that take into account both the generic and the particular: how individual experiences collectively form patterns, and how those patterns create meaning for individual patients in return. When considering AI-PROMs’ incremental movement towards “that dream science/technology of perfect language, perfect communication,” as Donna Haraway writes, “Only the god trick is forbidden”—don’t make the mistake of escaping to a view from nowhere. Our analysis must reconcile the roles that both generalization and personalization play in shaping one another. |
Context: |
The relationship between AI and PROMs is a growing and potentially transformative field. At the center are AI technologies that aim to make sense of how humans express, reason, and exchange knowledge, such as NLP and large language models (LLMs). The recent acceleration of these AI technologies has seen the emergence of journals dedicated to the applications of AI in healthcare, such as NEJM AI and Artificial Intelligence in Medicine. For these journals, as well as the research under review here, the literature is addressed to a readership that spans disciplines within the medical field, hinting at the emergence of new hybrid practices that encompass informatics and healthcare services, while catering to a range of specializations such as oncology and orthopedics. |
In clinical trials of AI health technologies, PROMs are used as both input (data to learn from) and output (predictions made). Outcomes evaluated over time using PROMs, from postoperative depression to substance use disorder, are being studied as predictable patterns that can be addressed preventatively. While the use of PROMs in clinical trials of AI health technologies is increasing, the use of PROMs in the assessment of AI health technologies as a trial endpoint falls behind the rate of PROM use across all clinical trials. Designers of AI health tools must maintain a broad understanding of how the tools impact patients’ quality of life. Integrating PROMs with AI health tools, whether functionally or as an assessment, helps to ensure that the patient perspective remains central to both clinical research and healthcare delivery. |
Results: |
A narrative of continual learning and optimization, with patient-centered outcomes as the primary endpoint, runs throughout the literature as a hallmark of AI-PROMs. AI excels at identifying patterns and correlations between features, both implicit and explicit. Integrating the predictive pattern-matching aspects of AI with the patient-centered nature of PROMs may help to support personalized care and advance the goals of patient-centered medicine. For example, designing AI-supported surveys that can dynamically adapt to match individual patients’ needs and capacities, patients may become more engaged participants in their own care. In turn, more engaged patients may contribute actively to the collection of data about health status and interactions with healthcare systems, leading to more accurate, reliable, and culturally sensitive PROM instruments that better reflect patient experience. |
Note: The indented content below has been hand-edited from AI-generated responses (refer to the Methods section above for details). |
What an AI-PROM would do: |
Improve diagnostics and make more accurate measurements: AI presents the opportunity to contextualize PROs “with a large number of clinical, biological, and psychological data.” AI algorithms using NLP can analyze text-based PROM responses, such as open-ended questions, to identify themes and patterns that may not be captured by traditional survey methods. AI algorithms can analyze vast amounts of medical data quickly and accurately, leading to earlier and more accurate diagnoses. This not only improves patient outcomes but also reduces the time spent in the diagnostic process, minimizing stress and discomfort for the patients. |
Assist in research and evaluation: PROMs are already used as an outcome in clinical trials to assess the effectiveness of treatments, interventions, and health services. AI-PROMs may extend this by contributing to the iterative improvement of PROM design. By analyzing response patterns, AI tools may be used to design PROMs that are more effective at capturing patients' experiences, while reducing the burden of lengthy questionnaires. Furthermore, AI-PROMs may help guide the design of clinical trials, identifying key patient-reported outcomes, predicting sample sizes required for statistical significance, optimizing study protocols, and analyzing results to determine whether interventions produce minimum clinically-important differences (MCID). Additionally, AI-PROMs can help identify areas for ongoing improvement in treatment quality by highlighting disparities and monitoring trends in PROs. |
Improve patient-provider communication: AI-PROMs may facilitate communication between healthcare providers and patients, showing clinicians a more comprehensive and context-rich representation of patient needs, providing cross-cultural and multilingual support, and assisting in timely input from providers on patients’ survey responses. |
Be available in real-time, all the time: A core value proposition of AI-PROMs is automation, in the form of fast, round-the-clock support from chat-based AI health assistants, symptom trackers, disease monitoring and management tools. Utilizing AI-PROM systems to attend to patients' symptoms, quality of life, and other health outcomes in real-time may enable more effective disease management, as well as improved patient outcomes and experience. |
Support shared decision-making: AI-PROMs may help guide the process of shared clinical decision making, providing actionable insights into patients' symptoms, quality of life, values, medical history, and treatment outcomes, while also making health information legible and accessible to patients in alignment with their needs and preferences, |
Conduct data analysis: Collecting large-scale PROM datasets that are inclusive of diverse patient populations—from sources such as electronic health records (EHR) and patient registries (repositories of anonymized information related to specific populations or diseases)—will improve AI model performance and generalizability. Analysis by AI-PROMs may help to identify patterns and trends that may not be apparent through traditional statistical methods, such as identifying minimum clinically-important differences (MCID). AI analysis can be applied to a wide range of data, including extracting PROs from unstructured clinical narratives, as well as mapping the extracted PROs for phenotyping and clustering. |
Produce predictive models: By analyzing historical patient data, AI can predict potential health issues before they manifest or worsen. Early intervention based on these predictions can significantly improve patient experience by preventing complications and reducing the need for invasive treatments. Incorporating PRO data as an input for AI health models ensures that patients’ perspectives, not just clinical observations, inform the features from which the algorithm makes its predictions. |
Support personalized care: By integrating PROs into AI-driven decision support systems, healthcare providers can develop tailored treatment plans that better address patients' unique needs and preferences, as well as assist with patients’ health literacy and information needs. For instance, AI analysis may identify patterns that indicate the presence or impact of traumatic experiences on a patient's health outcomes. This information can then be used to inform healthcare providers about potential trauma-related issues and help them tailor care plans accordingly, incorporating trauma-informed practices. |
Support efficient resource allocation: AI can help optimize resource allocation in healthcare systems by predicting patient needs and allocating resources accordingly. This ensures that patients receive the appropriate level of care when they need it, reducing wait times and improving their overall experience. |
Reduce administrative burden: By automating administrative tasks such as appointment scheduling, billing, and record keeping, AI can free up time for healthcare professionals to focus on direct patient care. This not only enhances the quality of care but also improves the overall experience by reducing wait times and increasing face-to-face interaction with providers. |
Contribute to risk stratification: AI algorithms can identify high-risk patients based on various factors (e.g., medical history, lifestyle), enabling targeted interventions and improved patient experience. |
Current barriers to widespread adoption of PROMs in clinical research that may be addressed with the integration of AI tools: |
Resource constraints: While collecting, analyzing, and interpreting traditional PROM data takes time and requires trained personnel, AI-PROMs may reduce administrative work by assisting healthcare workers with administrative tasks such as summarization and routine analysis. |
Limited specificity: Validated PROMs may not be widely available for certain conditions or populations. AI may help to speed up validation processes, as well as to supplement existing PROMs by extracting patient-reported outcomes (PROs) from unstructured text such as interviews and clinical notes. |
Cultural and linguistic barriers: To provide meaningful data, PROMs need to be culturally sensitive and translated into languages spoken by diverse patient populations. AI may help capture linguistic context and nuance across languages. |
Patient burden: Completing PROMs can be difficult and time-consuming, especially for people who are already experiencing fatigue and high cognitive load: “Questionnaires can be burdensome to complete, especially when multiple domains of patient health are assessed at the same time.” Computerized adaptive testing (CAT) adjusts the survey instrument to simplify the test-taking process in the moment, matching patients’ capacity while preserving test validity. Passive data collection—through devices and apps, for instance—is another low-impact source of patient-generated health data that can provide context for PROMs through data-driven analysis without contributing adversely to patient burden. |
Data quality concerns: Incomplete or inaccurate data collection can compromise the validity of results. AI methods for working with sparse data, filling in gaps, and generating synthetic or proxy data may help glean meaningful patterns, even from partial results. |
Ongoing concerns that may inhibit adoption of AI-PROMs: |
Lack of standardization: While separate guidelines exist for the use of PROMs and AI in clinical trials (such as SPIRIT-PRO and SPIRIT-AI, respectively), standards for their integration remain to be developed. Furthermore, while individual PROMs undergo validation for how they measure different aspects of patient experience, there is no interchangeable standard for PROMs such as there is for other forms of health data, such as EHR. This makes it challenging to compare results across studies, and may contribute to a lack of clarity on actionable steps following assessments provided by PROMs. |
Entrenched behaviors in clinical settings: While integration of PROMs data implies a reorienting of clinical relationships to better align with patients’ subjective experience, in practice they often “function more as a tool to support patients in raising issues with clinicians than they do in substantially changing clinicians’ communication practices.” |
Patient privacy and confidentiality concerns: Using electronic PROMs, as with the sharing of any sensitive data, carries risks of exposure and theft. |
Concerns about the interpretability, explainability, and accuracy of AI algorithms, especially when making critical decisions affecting patient care. Difficulty understanding how AI decisions are made can lead to mistrust among patients, clinicians, or researchers. |
What AI-PROMs must do: |
Be lightweight: Assessment processes should be streamlined or made adaptive, with fewer questions, more focus, more presence. If PROMs are to be further integrated into clinical workflows, they must reduce workload for patients and providers alike. |
Be responsive: make results, interpretations, and communication available to patients, caregivers and healthcare providers quickly. In addition to real-time availability, AI-PROMs must be adaptable and accessible. |
Be transparent: be able to explain or assist in interpretation of analyses and results. Offer clarity on what to do next, and assist in shared decision making. Focus on improving trust and equity, and acknowledge the limits of automated decision making. |
Advance healthcare equity: AI-PROM systems must account for social determinants of health and other factors affecting health disparities, promoting more equitable patient outcomes, and leading to more complete understanding of the burden of disease, in alignment with research goals. |
Involve patients: Patients' perspectives and experiences should be included throughout the process of developing and evaluating AI-PROMs, to ensure that AI health technologies align with their needs and values. |
Speak all languages: Learn from data in whatever form it comes in, with multimodal understanding. Combining PROMs with other types of health-related data (e.g., medical imaging, genomic data) may create more comprehensive AI models. Furthermore, learning relationships from a wide array of contexts and across languages—structured or unstructured text, voice audio, data from wearable devices and EHR, visual diagrams and graphic aids, social media, and so on.—may contribute to better accessibility and health literacy, provide a more comprehensive understanding of patient experiences and outcomes in context, and a more intuitive grasp of what is meaningful to patients. |
The hopeful promise of integrating AI with PROMs is that tools which offer broad leaps in adaptability, pattern-matching, personalization, real-time responsiveness, and continuous learning will lead to better patient-provider communication, increased trust, and, for patients, a sense of ownership and motivation in managing one’s health. Taken together, these factors may contribute positively to a range of outcomes, including alleviating patient stress, better data collection and interpretation, and better-informed clinical decision-making. |
Fundamentally, AI-PROMs must extend the core principle of PROM instruments, that of validating patient experience. Care providers who use AI-PROMs should center patient concerns by acknowledging that their symptoms and health concerns are legitimate and worthy of attention, showing understanding and compassion for what they are going through, and confirming that their experiences are real and will not be dismissed or minimized. |
(Unanswered) Questions |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.