text
stringlengths 1
1.75k
|
|---|
Prompted to Simply ask a question by the AI agent, I pause to remember myself and what my questions consist of. |
[...] |
I keep returning to my oncologist’s admonition to “not become a professional patient.” To engage with medical care beyond preventive and routine maintenance, to be a complicated problem, to hold multiple conditions, co-morbidities and confounding factors, requires professionalizing oneself as a patient—to become information, to perform legibility. |
Biological mediation, in Ebony Coletu’s formulation, “refers to any structured request for personal information that facilitates institutional decision-making about who gets what and why.” |
The varieties of functional life writing, from job and school applications to requests for social support, scholarships, grants, and so on, entail a continual retelling and reshaping of one’s biography to match institutional patterns and categories of need. |
As “a recurring form of self-disclosure required when requesting assistance or cooperation to achieve life goals,” this kind of life writing “transform the ways we speak of opportunity,” as disabled, sick, poor, incarcerated, and any other marginalized people know deeply. |
For example: in the U.S., to be recognized as disabled requires the performance of belonging to a formal category. To receive assistance through public support requires—alongside the formal application process—the loss of privacy, the continuing review of resources and living circumstances, and adherence to a low material standard of living: “ceremonies of social degradation,” to use disability historian and activist Paul Longmore’s term. Each of these burdens serves only to maintain the legible authenticity of the applicant. |
What does legible mean in the context of patient-reported outcome measures? What are the consequences for a patient who returns PROMs that are incorrect or incomplete? What makes a good or responsible completion of a PROM survey? How are outliers folded back in to maintain meaningful patterns? What use is an ambiguous response? |
[...] |
Performance studies scholar José Esteban Muñoz’s use of the term disidentification, following the work of linguist Michel Pêcheux, is crucial for working with productive ambiguities: |
Interrupted in the activities of daily living by a prompt—an interpellation—to fit a pattern, one can either adjust their behavior and presentation to match the pattern; or resist, continuously pushing back with a mirror pattern of their own. |
The third option is to disidentify, where the subject “neither opts to assimilate within such a structure nor strictly opposes it” |
Instead, the encounter with dominant patterns becomes an enabling misreading, where a subject is able “to read oneself and one's own life narrative in a moment, object, or subject that is not culturally coded to ‘connect’” |
Disidentification reframes contact between dominant patterns and subjects as something like a missed encounter: instead of forming either matching or mirroring patterns, we stay incommensurable. |
[...] |
Dismediation, as proposed by Mara Mills and Jonathan Sterne, extends Muñoz’s usage of disidentification to understand the ways disability and media shape each other. The process of dismediation, as Mills and Sterne describe it, takes “disability as method, not simply as content” for study. Instead of holding onto universal models of communication, dismediation “begins from a presumption of communicative and medial difference and variety.” Instead of using media as “the tools to repair a damaged or diminished condition of human communication,” Mills and Sterne advance a basis for “communication as something fraught, supplemented, and interdependent in all of its many forms.” |
Dismediation, as an analytical process, begins with the premise that there is no ground truth, no world that is available to the senses without media, and that “every media form is built around different ideas of the natures of human subjects and bodies. |
[...] |
What is relevant, what gets discarded? Why has this record of my life (health, experience) been edited in the way it has? Or, what model would I prefer to capture myself with? |
Life writing, as the continuous interchange between what we carry with us (our prior knowledges, our hupomnemata: ‘notebooks’, as Foucault puts it), plus what we experience as new in the present moment, is a tangible way of feeling how representations are not static, but have their unique movements and flows. |
Social, cultural, and technological contexts each contribute to the evolution of what we model as natural language. While this suggests a continuous requirement for upgrade and improvement of natural language models, there is equal demand for better awareness of the factors that create these contexts, and the influence of feedback between the models and what they represent. |
This kind of inferential logic is called abduction—the tacking back and forth between futures, pasts and presents, between ideas and observations, as a way of building theories. |
Adele E. Clarke traces the genealogy of the term to pragmatist philosophy that embraced the potential of abductive inquiry—in contrast with inductive or deductive logics—to produce concepts out of deep awareness rooted in tangible evidence. As a form of educated guessing, abduction “is not solely intellectual or cognitive, but also experiential.” |
For the foreseeable future, abduction, creative inquiry, and human-in-the-loop computational processes remain critical to the project of embedding patient-centered outcomes at the center of AI health tools. |
Good representations: |
Analysis is the technique by which representations are made. What is a good, or necessary, representation? |
Consider the current tools of generative AI: Transformers learn what is important and pay attention to it. Diffusers destroy structure, then learn how to iteratively restore it. Variational Autoencoders learn what is essential to make a compressed version of something, then reverse the process to learn how to make it whole again. Generative models use latent space to creatively explore complex things in a simpler and more meaningful form. Meaningfulness is in proportion to noise. A signal is meaningful when it carries more information than noise. Each of these tools, when used generatively to resynthesize from a learned representation, provides tangible evidence of what has been learned, what properties of the original data have been captured, and in the gaps: what properties have been lost. |
[...] |
The goal of analysis is to produce generalizable insight into the original object, or data, being studied. |
Natural language processing, or how computers come to understand and use language as humans do, has driven what we think of as the capacity for AI to effectively communicate meaning. Multimodal AI, or the modeling of relationships between different representational modes—sound, image, timbre, texture, tone, vibe, scene, gesture—extends this capacity, tacitly acknowledging that concepts of natural language are inherently incomplete, that meaning is contextual, relational, and sensory. |
[...] |
Can we speak in terms of the texture of illness? Of pain? The texture of neurodivergence, or a non-normative bodymind? Or even the texture of evaluative methods, such as patient-reported outcome measures? How might these textures be represented, or synthesized, so as to better understand them? |
The task of distinguishing one texture from another, as an aspect of images, language, and of signals more generally, has long been a focus of AI and computer sensing. |
Image description: A black and white image containing eight unique textures tiled together as a mosaic. The image is designed as a test card to prepare computer vision applications to discern transitions between textures along non-horizontal and non-vertical boundaries. The accompanying map of texture regions is shown below. |
Do you see it? The textures are present in the image in approximately equal proportions. The upper-left portion contains three textures in an arrangement where two textures converge along curved paths against a background of the third texture. The upper-right portion of the mosaic contains regions with non-vertical and non-horizontal boundaries, both straight and slightly curved. The bottom half of the image is made up of the eight textures in irregularly shaped regions of approximately equal size. |
Texture—how a thing feels, the attributes of surface and structure as they appear in a consistent way to the senses—is modeled as transitions between local and global attributes, as territory (topos). |
[...] |
I will for a moment draw a parallel between instruments that measure (medical, scientific) and instruments that make (music, images). |
To create, or modify the audible textures produced by a musical instrument, a musician may employ what are called extended techniques: Precisely coordinated movements and positioning of bodies, lips, tongue, breath, fingers, lungs, throat. Some extended techniques obscure the player’s physicality, masking or rerouting the sound’s perceived origins. Others draw added attention to the player’s physical presence, the shape and posture of their body, the grain of their voice, their various capacities and endurances. |
Extended technique suggests the extension of a world, making it more expansive in space, time, and imagination. At the same time, naming something as extended removes it from a set of otherwise standard or normal practices. This distinction is not self-evident: Instruments invite play, discovery is within their regular use, if not somehow against the grain of habit. |
Thought of as playing different or out of bounds brings extended technique into the shared metaphorical space with notions of (un)natural places and bodies that Eli Clare writes about. This is about pushing at the boundaries of a representation, the space of a manifold, not to erase the contours of constraint, but to soften them with practice, to make their shape more evident—more textured—and inclusive of variation. |
[...] |
Classification, sorting things continuously into one or another category, needs good representations. Working towards better mathematical representations of a signal’s texture, researchers have formalized the requirements of a good representation. |
A good representation should meet the following criteria: |
1. If the signal is translated, the representation should not change. |
2. If the signal is deformed, the representation should be deformed in a proportional way. |
3. Signals that are different should be represented differently. |
This can be learned. |
[...] |
Unless we think of everything as a signal, the goodness, or necessity, of a representation depends on its purpose. |
As the purpose of representation moves from simulation to finding and explaining patterns in data—from discrimination (what number did you write?) to recognition (what patterns in your behavior define you?)—measures of goodness shift from interpretability, and the ability to recall with precision, to the accurate predictions, the capacity for reason and action. |
Before integrating AI and patient-reported outcome measures, consider what the core tasks of making representations will be, the different consequences of each task, and defining the task shapes how the tools are used: Are we extracting information from a noisy signal, or summarizing, making abstractions and translations? Generalizing from specifics, or fine-tuning a general model to meet local needs? What kinds of prediction will be made? |
[...] |
“The myth of prediction,” as outlined by writer and critic Nora N. Khan, forecloses imagination—it “shapes our sense of possibility to an extreme degree” through logics that orient us “in bounded, limited ways toward people, places, the possibilities of how our lives can unfold.”[100] |
[...] |
The goal of a representation is to make our own latent selves recognizable. |
Meaningful Analysis |
In search of a working definition of analysis, I turn to Audre Lorde’s “Poetry is not a Luxury.” Poetry, she tells us, is a form of self-analysis, “the quality of light by which we scrutinize our lives.” Poetry, as analysis, is the latent space of what we can sense and know: “It is through poetry that we give name to those ideas which are, until the poem, nameless and formless-about to be birthed, but already felt.”[101] |
An example image from Christina Morgan and Henry Murray’s set of Thematic Apperception Test (TAT) cards (1935).[102] What story does this picture tell? Image description: a vertical rectangle of white, to the left of center. Framed as a silhouette in this rectangle are the outline of an open window or door, and a partial human form, which occupies the entire bottom right side of the white frame, occluding the corner and continuous with the black background. There are no details or breaks in the image apart from the silhouette, rendering the entire image in only black and white, no shades of gray. The figure’s arm, outstretched horizontally, bisects the frame at roughly the vertical midpoint. The figure’s head, seen in profile facing to the left from the right side of the frame, tilts slightly up. At the bottom of the frame, an ambiguous shape interrupts the white rectangle, continuous with the silhouette of the figure. |
I see Lorde’s “quality of light” in the processes described by artist and writer hannah baer as intrasubjective restoration. In this frame, baer is drawing a comparison between generative AI and the psychoanalytic use of ‘projective’ tools: ambiguous forms, such as Rorschach inkblots, or indistinct narrative illustrations, like those included in the Thematic Apperception Test (TAT). These images are not particular—they don’t index anything in reality, and their meanings are not fixed. And yet they are recognizable through the affect responses they produce and the narratives they conjure. We relate to them. |
Projective instruments do not validate well: there is no demonstration that they consistently measure constructs as intended. However, they serve another purpose: by rerouting both the clinician and the patient’s attention through the image, they give us something to talk about. We can ask questions and fill in missing data. |
AI, as either the simulation or resynthesis of familiar patterns, can be a catalyst for generative misrecognition—a tool for defamiliarization (dépaysement) and disidentification. As processes of affective disentangling, we may (optimistically) use AI to apprehend gaps or biases in representation, filling in data from our own experience to understand the tools as non-innocent, and to disassociate from normative visions of selfhood. |
If, as baer asks, projective instruments such as Rorschach or TAT, “don’t contain particular images and instead just help us tell our own stories, is that also what we’re doing with the ambiguous figure of AI?” |
Our engagement with the ambiguity of AI is not innocent or un-mediated: As Meredith Whittaker warns, the computational systems that produce these engagements are often private and pursuing the capture of attention, access, and control, “threading through our public life and institutions, concentrating industrial power, compounding marginalization, and quietly shaping access to resources and information.”[103] |
Where AI is applied as a mediating layer between individual persons and health infrastructures, it is all the more critical to direct our attention to how power works on and through each point of engagement. It is also imperative to push humanist sway onto emergent logistics of evaluation, optimization, and management—tilting towards what baer calls “a world where deep transformation—creating something that connects us more deeply to ourselves and one another, redrawing our self-image—is the tendency,” rather than reproduction of poorly fit and inequitable categories.[104] |
[...] |
Ensuring that person-centered and prosocial outcomes are at the heart of AI technologies, particularly in healthcare, requires specific and coordinated work.[105] This is the work of advocacy, of design, and policy-making, but it is also the broad theoretical work of tending to categories, and critical methods such as what Adele E. Clarke “work is the link between the visible and the invisible.”[106] |
Considering the shift from actuarial analysis to predictive analysis and data-driven forecasting across science and industry, Clarke elaborates on anticipation as the synthesis of abduction, simplification, and hope. |
Abduction, discussed previously in this chapter, is the inferential task of tacking back and forth between emergent theories and observations. Simplification—editing, sorting, arranging, representing, abstracting and generalizing—includes all the tools and techniques for making something complex more manageable, including what Clarke, Leigh Star, and others in feminist science and technology studies have pointed to as deleting the work.[107] Hope, finally, is the affective dimension of anticipation, a driving force, an outcome, and a commodity within anticipation work. |
[...] |
The work of maintaining categories: The work that curtains do. Osmosis and diffusion. Voter district remapping. Breathwork. Border walls. HVAC systems. Difference without separation. Bodies without organs. Liminal categories. Lenticular logics. Single stream recycling. Immune response. Navigating personal boundaries. Compassion and intersubjectivity. |
[...] |
Classification has consequences.[108] What are the best practices for making and maintaining categories? |
Center the interpretive flexibility that humans do well, with all the bias and richness context our positionalities bring to the process. |
Temper the flow of interpretation, thinking in terms of boundary objects, a concept laid out by Leigh Star to describe arrangements that allow different groups to work together without consensus.[109] A map that guides different groups of people to experience the same site in different ways is one kind of boundary object. Many technologies are. A library, where people disagree about which books belong and which should be removed, is not one. |
“Often, boundary implies something like edge or periphery, as in the boundary of a state or a tumor. Here, however, it is used to mean a shared space, where exactly that sense of here and there are confounded.”[110] These objects work because ownership is ambiguous but each group who takes part finds their information needs satisfied. Knowledge is always partial: no one knows everything, but everyone knows something. As objects holding the shared information needs of patients, providers, researchers, and caregivers alike, patient-reported outcome measures are an ideal example of boundary objects whose use is already navigated separately and together by each group. |
Recommendations |
Take a walk together, as an alternative to explaining or making an argument. |
In her germinal text “The Rejection of Closure,” whose characterization of open forms is discussed above, Lyn Hejinian draws on Umberto Eco’s reasoning as to the ways author and reader come together to make meaning—a position outlined in his 1962 essay “The Open Work.” Speaking to the generative creativity both reader and writer bring in response to an open form, and the “polygendered impulses” it activates, Hejinian quotes Eco on what he called inferential walks, a peripatetic practice of reading and writing that resonates for me as an analytical technique. An inferential walk is a method for embedding intertextual ideas, their frames of reference, across worlds. As Eco puts it: “to identify these frames the reader has to ‘walk,’ so to speak, outside the text, in order to gather intertextual support (a quest for analogous ‘topoi,’ themes or motives).”[111] |
Topoi, the plural of topos, a group of ideas stretched out into space, a problem space, or a decision space. Importantly, the notion of taking a walk to engage with ideas, doesn’t need to be strictly metaphorical. I’m thinking of a dialogue between Judith Butler and the artist Sunaura Taylor, who uses a wheelchair because of a congenital physical disability. As the two stroll through San Francisco’s mission district, they turn to the subject of walking in the context of disability. Taylor offers: “I use that word [walking] even though I can't physically walk. I mean, to me, I think the experience of going for a walk is probably very similar to anybody else's: it's a clearing of the mind, it's enjoying whatever I'm walking past. And my body is very involved even though I'm physically not walking.” Butler invites in return: “Nobody takes a walk without there being a technique of walking. Nobody goes for a walk without something that supports that walk, something outside of ourselves.” Part of taking a walk is the everyday opportunity to “rethink what a walk is in terms of all the things that power our movement, all the conditions that support our mobility.”[112] So, in walking we clear the mind, enjoy the surroundings, live in our bodies—while coming to understand how the environment (physical, technical, social, etc) supports, impedes, and redirects our passage. |
A central problem that walking gets at, for me, is the question of navigating territories—In terms of discipline, access, literacies, roles—the transversal cutting across and within boundaries of Rosi Braidotti’s nomadic subjectivity or Maria Lugones’ “world”-traveling.[113] |
[...] |
The medical humanities offers a number of approaches to the therapeutic exchange of perspectives, building and strengthening trust and recognition between patients, providers, and caregivers. Key to these approaches are engagement with ritual, performative and artistic forms that produce meaning in the link between internal and external experience, such as writing and creative making and role-play. I’m thinking in particular here about how the politically-engaged theater of Augusto Boal, and the recent innovations in conceptual, live-action role-play referred to as Nordic LARP might be applied in healthcare settings.[114] |
[...] |
In the following chapter, “The Validated Instrument,” I take up the question of what artificial intelligence (AI) technologies—including adjacent tools and techniques such as passive data collection, ubiquitous computing and computerized adaptive testing (CAT)—integrated with patient-reported outcome measures (PROMs) would do. Drawing exhaustively from recent scientific literature, I aim to trace the underlying narrative assumptions motivating research at the intersection of AI and PROMs, and to further trouble the foundational confusion set up by using both patients’ lived experience and automated pattern matching as key endpoints for research. |
The Validated Instruments: A Narrative Review |
Introduction |
Background |
This chapter, with a broad view, seeks to understand the characteristic research narratives that bring artificial intelligence (AI) and patient-reported outcome measures (PROMs) into allignment. While this view encompasses a wide array of disciplines and settings, it finds a limited range of narratives. The ascendant prioritization of patient- and value-centered models of care, and the shifts in knowledge and culture these priorities endeavor to bring about, provide the motivations, inform the underlying assumptions, and shape the ethical and technical concerns that illuminate the path forward. |
There are, as of now, no holistic hybrids at this intersection, no full integration of artificial intelligence and patient-reported outcome measure has been validated. However, the literature shows many paths towards the integration of AI and PROMs, through iterative advances in methods, across diverse contexts, with barriers to overcome, and an increasingly coherent sense of what the benefits could be as well as what cautionary measures need to be in place. |
Objective |
Two primary questions motivate this narrative review: |
1. What would an integration of artificial intelligence and patient-reported outcome measure (AI-PROM) do? |
2. What questions are researchers pursuing this integration asking, to know they are on a good path? |
Subject matter: |
What are patient-reported outcome measures? |
Patient-reported outcomes (PROs) are timely records of a patient’s experience of illness—their inner thoughts, bodily sensations, and social experience. PROs are raw data coming directly from the patient, without interpretation of the patient’s response by a clinician or anyone else. As such, PROs are one kind of patient-generated health data (PGHD). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.