Blogs

This very interesting article by Gordon Shiff elegantly lays out a lament for primary care and for that matter the primacy of cognitive clinical medicine. It is a truism AI is a tsunami throughout the world’s economies and it is no surprise clinical medicine will be effected as well. However the changing industries incorporating robotics and other forms of AI raises the specter of impacts on workers and the type work they will do in the new economy. The value of AI as a vast marketing tool for social media, support for automation of repetitive tasks, causes concern that AI will have the same function on clinical person centered medicine. Here clinical medicine diverges from potential for automation at the person level unlike automation appropriate at industrial scale. AI unchecked at the person level will force the industrialization of clinical medicine. This is the source of Dr. Schiff’s lament as he witnesses how AI has impacted the cognitive practice of younger colleagues in training through the vehicle of AI enhanced electronic medical record.

The role of AI in our society is under intense debate about the spectrum of potential benefits and harm. The harm succinctly put happens when machines talk to machines without the input of humans; the benefit comes from structurally organizing vast amounts of information beyond the capabilities of humans to understand for purposes of operational efficiency in healthcare social and industrial processes. Uniquely known inputs may or may not be within the realm of machines but are within the realm of clinicians. Day to day features of clinical medicine such as wisdom, experience, patient preferences and other features of an endless list cannot be reduced to data points, and are therefore unseen when training sets are created for AI applications. Human intelligence can do discovery of unknown patterns, provide judgments under conditions of uncertainty, and build novel relationships among humans as collaborative entities. There are many human qualities that resist pattern making by machines, but this is not to say  algorithms cannot do a reasonable job of presenting data to humans functioning as communities of knowers to add the human dimension to AI output.

Dr. Schiff presents the hazards of AI in the clinical context, which is clearly not the context of the transactional, industrialized business of healthcare. But , trends are disturbing, and while we have an opportunity to arm primary care to nudge AI toward more benefit and less harm, there needs to be a detailed understanding of what clinical medicine means within the frame of reference of  the data AI builds on.  More than precise definitions of clinical medicine as notes, codes, signs and symptoms, therapies, outcomes, etc, abstract non-codable entities, such as trust, relationships, pathways, preferences, uncertainty of outcomes, coordination, collaboration exist as counterfactuals. These counterfactuals are ignored in AI systems unless they are given importance and named by clinicians. The most important abstraction is relationships, which drive all the other counterfactual abstractions and assembly of transactions.

Herein lies the opportunity for primary care to insert their clinical skills and cognitive skills to not just mitigate the hazards of autonomous machines, but to bring clinical medicine into the line of sight of the lens of AI. Doing this requires reimagining what and how healthcare ecosystem standards apply to medicine. From the perspective of financialized and industrialized health care IT systems run the show; the lens of cognitive medicine, existing in all area of clinical medicine, does not fit into these IT systems except as outcomes of the complex and deep relationships and processes. These outcomes are simplified as transactions. The lens of cognitive medicine finds the “dark matter”, to use a metaphor from cosmology, which when focused on the personalized, can quickly adapt from this fundamental level to relationships pertinent to the clinical task at hand, then to the bigger world of abstractions that captures knowledge from populations, and most importantly measures and captures the complexity of relationships.  This is the opportunity of primary care, to accept the fluidity of the personalized if given the opportunity to do so by new and novel features of AI. Dr. Shiff recognizes this dark matter, included in his clinical note, but finds these unique characterizations of the patient encounter washed out by AI algorithms. What is left of the note serves efficiency, reimbursement, and supports the encounter as a transaction, though sterile, and nowhere close to the meaning that binds the patient and physician in a relationship of trust.

To amplify the opportunity for primary care, making the dark matter of fundamentals explicit and amenable to algorithmic formulations of AI will  bring into the information ecosystem the granularity of personalizing.  In other words, the clinical note as the Rosetta stone of the note as a holistic rich, energetic reflection of the complexity of the patient physician relationship, must capture the uniqueness of the time and place of the encounter, and maintain this granularity as the information migrates into the transactional world as aggregates of information.

This begs the question of how to make the specificity of the clinical note generalized to the healthcare ecosystem. To be clear, the ecosystem is a metaphor for many organizations that exist in various relationships and each of these entities perceive the complexity of the ecosystem as a maze. To consider the ecosystem in this way, relegates the clinical note as the Rosetta stone of fundamental, personalized data to a small, almost inconsequential entity in competition with larger transactionally oriented organizations.  The clinical note can be overlooked except as a conduit of limited transactional information that serves financial interests and other objectives of networks and populations.

To present the Rosetta stone as a metaphor for the fundamentals of personalized clinical care highlights the nexus of these notes with the entirety of the healthcare ecosystem. Meaning comes from details at the small level, and meaning is created and communicated in different ways as information flows through the hierarchy of healthcare reaching different audiences. Here is meant personalized patient centered care, coordination as networks, and information supporting healthcare policy. The potential to impact these organizations requires translation of meanings to support relationships.               

It is natural to believe that AI will assist and in fact define this process of information transfer. But will AI do what wise clinicians do when they assemble the personalized Rosetta stone? Can AI enter the world of dark matter of the unobservables, unknown influencers, preferences and all other features of the unobservables. Can a wise clinician enter this world? If not then how can AI do this? Here a time out is in order to step back from AI technology as a solution and consider what may be called a philosophy of information. This is necessary to collapse the maze of the ecosystem to small actionable relationships and create a space for the small and not just for the massive amount of data needed for training sets. To be clear the opportunity for primary care can be and will be to assume the task of ferreting out the encounter of granular information from the ecosystem information flow writ large when the flow originates from the personalized platform of the Rosetta stone. This is the role of curation, a process that is antecedent to probabilities, the tool of AI.

Raising fundamentals to a secure position in the information flow of the ecosystem requires probabilistic thinking (ref 3^rd Millenium) This allows flow of information from the smallest to the largest entities in the ecosystem. Importantly, placing fundamentals (Rosetta stone) at the smallest scale , represented by primary care as a representation and avatar of the  patient makes the source of primary information into the ecosystem as the potential source of transformation by disrupting or modifying the normal flow of information of the ecosystem.

Although it is assumed AI require large numbers (N) for pattern recognition, AI can in fact support the entre of information into the ecosystem with small or N of 1. This raises the question of what AI can produce from N of 1? Consider the granularity of information as attested by a PCP; as the patient curator, he  (she) will recognize patient features that are probabilities that represent parameters of populations. This is the meaning of risk factors; in fact, a single patient can claim membership in multiple populations  risk; curation amounts to judging the impact of multiple risks at the patient level. But more than quantified risk, PCP’s judge the impact of counterfactuals, some of which are identifiable, and some are not. Support for adding counterfactuals to the process of patient data curation comes from the concepts of necessity and sufficiency, elaborated as the quantity Probability of Necessity and Sufficiency, PNS. Added to observables, population defined clinical risk, allows curation to extend the package of patient centered fundamentals.

To elaborate further on fundamentals to explain the utility of PNS, more needs to be said about counterfactuals. These can remain unknowable and unmeasurable, or they can be known in contexts of groups of patients, though not available as actual fact in a particular instance.  When populations are the focus of measurement, counterfactuals can exist as indirect contributors or confounding biases to statistical estimates from probability mass functions (PMFs). Consider how astronomers discover planets; not by directly observing them, but by measuring the effects on spectroscopic signatures of the star caused by a planet. So goes clinical medicine; outcomes can reflect causes interacting with other direct causes – this is an indirect mediation of cause and effect. This unseen cause can be teased out by careful observation by cognitive agents, but the difference when the measurement, prediction if you will, is personalized, not the population, data of outcome as effect of cause does not exist, because the potential effect of the cause did not happen. This is the status of a personalized encounter that qualifies this encounter as fundamental, this lens in a moment of time must include realized in contrast to potential data. If this formulation of clinical decision making sounds esoteric, or at most unrealistic, it reflects how wise clinicians do their job. An other way of saying this is to respect how clinicians expect to deal with uncertainty, where counterfactuals are always around the corner.  The trick is to identify spurious causes and confounders without the assistance of randomization. Causal methods have the tools to do this to assist the curation process of “wise clinicians”. The most important and useful tool in this regard is graphical equation modelling, a visualization tool to capture and account for confounders and spurious causes which allow the fundamental level of clinical information to be where cause is distinguished from correlation. The utility of PNS is to quantify the details of the personalized. (Pearl Personalized article 2023)^[1]

Curation as the fundamental tool of personalization, captures the benefits and limitations of AI. Wise clinicians populate the Rosetta stone of the personal, inaccessible to AI because only observable data supports AI; clinicians deal in the broader scope of the personal, with potential as well as actual observable data as well as abstractions that inform explanations and actions leading to effects (outcomes). Once personal decisions are made, the data migrates to populations as aggregates of actual data, which is in the realm of AI. But first curation by the wise informed clinician builds the fundamental source of personalized information that is the basis of information propagation through the ecosystem.

Dr Shiff is clearly the exemplar of the wise clinician, and it will be a good day for clinical medicine when he and his “community of knowers” are in the driver seat. In a world rearranged by the personal, the flow of information becomes the power of information.

Robert Ripley MD