Through Thick and Thin

The philosopher Gilbert Ryle developed the framework of thick and thin descriptions as a means to describe and understand human behavior. He illustrated it via the parable of the wink and a twitch.  Although both actions can be superficially described and categorized with a thin description such as ‘closing and reopening the eye,’ a deeper, more nuanced thick description would not only differentiate between the two, but would also make finer-grained descriptions such as a “conspiratorial wink,” a “flirtatious wink”, or a wink that serves as a “signal in a game.” In this framework, the former would be a thin description and the latter a think descriptions. Thick descriptions use knowledge of context, infer intent and motivation to understand the “why, what, who, and what” of the action. They don’t shy away from subjectiveness and interpretation, but also afford deeper analysis and categorizations. Clifford Geertz saw the value of such a framework for anthropological research and popularized it with his book, The Interpretation of Cultures. Since its publication in 1973, thick descriptions have become the gold standard for qualitative social science research. Although not typically viewed through this framework, a S.O.A.P note is a combination of thick and thin descriptions that includes facts, interpretations, and explanations.

Thick and thin descriptions also bear a resemblance to the concept and techniques of compression in biology and information science. Although immaterial in appearance, information bears costs in the currency of time, memory, attention, and energy. All successful organisms must navigate the costs and benefits of information acquisition, storage, interpretation, and transmission. Balance the trade-offs between false-positives and false-negatives. The dimensionality of information is compressed by ignoring, not capturing, or removing information that is deemed irrelevant. For example in a technique known as lossy compression, data that is deemed irrelevant to the application is removed from an image to decrease file-size. Although this makes images more grainy, in applications or contexts that are transmission speed dependent, storage or memory constrained, or image quality agnostic, lossy compression is a pragmatic solution imposed by the materiality of information. Clinical documentation is an exercise in lossy compression. A clinical encounter that is typically information rich is compressed into a clinical document. Physicians attempt at lossy compression – excluding from documentation information that is perceived to be inessential, low value, and lacking explanatory power. In the modern era of industrialized, bureaucratic, and metric tyrannized healthcare, it is subjective and intersubjective information that becomes neglected, genericized, and discounted at scale.

This compression has downstream consequences especially because health (and disease) are complex phenomena that emerge from the interplay between social, psychological, and biological systems. By exclusively focusing on biological systems we miss the sociology and psychology of disease processes. Additionally, neither the healthcare system nor medicine has the coverage or depth of knowledge to map or track most major disease classes (future essay). The toolkit of biomarkers and radiological imaging afford only superficial legibility to the pathological processes that are occurring at the subcellular level (future essay). The fragmented and episodic US healthcare system lacks the penetration to track disease progression with any regular cadence (future essay) or depth of insight. Thus, for the foreseeable future, thick descriptions of symptom descriptions, the felt experience of diseases, and meaningful physician-patient relationships will carry weight and be explanatory and prescriptive. They will continue to serve as priors (future essay) and can serve as category makers and markers.  However, EHR documentation has missed this opportunity and has moved in the opposite direction. EHRs datasets are thin descriptions that abound in compressed information in the form of diagnosis codes, dot-phrase documentation of shared decision making, and low-dimensional scales and scores that underrepresent complex phenomena such as depression or pain.

In summary, EHR datasets offer only a veneer of completeness and have systematic blind spots. They are thin descriptions – compressed for efficiency, misaligned incentives, and ease of use rather than utility. Valuable and relevant thick descriptions are missing at scale. This becomes even more consequential as EHR datasets become the primary source of training and testing data for machine learning and large language models. Will models trained on thin descriptions output “thin inferences,” coarser-grained classifications, and amplify the consequences of blind spots in EHR datasets (future essay)? Furthermore, as the EHR has ascended in importance, its representations are becoming a stand-in for the patient – a digital twin (future essay). The physician at the bedside is becoming a physician at the computer side. Thick interactions have been replaced by thin ones. A profession of care and personalization is mired in-a-between state, groping towards the horizon of cure and precision (next essay).