Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Classification Systems in Neurosurgery
This chapter includes an accompanying lecture presentation that has been prepared by the authors: 
.
Key Concepts
-
Many classification systems have been described for use in neurosurgery.
-
No classification system for these classification systems has previously been described.
-
In evaluating a classification system, it is important to assess its consistency and validity as well as the strength of evidence supporting its utility.
-
We introduce a three-tier hierarchy for classifying clinical classification schemes. In this scheme, type 1 includes purely descriptive classifications. Type 2 includes “intermediate” systems that, in addition to being descriptive, have either a prescriptive or predictive component, but not both. Systems that have prescriptive capabilities are considered type 2a, and those that have predictive implications are considered type 2b. Systems that have all three features are considered “comprehensive” and are classified as type 3.
-
We provide a comprehensive list of all classification systems identified by authors of Youmans & Winn Neurological Surgery , 8th edition, annotated in part using this tiered classification system.
Classification systems abound in the neurosurgical literature and in the medical and surgical literature more widely. Some systems endure the test of time, either by withstanding empirical validation, undergoing iterative refinement, or persisting for reasons of convenience or stubborn convention in spite of limitations or inaccuracies. Others, and there are many of these, remain relatively obscure. Some classification systems have historical or honorific significance, associating the name of an individual, location, or institution with an eponymous framework for approaching a particular topic. Medicine and surgery, in their respect for memory and tradition, have tended to preserve such systems. Other classification systems, originating in the modern era of international and multilateral collaboration, evidence-based practice, and systematic reviews, cannot be attributed to the work of individual experts. Some classification systems are purely qualitative, while others have quantitative features. Some systems are purely descriptive , some are explicitly prescriptive in informing treatment, and others are predictive or prognostic. Classification may be referred to as staging, grading, scoring, scaling, or by a variety of other terms. The classification systems themselves, however, have not been systematically classified, and limited guidance is available informing the development of classification systems. This chapter represents an initial effort to compile and organize the neurosurgical classification systems in contemporary use and to classify them according to type and according to several criteria pertaining to their practical utility.
Descriptive, Prescriptive, Predictive
We begin by exploring the definition of a classification system from mathematical and statistical perspectives as well as from practical and utilitarian perspectives. We then discuss principles of classification, with an emphasis on the properties of classification systems that render them valid and useful. Classifying classification systems is of course itself a form of classification, and our aim is to provide a useful reference to the many existing classification systems in our field and to offer some organizing principles for generating, using, and refining such systems. As a concrete, initial contribution, we then present a scheme for classifying neurosurgical classification systems according to whether they are purely descriptive (type 1), prescriptive (type 2a), or predictive (type 2b), or have all three features (type 3). We apply this scheme to a comprehensive set of 488 classification systems described in the current edition of Youmans & Winn Neurological Surgery, annotating each system with a cross-reference to the chapters in which it is described in the current edition. Finally, we discuss issues of bias and limitations of this classification effort as well as implications for future refinements.
Defining Classification Systems
History and Philosophy
The term classification connotes a scientific organizational process, but its precise meaning differs from field to field. The Oxford English Dictionary offers the following definitions :
(1) The action of classifying or arranging in classes, according to common characteristics or affinities; assignment to the proper class. (2) The result of classifying; a systematic distribution, allocation, or arrangement, in a class or classes; esp. of things which form the subject-matter of a science or of a methodic inquiry.
The importance of classification in the natural sciences was recognized at least as early as Aristotle, whose Historia Animalium, written in the fourth century bc, contains a scheme for classifying all known forms of life. More systematic, modern classification efforts, however, are often traced to Carl Linnaeus, who described, in his Systema Naturae (1735) and subsequent works, a hierarchical taxonomy of living things that has had an enduring impact on science through modern biologic nomenclature.
In the late nineteenth century and early twentieth century, the notion of classification received careful attention among mathematicians dedicated to establishing the logical and philosophical foundations of mathematics. It is difficult to overstate the importance of these issues to the mathematical community, and they attracted talent through major contributions from John von Neumann, Kurt Gödel, David Hilbert, and other legendary figures. The concept of a “set,” a well-defined collection of objects, is fundamental to mathematical reasoning, and the early twentieth century saw the development of rigorous methods of defining and proving membership or nonmembership of mathematical objects in particular sets. Classification thus became an essential aspect of “set theory” (the branch of mathematical logic that studies collections of objects) and the foundational logic of all of mathematics.
The issues facing pure mathematics during its era of foundational crisis are not pure esoterica. They have bearing on the question of how to classify classification systems, and in particular on how well and comprehensively it is actually possible to accomplish such a task. In attempting to compile a classification of all neurosurgical classification systems, we are faced with a version of a logical paradox first discovered by Bertrand Russell in 1901 as a problem at the core of mathematical set theory. Russell, a polymath who later won the 1950 Nobel Prize in literature, essentially observed that a set of all sets that do not contain themselves is paradoxical. On the one hand, if the encompassing set includes itself, then it violates its defining characteristic (a logical contradiction); on the other hand, if the set does not include itself, then it does not contain all sets that exclude themselves (because the encompassing set is itself such a set, and yet is excluded). This paradox has been framed in many more mundane and intuitive ways. For example, “If the master barber shaves all those who do not shave themselves, and only those who do not shave themselves, who shaves the master barber?” ( Fig. 9.1 ).
Turning from the famous paradox of set theory to the related problem of classifying classification systems, we note that typical neurosurgical classification systems, being concerned with classifying physical objects or phenomena, do not include themselves in the classified material. In this sense, standard classification systems are analogous to theoretical sets that exclude themselves, and to the ordinary barbershop customers who do not shave themselves. But what about a classification of classification systems? Should it classify only existing classification systems, or should it be subject to its own classification scheme? Should it be subject to its own rules? Should it classify itself? We believe it should. A classification of classification systems should be structurally consistent with its own classification rules, just as consensus guidelines for systematic reviews and meta-analyses can be applied to review reviews and meta-analyze meta-analyses. This structural choice is deliberate, and we recognize that it makes the classification fundamentally different from ordinary neurosurgical classification systems (which cannot be used to classify themselves).
Modern Statistical Methods, Technology, and Applications
Developments in statistics and computer science in the late 20th and early 21st centuries have brought new methods to bear on problems of classification. The urge to classify is perhaps stronger today than it has ever been, and it will only become more so. In an era that has seen explosive growth in scientific data, to the point of rendering almost quaint the notion of “big data,” classification finds and imposes structure and enables efficient searching, labeling, indexing, and retrieval of information. The hash function, which is so ubiquitous in computer science, is a form of rigorous and mathematically defined classification.
Machine Learning
Contemporary machine learning techniques have highlighted a dichotomy between two types of classification: supervised and unsupervised . Supervised learning is hypothesis-driven and externally organized, and it facilitates automated classification according to extrinsically imposed labels. Unsupervised learning is often non–hypothesis driven, and it attempts to automate classification according to intrinsic properties of available data.
In supervised learning (classification) systems, a classifying algorithm is developed from initial “training” data. The training data are provided together with their class labels to facilitate development of the classification system. In a classic example, an algorithm being developed to classify Internet photographs is told which photographs in the training set contain cats and which do not. Once “trained,” the classifying algorithm is then used to classify new data in an automated fashion (the classic example distinguishes even between cat and non-cat photographs it has never seen).
By contrast, in unsupervised learning (classification) systems, an algorithm sorts data into classes without direct input from a “training” phase, purely on the basis of similarities among shared features in the data. These “unsupervised” systems identify data “clusters” on their own. To those more familiar with the world of traditional cell biology, this process is analogous in some ways to fluorescence-activated cell-sorting techniques in flow cytometry: the familiar “dot-plots” generated by cytometry display data “clusters” corresponding to cell types (or artifacts), and suggest strategies for sorting (classifying) cells into well-defined populations; the classification process sometimes reveals and defines a previously unidentified population. In a more direct, and perhaps nowadays familiar example of unsupervised classification, modern software may automatically sort your photographs according to the identities of individuals in the images. Early versions of this software ask you to label the subjects of the photographs; newer versions sometimes apply the name labels without user input.
Statistical methods have coevolved with classification processes to quantify aspects of both the process and the results of classification and to address the following questions: How reliable is automated classification? How much error (misclassification) can or must be tolerated? These and related questions become all the more salient when the process of classification is fully automated. However, a salient question remains: Whom do we trust to define and perform the classification process?
Contemporary classification processes in medicine are evolving to adopt all of these techniques. Legacy systems, which remain predominant and, of course, highly useful, constitute the majority of those classified here.
You're Reading a Preview
Become a Clinical Tree membership for Full access and enjoy Unlimited articles
If you are a member. Log in here