div#ContactForm1 { display: none !important; }
Hyper Smash

IONM Special Editorials

This Special Editorial column will host writings, reviews and invited guest Editorials from some of the top neurophysiologists contingent upon my invitation or proven quality of a work that may need wider exposure and some stirring up of minds in the field of neuromonitoring. The goal is simple and straightforward "Where do the IONM stand?, meaning where do the Intraoperative Neuromonitoring heading", what is the current status and future prospects?.
Dr. Joe Salamy, a doctorate in neuroscience field and one of the well known Neurophysiologists is writing the very first special Editorial. Dr.Joe is one of the finest and experienced members of the intraoperative neuromonitoring that many of you modern day IONM technologists and neurophysiologists would not have heard much, because he is not the self promoting, stage talking and conference bragging type but he is a practitioner and mentor, he trained several techs and Neurophysiologists engaged in IONM in the state of California and Nevada. He perhaps performed thousands of different spine and brain surgical procedure monitoring in the last couple of decades. Just retired few months ago, until his retirement, he himself performed IONM insider ORs and occasionally supervised techs by remote monitoring.  His knowledge and understanding is uniquely powerful and, an outstanding Neurophysiologist of our times.
Editorial Special:

IONM: Neurological Testing or Neurophysiological Divining? 
J. G. Salamy, Ph.D.  
VERTECz: surgical neurophysiology, Las Vegas, Nevada 
2012 All Rights Reserved
Email: jgsalamy@gmail.com
During the last three decades IONM has followed an objective probability-based clinical model which focused attention almost exclusively upon the detection of specific events.  IONM practices, however, do not neatly conform to those of conventional clinical testing procedures.  It is suggested herein that future efforts be directed toward developing additional models which elucidate the dynamic and subjective qualities of IONM and recognize the importance of sequences and their influence upon decision making.  Heretofore, the temporal and linguistic aspects of IONM have been largely ignored.  It is proposed that we examine IOMN as an ongoing interactive process, and develop new tools to help accomplish this task.
The 50/10-Rule
Brown, et al.,1 state in their seminal 1984 paper, “In this comparison (to a pre distraction reference average), decrease in peak-to-peak amplitude in excess of 50% relative to the control records is generally taken as a warning of the possible onset of a neurologic compromise, as is an increase in latency of more than 3 msec.”  This general guideline, offered in the then fledgling field of SSEP neuromonitoring, subsequently took on the mantle of ‘risk criterion,’ and thenceforth was dubbed the ‘50/10-rule’ (decreased amplitude by 50%, increased latency by 10%).  Imbued with the authority of a ‘RULE,’ 50/10 has, for some, become an axiomatic prescription for issuing an SSEP warning.

The premise that the status of central nervous system can be expressed by this simple ratio warrants consideration of some inherent practices and assumptions of intraoperative neurophysiology.  It is customary in routine clinical diagnostic SSEP testing to compare an acquired response to a previously obtained distribution of waveforms and thereby determine wherein a value falls. Much of intraoperative neurophysiology is a direct generalization from the Evoked Potential Clinic.  It is therefore often tacitly assumed that in the OR, we are, in effect, conducting Medical procedures tantamount to those performed in the Clinic.  Medical Tests are themselves replete with assumptions primarily derived from a strictly probabilistic model.  Intrinsic to probability theory is the notion that every event has a number attached to it: the probability that the event will occur (Kosko, 1993).2 The likelihood of incidence confers its significance. 

Thus the question becomes, “How improbable is the occurrence of 50/10?”  First, it must be decided how the 50/10-rule is to be calculated, i.e., 50% down from where?  Brown, et al (1984) explicitly selected a pre-distraction reference as the ‘control’ average, and specified the peaks of interest.  Is this comparable to clinical testing?  The intraoperative SSEP literature devotes remarkably little attention to the procedure for establishing an appropriate baseline for subsequent comparisons.   Is it sufficient to select a particular average and simply proclaim it to be the baseline?  If so, which average?  One could collect several averages and choose: a) the one with the “best” morphology, b) the “worst” morphology, c) the one that mostly resembles the ideal from the literature, d) calculate a grand average, e) randomly select an average, e.g., the 15th epoch acquired, or f) use another measure of central tendency such as the median response, etc. Next, the decision of which peaks from which channel(s) to measure must be made.  Finally, when is the best time to set the baseline?  Should it be chosen immediately following positioning prior to incision, during exposure when anesthetic levels are equilibrating, or right before decompression, instrumentation, etc.? What about an anterior-posterior procedure?  Is the anterior baseline suitable for the posterior phase, too?  It would seem that in order to impute a causal connection between an EP change and a surgical event precise calibration of “before” and “after” responses is required.

While the process of setting a “baseline” appears analogous to procedures used in routine clinical testing there are marked practical and conceptual differences between the Clinic and the OR.  In principal, clinical SSEP tests are conducted upon an individual patient whose averaged waveforms are compared to those of a neurologically unimpaired sample. The procedure is objective, standardized, and static.  Clinical tests work because measuring large numbers of individuals on any given characteristic generally produces a normal distribution of measurements.  Neuromonitoring, on the other hand, is performed in a quasi-objective, semi-quantitative, atheoretical manner.  In contrast to the clinic, the OR neurophysiologist is processing voluminous data generated in real time within a single subject, which are oft times asymmetrical and highly dynamic. The patient must serve as his own control and repeated measurements, which are not normally distributed, are made in the presence of continual anesthetic flux, homeostatic adjustments, vascular modifications, and surgical manipulations, all of which influence the quality of EP recordings. The OR neurophysiologist must deduce limits of variation in data trains from multiple channels and multiple modalities as they unfold.  In sum, in the Clinic the ‘probability distribution’ is handed to the diagnostician. In the OR the ‘probabilities’ must be inferred “on-the-fly.” That which may be regarded as ‘error’ or ‘noise’ in clinical data is the mainstay of intraoperative data.  Thus, the Clinic deploys a strictly norm-referenced strategy, whereas the OR must defer to a self-referenced approach.  Repeated measurements over time, within the same subject, entail serial dependency, which thereby precludes routine statistical testing (means, standard deviations, t-tests, etc.), which insists upon the independence of error components. 

Change, Probability, and Significance

            The 50/10-rule, like the clinical SSEP test, relies exclusively upon a “significant” difference in level from the reference sample.  The OR neurophysiologist cannot be content with a mere change in level, but must assess events as they unfold over the course of surgery.  In order to draw meaningful conclusions, patterns must be extracted, and event sequences delineated.  Toward this end, time-series analyses offer a wide range of methodological options (see Gottman, 1981).3 Time-series procedures, however, have yet to be exploited in routine neuromonitoring situations.  Box and Jenkins (1970)4 elaborated particular procedures called Autoregressive Integrated Moving Averages (ARIMA) models, which specifically correct for the presence of serial dependency, and thus allow proper application of the t-statistic.  ARIMA models permit significance tests to be made for differences in ‘level’ between pre- and post intervention data, for the presence of ‘deterministic drift’ in the data, and for changes in drift from pre- to post-intervention data (see Glass et al., 19755; Jones et al., 1977).6

Level is a term applied to autocorrelated data as the mean is used to describe central tendency in uncorrelated data.  A change in level refers to a change occurring at the point of intervention and represents a discontinuity from one phase to another.  Deterministic drift is simply a persisting trend or linear slope in the observations over time. Thus, the significance of the difference in the rate of increase or decrease from one phase to the next can be evaluated.  Change in drift refers to a deviation in slope occurring at each intervention point (see Jones, et al., 1977). See fig. 1.
While providing appreciable information, time-series analyses require prodigious computational algorhithms well beyond the capacity of most commonly used intraoperative monitoring systems.  Less sophisticated alternatives, which also compare data over time for separate phases are available, however.  One such procedure, the split-middle technique (White, 19727, 19748), allows examination of trends within and across phases of data collection (e.g., exposure vs. decompression).  Although the split-middle technique was developed to assess the rate of behavioral changes (frequency/time), it is also applicable to discrete categorization and may thus prove useful to the OR neurophysiologist.  The split-middle technique estimates the trend, referred to as the line of progress, or celeration line.  The celeration line predicts the direction and/or rate of change based upon median values in pre-determined quadrants.  Inferential statistics can be applied once the split-middle lines have been determined (see Kazdin, 1982).9 The null hypothesis posits no change across phases.  The baseline phase is therefore an accurate estimate of the intervention phase celeration line.  Given the null hypothesis, the probability of a data point in the intervention phase falling above the projected baseline celeration line is p=0.5.  A binomial test is then used to determine if the number of data points above or below the projected slope is sufficient to reject the null hypothesis (see Kazdin, 1982). 

The assessment of successive data in OR applications can be greatly facilitated by demarcating adjacent segments of surgical activity which entail varying levels of risk, e.g., pre-incision, exposing, distracting, decompressing, instrumenting, etc. Baseline data thus become an actual line rather than a single arbitrarily selected point. Systematic knowledge of profiles within and between surgical phases could provide useful information about trends and changes which might portend or reflect neurological compromise. See fig. 2.

Fig. 2.  Profiles of different surgical phases

Changes in level, drift, and slope can sometimes be discerned in graphed data if they are being sought.  Such information can be useful in assessing the direction of a case irrespective of statistical significance.  Noting unexpected changes can obviate the need for quantitative inference and place the OR neurophysiologist and surgical team on alert.  Visual inspection is not without shortcomings, however. These include: subjectivity of judgments, absence of a measure of reliability, subtle changes might be overlooked, specific rules for decisions are lacking, and no theoretical framework guides interpretation.

Methods need to be developed which elucidate qualitative dynamic processes.  Obviously, an abrupt total loss of EPs requires no statistical tests and inspires little controversy.  Problems arise, however, when responses are highly variable and markedly asymmetrical from the outset.  In the face of this ambiguity new tactics need to emerge.  A major concern for the OR neurophysiologist is the sheer volume of data that must be processed, more or less, instantaneously. One possibility that might make the data more manageable would entail automatic computer implementation of a coding scheme (see Bakeman & Gottman).10  Latency change, for example, could be coded as follows: 1-2 ms=code A, 3-4 ms=code B, 5-6 ms=code C, 7-8 ms=code D, etc.  This simple coding scheme could provide possibly important frequency and relative frequency information (rates of occurrence) about events of interest.  The patterns of such codes could then be used to assess anesthetic and/or intervention effects.  Moreover, plots of code sequences could help Remote Readers follow one or more cases effortlessly.  Rates, frequencies, probabilities, and percentages of event occurrences can easily be determined and subjected to routine statistical procedures via simple contingency tables (e.g., code by surgical phase).  Observed frequencies could be compared to theoretically “expected” frequencies or to those derived from other similar surgeries and chi- square or z-score binomial tests applied.  The issue of stochastic dependency arises again, however.  It could be argued that successive determinations (codes) are made independent of each other.  That is, one code is not influenced by previous codes assigned.  Thus, routine statistical tests could be used to infer significance.

In considering the ‘importance’ of a finding it is necessary to stress the distinction between statistical and clinical significance.  The former alludes to the improbable, while the latter denotes appreciable.  Obviously, the criterion for each is quite different.  The statistical procedures discussed above are primarily designed to evaluate an “experimental criterion,” namely, has a veridical change been demonstrated, and can it be attributed to the intervention (e.g., decompression, instrumentation, etc.).  We generally assume that changes beyond the statistically expected are also clinically relevant, but that requires the demonstration of new post-surgical deficits. Just as IONM procedures diverge from routine clinical testing conventions, however, so too do they depart from standard experimental methods.  Moreover, IONM data may not satisfy the assumptions of the General Linear Model.  

Subjectivity and Fuzzy EPs

            The statistical suggestions above, while addressing the serial nature of neuromonitoring, are nevertheless still tied to a probabilistic model.  Regardless of how sophisticated the statistical analysis the results are meaningless without immediate interpretation of well-trained experts to make sense of the findings.  Human intelligence, however, entails subjective, imprecise, and non-quantitative reasoning.  These attributes are ignored by the classic probability based model, yet they account for a great deal of the OR neurophysiologists job.  Fortunately, there are other means of quantifying uncertainty.  Lofti Zadeh, (1965)11 observed that most classes or collections of objects encountered in the “real world” have uncertain borders.  Despite imprecision, human communication is not impaired.  Zedeh introduced the concept of fuzzy sets and fuzzy logic to deal with ill-defined class membership.  Rather than ask the probability of an event’s occurrence, he asked, “What is the degree of membership in a given class?”  Thus, a fuzzy set is not statistical.  Fuzziness is not randomness.  Zedeh’s fuzziness represents vagueness due to human intuition, not probability.  The proclamation to a surgeon that the evoked responses are “down” is not understood in terms of likeliness. Nor does it necessarily signify transgression of a mystical 50/10 line.  Most often, the surgeon’s reply is, “What does that mean?” 

What does “down” mean?  The neurophysiologist has been studiously looking for a hint that a change in the pattern of neural excitation has occurred and that it is linked to surgical activities.  Of course every change could be reported.  That, however, would produce an inundation of false positives, which renders neuromonitoring useless.  The knowledge that is used to determine that a response is ‘down’ is often subjective, uncertain, and the basis for that conclusion unspecified, as are the antecedents and consequences of the event.  Moreover, the judgment is communicated in vague everyday language.  Its meaningfulness, however, is not misunderstood, and the attitude and activities in the OR change dramatically.  Terms like, “good,” “OK,” “down,” “no change,” etc., are ambiguous linguistic variables.  Just as probability theory is used to model random uncertainty, fuzzy logic is used to model linguistic uncertainty. 

Returning to our hypothetical coding notion of the previous section, we could apply a fuzzy logic approach instead of a probability based approach to develop ‘if-then’ rules which map EP measurements to linguistic variables.  Suppose in our fuzzy format we agree to call a small increase in latency 1-3 ms, a medium increase 4-6 ms, and a large increase >6 ms. Similarly, we could assign 1-5uv, 6-11uv, and >12uv to small, medium, and large amplitude changes.  The results can then be arranged into a table of rules of correspondence to change as follows:

An advantage of the fuzzy approach is that the neurophysiologist need not be constrained by values of milliseconds and microvolts.  The evoked potential is much more complex than individual peaks or troughs.  The architecture of the entire waveform often provides the first indication of impending change, notwithstanding specific latency and amplitude measurements.  Thus, rather than treat the EP as though it represented a single ‘telephone’ line, we can recognize its complexity and be aware that numerous synchronized neural elements contribute to its construction.  Subtle alterations in organization which may involve non-neural and vascular components may precede changes in specific peaks and valleys.  The fuzzy approach offers a method by which morphological changes can be estimated.  Various distortions of waveforms can be articulated and classified using experts and ‘information mining’ techniques to establish a “footprint of uncertainty” (see Mendel, 2001).12 

We have advised caution with respect to the 50/10 – one size fits all- strategy.  It may be inappropriate due to large within subject variability, pre-existing deficits, idiosyncratic non-linear reactions to anesthetics, persisting trends in measurements over time, and the complexity and dynamic properties of the EP waveform may not be distillable into one or two simple parameters.

The OR neurophysiologist’s task is not that of a 50/10 detecting instrument, but rather one of deciphering latent patterns in a complex array of unfolding electrophysiological activity from which statistical regularities must be inferred.  They must be astute, dedicated observers, and must extract relevant neural ‘signatures’ despite the vicissitudes of anesthetic and surgical activities. The 50/10-rule, by focusing attention upon a single feature may hinder recognition of vital trends in the data that portend critical changes.  This feature, however, is part of the process, and considered in context, should not be summarily dismissed.  It is a useful guide, but not an immutable rule. The OR neurophysiologist does not adhere strictly to the protocols of the clinic or the research laboratory.  Nor do they enter the surgical theatre armed only with a neurophysiological divining rod.  Rather, fortified with the prodigious integrative capacity of the human brain’s visual-perceptual system, they make painstakingly careful observations which can guide surgical decisions and prevent catastrophic iatrogenic injury. The systematic cataloging of these observations and the elucidation of patterns in the data stream can provide the stimulus for innovation in research and technological development necessary for the advancement of our field.

1.      Brown, R.H., Nash Jr., C.L., Berilla, J.A., and Amaddio, M.D. Cortical evoked potential monitoring: a system for intraoperative monitoring of spinal cord function. Spine vol. 9, num. 3, 1984.
2.      Kosko, Bart, Fuzzy Thinking, Hyperion, NY, 1993.
3.      Gottman, J.M. Time-series analysis: introduction for social scientists. New York, Cambridge, University Press, 1981.
4.      Box, G.E.P., & Jenkins, G.M. Time series analysis: Forecasting and control. San Francisco: Holden-Day, 1970.
5.      Glass, G.V., Wilson, V.L. & Gottman, J.M. Design and analysis of time series experiments. Boulder: Colorado Associated University Press, 1975.
6.      Jones, R.R., Vaught, R.S., & Weinrott, M. Time-series analysis in operant research. J. appl. Behavior Analysis, 1977, 10, 151-166.
7.      White, O.R. A manual for the calculation and use of the median slope – a technique of progress estimation and prediction in the single case. Regional Resource Center for Handicapped Children. University of Oregon, Eugene, Oregon, 1972.
8.      White, O.R., The “split-middle” a “quickie” method of trend estimation. University of Washington, Experimental Education Unit, Child Development and Mental Retardation Center, 1974.
9.      Kazdin, A.E. Single-case research designs: methods for clinical and applied settings. New York, Oxford University Press, 1982
10.   Bakeman, R. & Gottman, J.M. Observing Interaction: an introduction to sequential analysis New York, Cambridge University press 1986.
11.   Zedeh, L.A. “Fuzzy Sets.” Information and control 8 1965, 338-53.
12.   Mendel, J.M. Uncertain Rule-Based Fuzzy Logic Systems Prentice Hall PTR, Upper Saddle River, NJ. 2001

Q & A: Do you have a question or clarification, or Do you want have a positive feedback or critic, do not hesitate to contact the author, write to the Author Dr Joe at: jgsalamy@gmail.com, and readers can also write to the Blog author and Editor: brainspinellc@drmuni.com

Creative Commons License
IONM: Neurological Testing or Neurophysiological Divining? by Dr.Joe Salamy, Neurophysiologist is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 United States License.
Based on a work at VERTECz: surgical neurophysiology, Las Vegas, Nevada.
Permissions beyond the scope of this license may be available at drmuni.com.   

Note: Special Editorial column, a Seperate Blog will also be created for this purposes to let my colleagues and top Neurophysiologists participate and contribute. 

1 comment:

  1. This was very helpful and insightful. Easy to read and understand. Statistical reasoning is a method of reasoning in the context of uncertainty and incomplete knowledge.

    statistical analysis using the fuzzy approach