Diagnosing spells

In the current era of technological advances in medicine, public interest seems focused on advances in laboratory testing, imaging, and surgical instrumentation. Modern-day expectations in the diagnostic part of medicine appear to demand answers which are instant, specific, and without ambiguity. This is consistent with the idea of developing the science of medicine. However, many still consider obtaining a diagnosis through a careful and thoughtful history as an example of the art of medicine.

Clinicians have supported the notion for decades that a properly taken history is far superior to diagnostic testing in establishing the cause of spells such as syncope. Extensive testing which is frequently performed in many patients has been considered to be highly expensive, inappropriate, and unnecessary in evaluating patients presenting to the emergency department with loss of consciousness because it still does not lead to a definitive diagnosis.¹ Blood work, imaging, and physiologic testing will continue to play some role in diagnostic evaluation, but history taking and testing will still leave many of these patients without an identifiable cause for their spells. This is further complicated by the fact that videoEEG (VEEG) and cardiac monitoring which may be considered the gold standards of diagnosis are only able to record the events in a small subset of this patient population.

VEEG recording has been used extensively to evaluate patients with medically intractable epilepsy as part of presurgical evaluations. Highly sensitive and specific descriptive characteristics of seizures from different brain regions have been published to guide the clinician in their history taking.² Patients with less frequent spells are less often recorded to enable a definitive diagnosis leaving many with diagnostic uncertainty and a lower quality of life.

In this issue of Neurology: Clinical Practice, Wardrope et al.³ address the important question of the role of machine learning as it applies to making a diagnosis in cases of transient loss of consciousness (TLOC). They identify a great need because over half of the population may experience TLOC over a lifetime, and it may be their initial or only contact with the health care system if they present to the emergency department. The authors cite the need for other diagnostic approaches because of the fact that although there are a few symptoms and signs that may be considered to be moderately specific for diagnosis, they lack significant sensitivity. In addition, despite the prevalence of TLOC, the clinical decision rules for syncope fall far short of expectations.⁴ Although they do not address all of the possible diagnoses for TLOC, they do establish a framework for differentiating between syncope, epilepsy, and nonepileptic seizures which are considered to be the major causes.

The ability to discriminate between subcategories is a limitation of this study because there appears to be a greater focus on distinguishing between epilepsy and psychogenic nonepileptic seizures based on the content of the questionnaire. Another limitation of this study is that the cases studied with VEEG are probably not typical of the patients presenting after a single episode, which represents the more typical scenario in clinical practice. In addition, the low number of responders to the request to participate in the study and the high number of incomplete questionnaires may limit their usefulness when implemented in real-world scenarios. Recall of historical features from events over several years is another variable that may affect the consistency of the data accrued in this study and therefore its wider application. This is further complicated by the fact that interrater reliability may be low among physicians reviewing VEEG-recorded events.⁵

Simple point-based algorithms using apps have been claimed to differentiate epilepsy from nonepileptic events in up to 87% of patients in underserved areas of the world. The “validation” is based on high agreement with subsequent clinic visits but not confirmed by VEEG.⁶ Machine learning has been tested by comparing clinical data gathered at the beginning of an epilepsy monitoring unit admission to the results of VEEG with high concordance to the prediction of the algorithm.⁷ Neurophysiologic and related biological measures from VEEG-confirmed seizures have been used to detect seizures but have not been applied to making an initial diagnosis to the same extent.⁸

The potential for artificial intelligence to contribute to making the diagnosis of episodic events is promising, but testing carried out to date is relatively limited and considered to be in its relative infancy. With expansion of the classification of different types of syncope and seizures, the likelihood of a one-size-fits-all diagnostic method is unlikely but should not deter the very important research in this field that needs to be performed. Clinicians should not fear that computers will take over the skills of medicine, but those who do not use technology will definitely be at a loss.

Author contributions

Drafting/revising the manuscript.

Study funding

No targeted funding reported.

Disclosure

The authors report no disclosures relevant to the manuscript. Full disclosure form information provided by the authors is available with the full text of this article at Neurology.org/cp.