The Glasgow outcome scale in vegetative state: A possible source of bias

Artificial intelligence procedures (classification and decisional trees) [1–3] identified four significant neurological signs allowing early, reliable prognosis in a large (n= 333) sample of vegetative state (VS) inpatients [4]. Patients were admitted to a one dedicated intensive care unit (1998–2006), after at least 2 weeks from brain injury, and were monitored by standardized clinical criteria. The significant signs were: recovery of spontaneous motility, eye tracking and oculo-cephalic reflex, disappearance of oral automatisms. The approach proved efficient; the model mean accuracy across the observation time points was 79% [4].

At closer scrutiny, however, the overall error of the model matrix (ranging from 26.7% at T₀to 16.8% at T_{180 days}) was not evenly distributed across the patients’ aetiologies (post-traumatic or non-traumatic) or Glasgow Outcome Scale (GOS) classes. Instead, portions of error rating as high as 82% in the post-traumatic patients’ sample at T₀(with a monotonic, asymptotic decrease in this patients’ group as a function of time to 50% at T_{180 days}) were accounted for by a discrepancy in the criteria of attribution of the GOS class 3. Independent from aetiology, subjects classified as GOS3 by conventional clinical criteria were allocated by the decisional trees procedure to GOS classes 2 or 4–5, with a disproportionate increase as a function of time of the misattributions to GOS4–5; misattributions were less frequent (3–31%) in the other GOS classes. The error was greater for patients in VS due to traumatic brain injury than in those with massive cerebrovascular lesions, prolonged anoxia-hypoxia, etc. The percentages of misattributions at different time points (expressed in days) are reported in Figure 1for both post-traumatic and non-traumatic patients; in each column, misattributions to a better GOS class are in darker colour.

Figure 1. Percentages of misattributions at different time point, for post-traumatic and nonpost-traumatic patients.

The error was not casual, nor should it be understood as indicative of inappropriate methodological approach or data treatment [4]. Instead, the misattributions by artificial intelligence to GOS classes 2 or 4–5 rest mainly on the association of items assessing vigilance or movement among the criteria of attribution to GOS class 3. In a pathophysiological condition such as the VS, in which severe impairment or deep loss of consciousness is the main feature and recovery of consciousness the significant indicator of outcome, recuperation of voluntary movements may have lesser relevance than in other conditions. In this regard, GOS may become a possible source of bias and caution is advisable when using it for prognostic evaluation of patients with severely impaired vigilance, as in the case of VS.