Response to: “Questioning the evidence for BCI-based communication in the complete locked-in state”

摘要

Patients in completely locked-in state (CLIS) have no means of communication and present a highly challenging and daunting problem for the neuroscientist [ 1 – 3 ]. Until today, few groups have attempted to solve this problem, and only some have reported success in advancing the goal of providing a means of communication to patients in CLIS [ 4 – 7 ]. In his commentary, Dr. Spüler raises doubts about all the research efforts towards this goal but primarily about the results published in 2017 by Chaudhary and colleagues. Dr. Spüler bases the commentary on 2 main calculations: Absence of hemodynamic differences between “yes” and “no” thinking and Chance-level classification across all the sessions in the 4 published cases with CLIS. In this commentary, we address the issues raised by Dr. Spüler. 1. Absence of hemodynamic differences between “yes” and “no” thinking In his commentary, Dr. Spüler claims that, in the paper by Chaudhary and colleagues [ 6 ], the change in the concentrations of oxy-hemoglobin (O _(2)Hb) acquired from 20 different functional near-infrared spectroscopy (fNIRS) channels were averaged, and then further averaging was performed across trials and sessions. Chaudhary and colleagues [ 6 ] presented the averaged change in relative concentration of O _(2)Hb separately for each of the 20 channels used during the study, as shown in their Fig 1 [ 6 ]. In none of the fNIRS literature published to date have fNIRS channels placed across such disparate regions been averaged [ 8 – 12 ]. The reason behind not averaging the channels is the fact that different channels represent metabolic information from the respective underlying brain region. In Chaudhary and colleagues’ paper [ 6 ], therefore, first signal acquired across different trials—i.e., “yes” and “no” thinking—were averaged separately for the different channels and were then averaged across sessions as shown in Fig 1 of Chaudhary and colleagues [ 6 ]. Fig 1 of Chaudhary and colleagues’ paper shows the averaged relative change in O _(2)Hb from all the 20 channels; if all 20 channels were averaged, then we would have had just 1 time-series of relative change in O _(2)Hb and not 20 different time-series of relative change in O _(2)Hb, each corresponding to a channel, as depicted by Chaudhary and colleagues. To further elucidate the difference in hemodynamic response between “yes” and “no” thinking, general linear model (GLM) analysis was performed as shown in S1 Text . Dr. Spüler’s claim of a lack of difference between the 2 response categories “yes” and “no” thinking is thus unfounded and not comparable to that of Chaudhary and colleagues. As reported by Chaudhary and colleagues, channels were treated separately for classification and model building for online feedback session as written on page 18 and 19 of Chaudhary and colleagues’ paper. According to Chaudhary and colleagues (page 18), “The mean of relative change in O _(2)Hb across each channel was used as a feature to train the SVM model through a 5-fold cross-validation procedure.” On page 19, Chaudhary and colleagues further state that, “During an online feedback session, fNIRS data acquired online corresponding to each ISI was processed to obtain the relative change in O _(2)Hb, as described above, across all the channels. The mean of the relative change in O _(2)Hb across all the channels was used as test feature to map onto model space.” 10.1371/journal.pbio.3000063.g001 Fig 1 Bar graph of offline classification accuracy results obtained from the sessions performed by the patient B and published by Chaudhary and colleagues [ 6 ]. (A) The offline classification accuracy of sessions performed by patient B using the method reported by Chaudhary and colleagues, i.e., the mean O _(2)Hb response was used as input feature for the linear support vector machine classifier. (B) The offline classification accuracy of sessions performed by patient B using the method suggested by Martin Spüler. (C) The offline classification accuracy of sessions performed by patient B using the CSP for feature extraction and using linear SVM classifier. CSP, common spatial pattern; O _(2)Hb, oxy-hemoglobin; SVM, support vector machine. 2. Chance-level classification across all the sessions Spüler raises doubts about the classification results based on the method he employed to calculate the offline classification accuracy of each session. It is well known in the machine learning literature that application of different machine learning algorithms and features results in different outcomes, as is obvious from the result presented by Dr. Spüler. We can argue on the method that can and should be used for classification, but that does not invalidate the results presented by Chaudhary and colleagues [ 6 ]. It has also been argued that the sessions should be combined randomly to build a model; we argue that such a method might be valid for stable and invariant data but might be completely misleading for patients in CLI