Independent calibration of perturbational complexity for unresponsive patients

Paper: Silvia Casarotto and colleagues, “Stratification of unresponsive patients by an independently validated index of brain complexity.” Annals of Neurology 80(5), 718–729. DOI: 10.1002/ana.24779. Complete article at PMC. Read and summarized 29 September 2026.

Reading status: complete for this article and its identified written scientific supplement. Read the entire archived PMC article, including Methods, results, discussion, all captions, main Table 1, references and declarations. Visually inspected every panel of Figures 1–4 and the complete main table. Retrieved the sole written supplement, ANA-80-718-s001.docx, through the Europe PMC supplementary package; read all of Table S1, its 81 patient rows and legend, and visually inspected all three pages of a locally rendered PDF. The supplement contains no additional figures or extended Methods. The PDF is a reading derivative, not a publisher-supplied PDF. No referenced papers were opened as part of this reading. Raw data and analysis code were not reanalyzed. Source inventory and hashes.

What the study establishes

The authors calibrate a TMS–EEG perturbational complexity index against reports in healthy people and communicative brain-injured patients, then apply the resulting threshold to patients with disorders of consciousness. This separates calibration from the uncertain consciousness status of the unresponsive patients. The threshold identifies 36 of 38 patients diagnosed as minimally conscious and separates the 43 patients diagnosed as vegetative state/unresponsive wakefulness syndrome (VS/UWS) into groups with absent, low-complexity or high-complexity responses.

The important result is a promising transport of a report-calibrated physiological measure into a clinically difficult population. It is not a demonstration of 100% diagnostic accuracy for actual experience in unresponsive patients. The benchmark's perfect separation is measured on the data used to select the cutoff; the patient cohort supplies a test of sensitivity to behavioral MCS and a physiological stratification of UWS, where experience itself is unobserved.

The benchmark and what “independent” means

The benchmark contains 150 unique participants: 102 healthy people and 48 conscious brain-injured patients. The latter group comprises five with locked-in syndrome, 16 with subcortical stroke, 18 with cortical stroke and nine who had emerged from MCS and recovered functional communication. Including communicative injured patients reduces one obvious mismatch between a healthy benchmark and the intended clinical population.

All 102 healthy participants were measured awake. Forty-eight also contributed unresponsive conditions. Those condition counts total 50 because observations are not all independent people:

Thus the analysis has 200 subject-condition maxima, derived from 540 individual stimulation-site measurements, rather than 200 independent participants. There are 164 maxima in the immediate/delayed report group and 36 in the no-report group. The report categories are the study's provisional consciousness labels, not independently observed experiences. In particular, a delayed report also depends on memory and the accuracy of later reporting. The authors explicitly acknowledge forgetting and confabulation.

ROC analysis selects an empirical cutoff PCI = .31*. The benchmark achieves reported AUC, sensitivity and specificity of 100% under these labels. The paper does not describe a held-out benchmark partition or cross-validation for selecting and evaluating that threshold. The same benchmark supports both operations. The subsequent DOC cohort is distinct from the benchmark, so the cutoff is independent of its behavioral diagnostic labels; “independent” should not be expanded into independent investigators, an entirely new cohort relative to earlier PCI publications, or a prospective external replication. Forty-seven benchmark participants and 12 DOC patients had already appeared in earlier publications, including Casali et al. (2013).

The study broadens the earlier PCI evidence by including disconnected but retrospectively reportable experience, heterogeneous conscious brain injury and a larger clinical application. It does not repair the still-open reading gap in the Casali 2013 scientific supplement, nor does it reproduce every algorithmic detail delegated to that earlier paper.

The marker includes the acquisition and sampling protocol

PCI here quantifies the compressibility of a statistically thresholded pattern of TMS-evoked cortical source activity. The analysis uses Lempel–Ziv complexity normalized by source entropy. The intended interpretation is that a differentiated response that also propagates through a system expresses both differentiation and integration. This is a physiological operationalization, not a direct computation of IIT's intrinsic integrated information, Φ.

The protocol uses 60-channel EEG, individual MRI acquired within the preceding week, and neuronavigated stimulation of eligible bilateral frontal BA6/8 and parietal BA7 targets. Investigators avoid visibly lesioned cortical targets and sites compromised by skull breaches or drains. Earplugs deliver masking noise. Stimulation starts around 120 V/m and may increase to 160 V/m if an adequate early response is absent. At least 200 trials are acquired at an accepted site; unavoidable muscle or magnetic artifacts cause abandonment of that site. Visual inspection rejects bad trials/channels, followed by ICA. Recordings with fewer than 80 good trials or more than 10 bad channels are excluded. Subsequent filtering, downsampling to 362.5 Hz, source estimation and statistical processing are automated according to the earlier PCI procedure.

If fewer than 1% of spatiotemporal activations survive the statistical procedure, the algorithm sets PCI to zero. A zero is therefore an operational result under this acquisition and significance pipeline, not proof of literally absent neuronal activity everywhere in the brain. The authors explicitly state that absolute PCI values depend on the acquisition and analysis protocol.

The diagnostic variable is PCImax: the maximum over the eligible measurements in that individual and condition. This aims to detect preserved complex responses even when other sites are damaged or uninformative, and parallels the use of a patient's best behavioral score. It is not a typical-site or mean-brain PCI. In awake benchmark measurements, 13 of 461 individual-site values fell in the low range despite every subject's maximum exceeding the cutoff. In MCS, 37 of 112 individual-site values were below threshold despite 36 of 38 patients having a positive maximum.

Site coverage is consequently part of the marker definition. The patient dataset contains 254 site measurements: a mean of 3.30 per UWS patient and 2.95 per MCS patient. Table S1 documents two to four sites in most patients, but only one in patient 33. The study should not be described as uniformly testing four sites in every patient. A future implementation must specify eligible sites, quality control, stimulation intensity, how many attempts are permitted and the maximum-selection rule before transferring the numerical threshold.

The clinical test population and reference labels

The 81 patients comprise 38 MCS patients—17 MCS+ and 21 MCS−—and 43 VS/UWS patients. Etiologies include anoxic, traumatic and vascular injury. Testing occurred at least 20 days after onset and at least three days after withdrawal of sedation. Patients with medical instability, refractory generalized seizures or specified neurological/psychiatric histories were excluded.

The behavioral reference is four CRS-R assessments over one week. Any observed nonreflexive behavior supports MCS; patients showing only reflexive behavior across all assessments are classified VS. The best observed score is retained. TMS–EEG occurred in the same week, during daytime, with eyes open spontaneously or through stimulation; drowsiness prompted interruption and arousal facilitation. Repeated behavioral assessment is valuable, but measurements from the same week do not establish that every patient's conscious state was identical during behavioral testing and during TMS.

The article and supplement do not report diagnostic, EEG-classification or outcome-assessor blinding. Later analysis is described as automated after preprocessing, which does not establish that acquisition choices, manual artifact rejection or behavioral assessments were blinded. Nor is a prospective threshold registration or a detailed recruitment/exclusion flow supplied. These are reporting limits rather than proof that the corresponding safeguards were absent.

Sensitivity to MCS: 36/38, or 94.7%, have PCImax > .31. Both lower-scoring patients are MCS−. Table S1 identifies a subacute vascular patient with PCImax .29 and a chronic traumatic patient with .27; both have severely abnormal EEG and vigilance score 1. All 17 MCS+ and 19/21 MCS− patients exceed threshold. Low PCI therefore fails to rule out even behaviorally demonstrated consciousness in this dataset. A nonsignificant MCS+/MCS− difference does not establish equivalence.

The conventional EEG assessment provides a useful within-cohort comparison: a threshold between severely and moderately abnormal background patterns detects 81.6% of MCS cases. PCI detects five MCS patients with severely abnormal resting EEG. Figure 3 displays examples across severe, moderate and mild background abnormalities. Comparisons in the discussion to accuracies of other published methods are not head-to-head tests in this sample.

What the UWS stratification means—and leaves unknown

The high-complexity UWS scores range from .34 to .50. Figure 4 contrasts representative absent, stereotyped and differentiated response patterns alongside MRI. Their diagnostic significance is inferred by transferring the benchmark association to severely injured, noncommunicative brains. The nine patients are plausible candidates for preserved capacity for consciousness despite absent behavioral evidence. The paper does not independently establish their current experiences, contents of experience or covert command-following.

Calling the nine either confirmed true positives or confirmed false positives would assume the very fact the study aims to assess. Behavioral UWS is not a definitive unconsciousness label. Accordingly, patient specificity, positive predictive value and negative predictive value for actual experience cannot be computed from these data by importing the benchmark's 100% values. Nor can the lower-complexity UWS groups be declared certainly unconscious, especially given the two MCS false negatives relative to the behavioral reference.

The UWS groups do not differ significantly in best CRS-R score or the reported vigilance comparisons, but absence of a detected difference is not proof of equality. Their markedly different injury etiologies and EEG profiles also matter when interpreting the outcome association.

Follow-up is a separate outcome

At six months, six of nine high-complexity UWS patients had transitioned behaviorally to MCS, with one outcome unknown; five of 21 low-complexity patients had made that transition, with two unknown. None of the 13 no-response patients improved. These counts preserve the original group denominators and the missing outcomes; they should not be presented as though every outcome was observed.

This is evidence of an outcome association in these groups. It is not proof that the high-complexity patients were experiencing something at the baseline measurement, and it does not establish that raising PCI would cause recovery. No prespecified multivariable prognostic model or independently validated prognostic performance is reported. The authors themselves emphasize interpretation as current capacity for consciousness rather than as a prognostic marker. Table S1 contains baseline clinical and measurement details, not individual follow-up trajectories.

Causal and connectomic interpretation

TMS makes this stronger than a purely spontaneous correlation for studying evoked neural propagation: the brain's response follows an imposed cortical perturbation. It does not make the association between the resulting complexity score and consciousness a demonstrated causal relation. The study does not manipulate complexity while holding other determinants fixed, establish complexity's necessity or sufficiency for experience, or identify a uniquely consciousness-producing circuit.

The discussion proposes several explanations for lower complexity, including injury-related loss of effective interactions and state-dependent bistability. Those are mechanistic hypotheses here. Individual MRI guides stimulation and identifies lesions; the study does not reconstruct each patient's synaptic connectome or measure all thalamocortical causal interactions. It illustrates why anatomy, dynamic state, intervention and measurement conditions jointly constrain inference.

For a neural simulation, predicting these perturbation responses would be a useful target. A defensible comparison would reproduce the intervention, signal-generation/EEG observation process, preprocessing, source estimation, significant-activation matrix and maximum-over-sites procedure, and test new responses not used to fit the model. Matching only a scalar PCImax could conceal major waveform or spatial errors. Applying .31 to a different simulated signal or recording resolution without recalibration is unsupported by this study. Even accurate prediction of the biological marker does not by itself establish consciousness in the simulation, preservation of an individual's memories or personal continuity.

Reading audit and source caveats

• Main Figure 1: individual-site versus maximum scores, benchmark ROC and report contingency; Figure 2: patient maxima and CRS-R relationships; Figure 3: EEG categories and representative MCS evoked responses; Figure 4: the three UWS response groups. All panels and captions were inspected.
• Main Table 1: all condition counts and PCI/PCImax distributions were read as text and visually from a faithful rendering of the archived HTML table. Table S1: every patient row and its definitions were read as text and visually from the DOCX rendering.
• The main text contains an unresolved demographic inconsistency: Methods give 63 healthy females, whereas the Results sex comparison uses 45 female and 57 male healthy participants. This summary does not silently reconcile those figures.
• Direct PMC/Wiley supplement routes failed or returned non-document responses; one retry of the Europe PMC package resolved access before substantive reading. No failed HTML response is retained under a document extension.
• The main article delegates older algorithmic and sleep/anesthesia protocol details to references. Reading those references is outside this paper's identified supplement scope. The separate Casali 2013 supplement remains explicitly unread and unacquired in its own record.