The research, led by Danilo Strugarevic of the Academy of Applied Preschool Teaching and Health Studies in Kruševac, Serbia, together with Vojkan Nikolic of the University of Criminal Investigation and Police Studies in Belgrade, Dragan Perakovic of the University of Zagreb and Aleksandar Jevremovic of Singidunum University, took an unusually broad view of the testing process. Rather than examining a single signal in isolation, the team combined several distinct streams of information: textual characteristics of the test items, entropy-based measures computed from screenshots of the rendered displays, response times, candidates’ self-reported knowledge, and facial-expression features captured during the exam. The goal was to see whether these process indicators, taken together, could support a more accurate and more honest interpretation of electronic test performance than scores alone.

One of the most technically interesting elements of the study is its use of Shannon entropy, the foundational information-theoretic quantity that measures the unpredictability or information content of a signal. In this context, entropy served as a compact descriptor of how informationally dense a rendered test item appeared on screen. The researchers computed entropy measures such as Entropy_AVG, Entropy_SUM and logEntropy from screenshots of the displays that candidates actually saw. The intuition is straightforward: a visually or textually cluttered item carries more information per unit of screen space, and that density may shape how long a candidate needs to process it. The data bore this out in part. Both Entropy_AVG and QA_Total_Length, a measure of question and answer text length, were positively associated with response time, with correlation coefficients of 0.298 and 0.288 respectively, both statistically significant at p < 0.001.

Yet the relationship between time and performance proved far more elusive. Response time was not significantly associated with the final test result at all, showing a correlation of just −0.029 with a p-value of 0.435. In other words, spending longer on a question neither helped nor hurt the score in any statistically detectable way. This finding cuts against a common intuition in assessment design, where slow responses are sometimes read as a sign of struggle or guessing and fast responses as a sign of confident mastery. The study suggests that response time is a genuine process indicator, sensitive to the structural demands of the item, but it is not a reliable proxy for knowledge.

When the team turned to prediction, the picture became more modest. They compared global and segmented regression models that attempted to predict test outcomes from the combined feature set, and found that predictive performance was limited, with no single algorithm emerging as superior across all evaluation measures. This is an important negative result. In an era when machine-learning vendors promise to read competence out of behavioural traces, the study offers a reminder that the mapping from interaction data to knowledge is noisy and algorithm-dependent. Entropy_SUM and logEntropy showed weak positive associations with the final result, hinting that items with higher informational content may relate in some small way to performance, while Question_Length stood out as the highest-ranked negative structural indicator, meaning longer questions were associated with lower results.

The facial-expression analysis produced perhaps the most nuanced findings of all. Using automated facial-expression recognition, the researchers extracted features corresponding to emotional categories over the course of the test. As standalone predictors, these features had limited value: classification agreement was low and regression correlations were near zero. But a subtler pattern emerged when the team examined timing. Features computed from the first five seconds of each item, covering neutral, disgust and anger expressions, were more informative than averages computed over the full response interval. This suggests that the initial emotional reaction to reading a question, the moment a candidate first registers its demands, may carry more signal than the emotional state sustained while answering. Intriguingly, this early-window advantage did not hold across all expression categories, indicating that different emotions unfold on different timescales during problem solving.

The study’s framing draws on a rich tradition in psychometrics and learning analytics. Classical test theory and item response theory have long recognised that item characteristics shape observed scores, and decades of research on guessing in multiple-choice and true-false formats have shown that partial knowledge and informed guessing can inflate results. What is new here is the attempt to bring information theory into the item-review pipeline. Prior work has used entropy-based algorithms to quantify textual comprehension difficulty and to cluster visual question difficulty, and the present study extends that logic to the actual pixels a candidate sees, treating the rendered display as a measurable object in its own right rather than a neutral window onto the question text.

The multimodal approach also connects to a growing body of work on emotion in learning. Pekrun’s control-value theory of achievement emotions has established that emotional states are not incidental to performance but causally intertwined with it, and earlier studies have monitored facial expressions to gauge learner attention in e-learning environments. The Serbian team’s contribution is to test whether such signals can be harvested unobtrusively, from a standard webcam during a real exam, and to report honestly on how much predictive weight they can actually bear. Their answer, based on the data, is: some, but not much on their own, and only when timed carefully.

Methodologically, the study was conducted at the Department of Health Studies in Ćuprija, Serbia, with ethical approval from the institutional Ethics Committee and written informed consent from all participants, including consent to be video-recorded during testing. The authors note that the underlying datasets, which include performance records, behavioural traces, and image and video material, cannot be released publicly because of privacy and ethical restrictions, though anonymised and processed data may be shared on reasonable request subject to institutional approvals. All four authors contributed equally to the work, with Strugarevic leading conceptualisation and methodology and the team sharing data collection, analysis and interpretation.

The practical implications reach toward the next generation of digital assessment platforms. The authors argue that jointly considering item-structure measures, screenshot-derived entropy, response time, self-reported knowledge and facial-expression features could support smarter testing environments that combine conventional scores with interpreted process indicators, helping item-quality review and potentially helping to distinguish knowledge-supported responses from those involving guessing. Crucially, they resist the temptation to crown any single indicator as definitive. No webcam, no timer and no entropy calculation can certify knowledge, detect guessing, or read an emotional state on its own. What the study demonstrates is that these signals are complementary pieces of a larger puzzle, and that the honest path to fairer electronic assessment runs not through any single magic metric but through the disciplined, statistically cautious fusion of many imperfect ones. In a world increasingly examined by algorithm, that caution may be the study’s most valuable result.

Subject of Research: Multimodal human-computer interaction analysis for reducing error in electronic knowledge assessment

Article Title: Analysis of Human-Computer Interaction Aimed at Reducing Error in Assessing Students’ Knowledge

Article References: Strugarevic, D., Nikolic, V., Perakovic, D., & Jevremovic, A. (2026). Analysis of Human-Computer Interaction Aimed at Reducing Error in Assessing Students’ Knowledge. Mobile Networks and Applications. https://doi.org/10.1007/s11036-026-02538-0

Image Credits: AI Generated

DOI: 10.1007/s11036-026-02538-0

Keywords: human-computer interaction, electronic testing, Shannon entropy, facial expression recognition, psychometrics, response time, learning analytics, multimodal assessment, machine learning, item quality, guessing detection, e-learning

The research team, led by Sari Rahmawati Kusuma Dewi of Taipei Medical University together with colleagues from several Taiwanese institutions, enrolled 111 hospitalized adults with impaired consciousness following acquired brain injury between October 2022 and November 2024 at Shuang Ho Hospital. The cohort was clinically heterogeneous: 37 percent had ischemic strokes, 33 percent hemorrhagic strokes, 25 percent traumatic brain injuries, and the remainder brain tumors or aneurysms. Patients were on average 69 years old, nearly two-thirds were men, and hypertension was present in more than 70 percent. Each participant underwent task-based EEG during hospitalization and was followed for six months, with outcomes categorized by whether the patient maintained or recovered the ability to follow commands.

The EEG paradigm itself was elegantly simple. Patients heard alternating spoken commands, ten seconds of “keep opening and closing your hand” followed by ten seconds of “stop opening and closing your hand,” delivered through the E-Prime presentation software with event markers embedded in the recording. Critically, the investigators included all patients in the analysis regardless of whether overt movement was visible, because the goal was to probe command-related brain engagement rather than confirm motor execution. Signals were recorded from a standard 19-channel 10-20 electrode array at 500 Hz, band-pass filtered between 1 and 30 Hz for analysis, and cleaned of artifacts using independent component analysis with automated classification via the ICLabel algorithm, with conservative retention of ambiguous components to avoid discarding genuine neural signal in this fragile population.

From these recordings, the team computed three measures of functional connectivity: spectral coherence, the weighted phase lag index (wPLI), and a debiased squared version of wPLI. The choice of metrics matters. Coherence, while popular, can be inflated by zero-lag synchronization arising from volume conduction, in which electrical activity from a single source is picked up by multiple electrodes, creating the illusion of communication between distant brain regions. The wPLI sidesteps much of this problem by weighting the imaginary component of the cross-spectrum, emphasizing phase-lagged interactions that are more likely to reflect genuine neural communication. The debiased squared estimator further corrects for finite-sample bias, providing a robustness check on the primary findings. Connectivity was summarized across delta, theta, alpha, and beta frequency bands during both moving and resting command conditions, with mean and median values averaged across all channel pairs.

The headline result was striking even if statistically fragile. Alpha-band wPLI during the moving condition showed the strongest nominal difference between outcome groups: patients who went on to maintain or recover command-following had higher alpha connectivity (mean 0.374) than those who did not (mean 0.318), a difference with a nominal p-value of 0.006. Alpha activity, oscillating at 8 to 12 Hz, has long been linked to large-scale cortical communication and thalamocortical function, making it a biologically plausible candidate marker of preserved consciousness networks. Group-level connectivity matrices reinforced the picture, showing visually stronger wPLI in the good outcome group, particularly over posterior parieto-occipital regions during the movement command condition. By contrast, coherence showed no significant group differences in any band or condition, and the debiased wPLI estimator showed directionally consistent but attenuated effects.

Here the story becomes a lesson in modern statistical rigor. The alpha-band finding did not survive Benjamini-Hochberg false discovery rate correction across the full family of 48 EEG connectivity variables, yielding a corrected p-value of 0.169. In adjusted logistic regression, alpha-band moving wPLI retained a nominal association with good outcome after controlling for age and Glasgow Coma Scale total score (odds ratio 1.76 per standard deviation), but the confidence interval was wide, and the association attenuated toward the null when the models additionally accounted for injury etiology or baseline motor responsiveness as measured by the GCS-Motor score. A Firth penalized regression, designed for small samples, produced a borderline, directionally consistent result. The authors are transparent: this is an exploratory, hypothesis-generating signal, not an established biomarker.

Perhaps the most sobering comparison came when EEG connectivity was pitted against plain clinical data. Age and Glasgow Coma Scale score together achieved a cross-validated area under the receiver operating characteristic curve of 0.778, while the multimodal model adding mean wPLI achieved 0.733 in leave-one-out validation, actually performing slightly worse once overfitting was controlled. Even simplified four-predictor models, which had healthier events-per-variable ratios, matched but did not exceed the clinical model. Formal incremental value testing confirmed the impression: adding wPLI features to age and GCS produced no statistically significant improvement by DeLong test or likelihood-ratio test. Calibration was adequate for both model types, and decision-curve analysis showed only small, threshold-dependent differences in net benefit, mostly between 55 and 75 percent probability thresholds. In short, the brain wave data told a biologically interesting story but did not beat two numbers any clinician can collect at the bedside.

Why does this matter? Because the field of consciousness prognostication is hungry for scalable tools. Positron emission tomography and functional magnetic resonance imaging can reveal residual brain function that behavior alone conceals, including the phenomenon of cognitive motor dissociation, in which unresponsive patients show covert brain activation to commands. But these technologies are expensive, technically demanding, and largely confined to specialized centers. EEG, by contrast, is inexpensive, noninvasive, and available in virtually every hospital. If task-related connectivity measures can be validated as complementary markers, they could eventually enrich multimodal prognostic frameworks without requiring patients to be transported to imaging suites. The present study also adds methodological value by demonstrating that phase-based measures such as wPLI behave differently from coherence in clinical populations, and that bias-reduced estimators should be reported alongside standard indices.

The investigators acknowledge important limitations. The study was single-center and modestly sized. The cohort excluded patients with hypoxic or post-cardiac arrest injury, the population where many consciousness biomarkers are tested, limiting generalizability. The six-month outcome, based on command-following, collapses death and persistent unresponsiveness into a single bad outcome category, which is clinically meaningful but coarser than graded functional scales. Many patients were already obeying commands at baseline, so the outcome partly reflected maintenance rather than recovery of responsiveness, and the task conditions were defined by instruction markers rather than verified behavioral compliance. Sedative effects were captured only as a binary variable, and short-epoch connectivity estimates may be influenced by fluctuations in arousal during recording.

The bottom line is characteristically scientific: a promising, plausible, and honestly reported signal that demands replication. Alpha-band task-related wPLI captured something real about the recovering injured brain, but its incremental value beyond age and clinical scales remains unproven in this cohort. The authors call for larger multicenter studies with external validation and more granular outcome measures before any clinical implementation. For now, the study stands as a careful step toward the long-sought goal of reading the injured brain’s network integrity at the bedside, using little more than electrodes, spoken commands, and rigorous statistics.

Subject of Research: EEG functional connectivity as a prognostic marker for six-month consciousness-related outcome after acquired brain injury

Article Title: EEG functional connectivity measures for differentiating 6-month consciousness-related outcome after brain injury

Article References: Dewi, S. R. K., Chen, H.-C., Li, Y.-C., Yang, H.-C., Huang, C.-W., Chan, L., Tu, Y.-K., Kuo, T. B. J., & Lin, M.-C. (2026). EEG functional connectivity measures for differentiating 6-month consciousness-related outcome after brain injury. BMC Medical Imaging, 26(1), Article 461. https://doi.org/10.1186/s12880-026-02681-w

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02681-w

Keywords: EEG, functional connectivity, weighted phase lag index, disorders of consciousness, brain injury, prognosis, Glasgow Coma Scale, alpha oscillations, coherence, multimodal assessment, neurophysiology, BMC Medical Imaging