Unilateral spatial neglect, or hemineglect, is a disabling cognitive consequence of right-hemisphere stroke in which a person fails to orient toward or interact with stimuli in the contralesional, usually left, side of space despite having intact primary sensory and motor function. The condition can interfere with reaching, object manipulation and self-care, and it is associated with longer hospitalization and poorer long-term outcomes. HemiBot is intended for post-stroke neglect through a non-contact exercise: the end effector of PAL Robotics’ TIAGo mobile manipulator acts as a moving visual cue while a participant reaches toward it with the affected arm. The robot does not physically guide the limb, and the task is designed to combine attentional cueing with active movement into the neglected hemispace.

The underlying software is divided into two cooperating layers. A session-level Coach node loads a YAML session plan, arranges exercises, receives aggregate results, presents a real-time scoreboard to the therapist and selects the difficulty for a subsequent session. At the individual-repetition level, a Python Behavior Tree built with py_trees coordinates motion commands, skeletal tracking, metric calculation, session aggregation and emergency-stop handling. This modular arrangement allows components such as the hand-pointing metric to be replaced later without redesigning the overall orchestration. After each attempt, the robot also delivers a text-to-speech message in Italian chosen from predefined pools associated with acceptable, failed or high-quality performance.

For the pointing exercise, the TIAGo gripper travels along a semicircular arc centered on the participant’s left shoulder. Easy, medium and hard settings use arc extents of 20, 30 and 40 degrees, respectively, while sharing the same 0.85-meter radius, angular velocity, three repetitions per standard session and dwell time at the extremes. The end effector always faces the arc center. Its movement is generated through a custom C++ ROS 2 action server using MoveIt 2 Cartesian path planning, with a trapezoidal velocity profile applied to the joint trajectory. Limits on joint velocity and acceleration are checked before execution, and invalid trajectories abort the action. Progress feedback is published at 5 hertz, while the Behavior Tree’s main tick runs at 30 hertz.

The system also treats the participant as a moving obstacle. Skeletal joint estimates from the robot’s onboard RGB-D camera are converted into spherical collision objects and updated in the MoveIt 2 planning scene. An OctoMap built directly from the depth stream serves as a coarser fallback when skeletal tracking temporarily deteriorates. This combination is intended to keep the planned trajectory away from the participant’s current body configuration, although the reaching targets themselves are designed a priori to remain at a safe distance. Emergency-stop signals are handled both by the arm action server and the Behavior Tree; the software detects the command within approximately 33 milliseconds at its 30-hertz control rate, but actuator delay, physical stopping time and stopping distance were not measured. The authors therefore describe the safety mechanisms as design provisions requiring quantitative validation before clinical use.

Movement quality is estimated from two quantities captured at 30 hertz: the angle between the line from the participant’s shoulder to the wrist and the line from the shoulder to the robot’s end effector, and the Euclidean distance between the wrist and the end effector. A sample counts as on target only when both fixed thresholds are satisfied, reducing the chance that correct-looking pointing without an adequate reach, or a reach without correct direction, will be rewarded. The proportion of on-target samples becomes a score from 0 to 100, divided into Fail below 25, Acceptable from 25 to 75 and Great at or above 75. Failed attempts must be repeated. At the end of a session, a mean score above 75 triggers promotion, while a failed-attempt ratio of 0.50 triggers demotion; demotion takes precedence if both conditions are met. The thresholds were derived in consultation with rehabilitation specialists and were not tested in patients.

Sixteen healthy adults, nine women and seven men with a mean age of 54.5 years and a standard deviation of 6.8 years, took part in two sessions. Eight entered a high-performance group and completed all repetitions as instructed. Eight entered a low-performance group in which the experimenter directed a scripted pattern of correct, failed and partial attempts during the first session. That partial pattern involved following the robot through the rightward half of the arc and resting during the excursion into the left, contralesional side. It was a controlled software test rather than natural patient behavior, and the participants were capable of correct reaching throughout. The experiment produced 120 recorded repetitions in total, with the robot’s measurements compared against BTS SMART-DX EVO stereophotogrammetric motion capture.

Across the 88 repetitions independently identified as fully correct by the motion-capture reference, measurement error increased as the arm moved farther toward the lateral workspace. The error increase was significant between Easy and the two more demanding levels, while the difference between Medium and Hard was not significant, suggesting that the largest source of difficulty arose when the arm left the frontal workspace rather than accumulating steadily with eccentricity. Median root-mean-square error for arm alignment was 3.42 degrees at Easy, 4.29 degrees at Medium and 4.39 degrees at Hard. For hand-to-end-effector distance, the corresponding medians were 1.59, 1.99 and 2.77 centimeters. Transient wrist occlusions were identified as a cause of larger errors in some trials, particularly when the wrist approached the robot’s gripper or extended laterally. Despite this, the robot and motion-capture systems assigned the same performance band in all 120 repetitions: 24 Fail, 8 Acceptable and 88 Great, with 100 percent classification accuracy and Cohen’s kappa of 1.000.

The adaptive logic also behaved as programmed for every participant. All eight members of the high-performance group were promoted from Medium to Hard after achieving no failed attempts, while all eight in the low-performance group met the demotion threshold and were moved from Medium to Easy. The high-performance group’s first-session mean score was 95.8, with individual means ranging from 86.0 to 100.0 and a fail ratio of zero. The low-performance group had a mean score of 40.7, with individual means from 39.7 to 41.8, and a fail ratio of 0.50, calculated from 24 failed attempts out of 48 first-session repetitions. After demotion, all eight low-performance participants achieved near-perfect scores in the second session, with means from 97.3 to 100.0 and no failed attempts. These results confirm that data collected asynchronously by the robot’s tracking and execution nodes reached the session aggregator and Coach node correctly. However, the performance groups were deliberately separated from the scoring boundaries: only one repetition came close to a boundary. The study therefore cannot establish how the system behaves when real patients produce gradual or fluctuating performance near the promotion and demotion thresholds. Healthy participants also do not reproduce the fatigue, impaired movement or attentional fluctuations expected after stroke.

Several important gaps remain before HemiBot could be used clinically. Its present score is motor-kinematic rather than a direct measurement of where a participant looks, so the system does not yet quantify the attentional side of neglect directly; wearable eye-tracking glasses are being explored for that purpose. Occluded samples can affect scores because the tracker does not explicitly identify every occlusion, and no minimum number of valid observations is required before an attempt is scored. The therapist’s scoreboard can override a repetition, while decisions based on session-wide statistics reduce the influence of a single anomaly, but these safeguards remain untested in a patient population. The study likewise did not establish that the robot improves neglect, that its feedback is perceived as empathetic or engaging, or that the full multi-exercise system is safe. Planned work includes usability testing, a clinical trial in people with post-stroke neglect, eye-tracking integration, gamification and more sophisticated closed-loop methods for adjusting difficulty.

Subject of Research: Technical validation of an adaptive robotic coaching system for unilateral spatial neglect rehabilitation

Article Title: HemiBot: design and technical validation of an adaptive Robotic coaching system for Unilateral Spatial Neglect rehabilitation

Article References: Soldi, R., Guerra, B. M. V., Sozzi, S., Lunghi, F., Russo, L., Schmid, M., & Ramat, S. (2026). HemiBot: design and technical validation of an adaptive Robotic coaching system for Unilateral Spatial Neglect rehabilitation. Frontiers in Robotics and AI, 13, 1950821. https://doi.org/10.3389/frobt.2026.1950821

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1950821

Keywords: HemiBot, rehabilitation robotics, unilateral spatial neglect, adaptive difficulty, mobile manipulator, ROS 2, motion capture, stroke rehabilitation, technical validation, design, technical, validation

The mathematical reason is straightforward but easy to overlook. When change is estimated as the difference between a follow-up measurement and a baseline measurement, the error in that difference depends not only on how large the errors are at each visit but also on how the two errors relate to each other. Errors that persist across visits tend to cancel when the measurements are subtracted, while errors that vary between visits can manufacture apparent change out of nothing or mask real change. The study formalizes this with a variance identity: the spread of paired-change error equals the sum of the visit-specific error variances minus twice their covariance. Because single-image evaluation captures the variances but never the covariance, no amount of single-snapshot testing can reveal how a system will behave longitudinally.

To demonstrate that this distinction matters in practice rather than only in theory, the researchers, Hunter Colson of the University of Kentucky College of Medicine and Manuel Valdebran of the Medical University of South Carolina, built two synthetic lesion-area experiments with exactly known geometry. Synthetic targets were chosen deliberately: because the true area and true interval change are known precisely, any error in the measurement can be attributed to the measurement procedure itself rather than to uncertainty in the reference. The experiments do not rank specific commercial models or establish clinical performance, but they isolate the measurement behavior with unusual clarity.

The first experiment, called the repeated-camera experiment, imaged 72 synthetic lesions at two visits, with 36 lesions unchanged and 36 growing by 10 percent in area. The team crossed multiple acquisition and processing choices, including camera distance, image width, three sources of three-dimensional surface geometry, three sources of lesion boundary masks, and six calibration-marker arrangements, yielding 960 configurations per arm that included at least one trained AI model. Geometry came from either the renderer’s known surface or from monocular depth models MoGe-2 or Depth Pro; boundaries came from the known mask or from SAM ViT-B or MedSAM, which were given a box around the already-located lesion. Eight coded markers of known size provided the scale reference.

The results were striking. Because camera pose, distance, and lighting were repeated between visits, errors at the two visits tended to be similar and largely canceled on subtraction. Assuming independent errors, the conventional shortcut, overestimated the paired-change error standard deviation by a median factor of 6.4 across the 960 no-change configurations, with ratios ranging from 1.8 to 143.5. In one worked example using MoGe-2 geometry and SAM boundaries at a 25-centimeter distance, the cross-visit error correlation reached 0.987, and ignoring the covariance would have inflated the estimated change-error spread by a factor of 8.2. No single conversion factor could translate single-image error into change error, since the ratio varied enormously across configurations.

Rankings were also unreliable. Pooled single-image error and paired-change error were correlated, with a Spearman coefficient of 0.724 in the no-change arm, yet the configuration with the lower single-image error had the higher change error in 23.2 percent of comparisons with unequal single-image values. The reversal rate remained substantial even when the single-image advantage was clear: 23.8 percent for a 1.5-to-2-fold advantage and 13.8 percent for a 2-to-3-fold advantage. When single-image errors were nearly tied, rankings were essentially unstable, flipping in 46.3 percent of comparisons and rarely surviving bootstrap resampling of the 36 lesions. Choosing a measurement protocol on single-image performance alone, the study suggests, is a gamble.

The second experiment tested what happens when the camera and lighting are not repeated between visits, a condition far closer to real clinical follow-up. For the same 36 unchanged lesions, the researchers compared a matched follow-up image, captured with identical camera pose, distance, and lighting, against a resampled follow-up in which viewing direction varied up to 8 degrees, distance varied uniformly between 15 and 25 centimeters, and light energy varied by up to about 20 percent in either direction. Varying the acquisition increased paired-change error in all 48 configurations using a trained AI model, by a median factor of 6.2, while single-image error barely moved. The median cross-visit error correlation fell from 0.993 to 0.752, showing that the comfortable error cancellation seen under laboratory conditions evaporates as soon as imaging conditions vary.

The study also elevates a frequently ignored metric: pair availability, the proportion of scheduled cases that yield valid measurements at both visits. In the repeated-camera experiment, both measurements were obtained for 95.3 percent of 8,640 scheduled pairs, but all 423 final failures occurred at the marker-detection stage, and because nine surface-and-boundary combinations reused each detection, a single failed detection silently produced 3,807 missing measurement rows. In the worst configuration, with recreated markers, a 25-centimeter distance, 512-pixel images, and clustered calibration markers, only 61.1 percent of lesions yielded complete pairs. A protocol can therefore look accurate on the cases it completes while quietly failing on the difficult ones, a pattern the authors argue must be reported separately from error among successful measurements.

Building on quantitative imaging biomarker methodology, including the Quantitative Imaging Biomarkers Alliance’s distinction between cross-sectional and longitudinal performance, the authors propose a reporting framework built on three questions: how accurately does the complete procedure measure change, how often does it return both required measurements and where do failures occur, and under which acquisition conditions and software versions were these properties established. The framework specifies that claims should name the model checkpoint, prompts, preprocessing, calibration, postprocessing, and execution environment, and that substantial software updates should trigger repeat testing, since changes to any of these steps can alter the relationship between errors across visits even when single-image performance improves.

The authors are careful about scope. The numerical findings depend on synthetic targets, controlled acquisition ranges, and segmentation models that were handed the lesion location, so they cannot be read as clinical failure rates or as evidence about any particular product. Recovering known changes in synthetic lesions establishes technical performance, not clinical responsiveness or patient benefit, which would require validation with physical targets, then human images, and finally demonstration of clinical utility. Still, the core message is hard to dismiss: for any AI system intended to measure change over time, single-image accuracy is not just insufficient evidence, it may point in the wrong direction. Direct evaluation of paired-change error, pair availability, and the full measurement protocol, the study concludes, is the only sound basis for a longitudinal performance claim.

Subject of Research: A paired-error framework for evaluating longitudinal technical performance of AI-enabled quantitative imaging

Article Title: Single-image performance alone does not establish longitudinal technical performance: a paired-error framework for artificial intelligence-enabled quantitative imaging

Article References: Colson, H., & Valdebran, M. (2026). Single-image performance alone does not establish longitudinal technical performance: a paired-error framework for artificial intelligence-enabled quantitative imaging. Frontiers in Imaging, 5, Article 1976334. https://doi.org/10.3389/fimag.2026.1976334

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1976334

Keywords: artificial intelligence, quantitative imaging, longitudinal measurement, paired-change error, pair availability, lesion measurement, measurement error, QIBA, image segmentation, technical validation, medical imaging, repeatability

The research, led by Habib Francis of the Sydney Clinical Trials Centre at The University of Sydney together with colleagues across Australia, Hong Kong and India, drew on plasma samples from the Fenofibrate Intervention and Event Lowering in Diabetes (FIELD) trial, a landmark randomised controlled trial of nearly 10,000 adults with type 2 diabetes. Because the samples came from a large, well-characterised clinical cohort, the team could interrogate the behaviour of 18 S rRNA across hundreds of real biobanked specimens rather than a handful of laboratory controls. In total, 789 plasma RNA samples analysed across 17 separate OpenArray runs formed the backbone of the stability assessment, giving the study a statistical weight that most reference-gene validations never approach.

The technical challenge the authors set out to resolve is a genuine worry in multiplexed qPCR. Microfluidic OpenArray plates pack thousands of nanolitre-scale reactions onto a single plate, and the tiny input of plasma RNA, which is naturally dilute and dominated by fragmented extracellular species, usually forces a pre-amplification step. During those 14 cycles of targeted pre-amplification, all the primer pairs in the pool compete for the same enzymes, nucleotides and reaction space. Highly abundant transcripts such as 18 S rRNA, which can be present at levels orders of magnitude above the messenger RNA targets of interest, could in theory monopolise shared reagents and quietly suppress the amplification of scarcer targets. If that happened, Ct values would shift in ways that no amount of normalisation could correct, corrupting every downstream comparison.

To test this directly, the researchers ran a paired experiment on 23 plasma samples. Each sample was pre-amplified twice, once with a primer pool containing the 18 S rRNA assay and once with an otherwise equivalent pool lacking it. The team then measured four representative target genes, ACTG1, ACTB, CLU and PFN1, under both conditions and compared the raw cycle threshold values sample by sample. Because not every gene was detectable in every sample, the effective sample sizes ranged from eight to seventeen per assay, a limitation the authors acknowledge candidly. Even so, the paired design is the cleanest possible way to isolate the effect of a single primer from all other sources of variation.

The verdict was reassuring. Median paired Ct differences between the two pre-amplification conditions ranged from just 0.07 to 0.52 cycles, and none of the four targets showed a statistically significant shift, with all p-values exceeding 0.05. Bland–Altman analysis, which plots the difference between paired measurements against their average, confirmed the picture: the mean bias across all four genes combined was a negligible −0.18 Ct, and the data points for each gene were distributed similarly, mostly falling within the 95 percent limits of agreement. In plain terms, adding 18 S rRNA to the pre-amplification pool did not measurably distort the amplification of the co-amplified messenger RNA targets under the conditions tested. The feared competition effect, at least in this workflow, did not materialise.

The stability data were equally encouraging. Across the 789 samples and 17 runs, 18 S rRNA was consistently detectable, with a median Ct of 14.5 and an interquartile range of 13.4 to 16.3. The distribution of Ct values was unimodal, clustering tightly around the median without the bimodality or systematic shifts that would hint at batch effects or hidden technical artefacts. Inter-run variability was low, with a coefficient of variation of just 3.9 percent and no significant differences between runs. Notably, these results were obtained using archived plasma samples that had endured three freeze–thaw cycles, conditions typical of longitudinal clinical biobanking, suggesting that 18 S rRNA detection remains technically robust even when samples have been handled the way real-world studies demand.

The study also tackled a practical problem that haunts long-running biomarker projects: what happens when a primer pool runs out? The FIELD analyses had relied on an archived pre-amplification pool containing the 18 S rRNA assay, stored unopened at −80 °C for more than five years. The researchers compared 18 S rRNA Ct values in 774 paired plasma samples amplified with either that archived pool or a newly manufactured one. Amplification succeeded under both conditions, but the new pool produced slightly lower Ct values, a statistically significant shift with a mean bias of 0.53 Ct. More concerning was the spread: the standard deviation of paired differences was 3.82 Ct, with 95 percent limits of agreement stretching from −6.95 to 8.01 Ct, and variability appeared to widen at higher average Ct values, where template abundance is lowest.

That combination of a small average bias and wide individual-level variability carries an important lesson for anyone managing longitudinal studies. The authors conclude that data generated with archived and new primer pools remain broadly comparable at the group level, but the two pools should not be treated as directly interchangeable for individual paired measurements. They recommend practical safeguards: keeping detailed records of primer sequences, concentrations and manufacturing lots; running bridging experiments on a representative subset of samples whenever a pool must be replaced; and using paired comparisons with Bland–Altman analysis to quantify any systematic bias before transitioning. Where the bias is small and consistent, calibration may suffice; where variability is large, reanalysis with a common pool or an alternative validated assay may be the only defensible option.

The authors are careful to frame what this study does and does not prove. Their findings establish the technical performance of 18 S rRNA within plasma-based OpenArray workflows, not its biological stability across disease states or experimental conditions, and they cannot be assumed to extend to other sample types, platforms or pre-amplification chemistries. Plasma is a peculiar analytical matrix, rich in fragmented extracellular RNA and poor in total messenger RNA, and pre-analytical variables such as blood collection tube type can reshape extracellular RNA profiles entirely. Still, by combining large-scale stability assessment, paired technical comparisons and a real-world test of reagent longevity, the study offers a template for validating endogenous controls in any high-throughput qPCR setting, and it gives biomarker researchers working with scarce clinical plasma a solid reason to trust one of molecular biology’s oldest internal controls.

Subject of Research: Validation of 18 S rRNA as an endogenous control gene in high-throughput plasma RNA qPCR workflows

Article Title: Evaluation of 18 S rRNA as an endogenous control in OpenArray® plasma RNA qPCR workflows

Article References: Francis, H., Januszewski, A. S., Mangani, A., Ema, F. K., Hardikar, A. A., Joglekar, M. V., Huang, M. L., Sullivan, D. R., Ma, R. C., Galande, S., Gebski, V., Simes, R. J., Jenkins, A. J., Molloy, M. P., & Keech, A. C. (2026). Evaluation of 18 S rRNA as an endogenous control in OpenArray® plasma RNA qPCR workflows. Molecular Biology Reports, 53(1), Article 1641. https://doi.org/10.1007/s11033-026-12820-9

Image Credits: AI Generated

DOI: 10.1007/s11033-026-12820-9

Keywords: qPCR, 18S rRNA, OpenArray, plasma RNA, endogenous control, pre-amplification, reference gene, Bland-Altman analysis, FIELD trial, biomarkers, gene expression normalisation, technical validation