Beibit Abdikenov

Background/Objectives: Parkinson’s disease (PD) is a rapidly growing cause of neurological disability, yet diagnosis is often delayed in settings with limited access to movement-disorder specialists, including Kazakhstan and other Central Asian countries. This study aimed to develop a brief, self-administered, two-stage machine-learning-derived questionnaire based on binary symptom responses to identify individuals with possible clinically manifest PD who may benefit from specialist neurological assessment, and to evaluate its performance. Methods: From outpatient records of 402 patients with confirmed PD, a preliminary 100-item yes/no questionnaire was compiled. In a development cohort of 389 participants aged 60 to 90 years (146 with PD, 243 controls) recruited in Almaty, Kazakhstan, we used ElasticNet logistic regression, random forest, and gradient boosting to select features, reducing the instrument to 35 items (15 in Stage 1, 20 in Stage 2). We assessed performance on a held-out test set (n = 78) and in an interim prospective external-validation cohort (n = 290) recruited independently, with neurologists blinded to questionnaire results. Results: On the held-out test set, sensitivity was 86.2%, specificity 89.8%, accuracy 88.5%, and Stage 1 AUC 0.9557. In the interim external validation, sensitivity was 95.1%, specificity 94.4%, accuracy 94.5%, NPV 99.2%, and Stage 1 AUC 0.9962; only 12.1% of participants required Stage 2. Conclusions: The two-stage questionnaire showed strong case-finding ability in identifying individuals who warrant neurological evaluation. Findings are interim; completion of external validation, larger geographically diverse cohorts, and comparison against existing instruments are needed before wider implementation.

Digitizing the historical Chagatai manuscripts of the Gunnar Jarring Collection is severely hindered by cursive orthographic complexity and extreme data scarcity. This study establishes a rigorous empirical HTR benchmark, comparing two distinct paradigms under identical, highly constrained conditions: a convolutional recurrent neural network (CRNN) trained from scratch and pretrained Vision-Transformer-based architectures (TrOCR Base and Large). To eliminate intra-page data leakage, a strict page-level 80/10/10 partitioning protocol is enforced across the 2531 manually annotated text lines derived from 230 pages. To guarantee statistical reliability, all configurations are evaluated across three independent random seeds. The fine-tuned TrOCR Large model with architecture-specific data augmentation achieved the highest accuracy, yielding a Character Error Rate (CER) of 9.82 ± 0.05% and a Word Error Rate (WER) of 33.27 ± 0.77%, outperforming the from-scratch CRNN baseline (18.07 ± 0.79% CER) by 8.25% absolute. Our decoupled evaluations reveal that while pretraining provides a robust prior that stabilizes learning, spatial augmentation acts as a crucial regularizer. Our decoupled evaluations reveal that the performance advantage of fine-tuned Vision Transformers over scratch-trained CRNNs stems primarily from the stabilizing prior of pretrained transfer learning, while spatial augmentation acts as a crucial regularizer under extreme data scarcity. Character-level error taxonomies identify systemic failure modes, including deletion biases in Transformers and insertion-heavy “diacritic babbling” in connectionist temporal classification models.

Breast dynamic contrast-enhanced MRI acquires one image volume before contrast injection and several after it; each volume is a contrast phase. Which phases a tumour segmentation model needs has not been systematically tested. In 980 examinations from 22 sites of ISPY2, a neoadjuvant trial cohort in the public MAMA-MIA benchmark, we trained a 3D U-Net on all fifteen combinations of the first four phases, five seeds each, and measured overlap with the expert-corrected reference in Dice points (0–100). One comparison was pre-registered as confirmatory. The confirmatory comparison was inconclusive: a pair chosen before the sweep by a phase-selection model exceeded the clinical pre-contrast plus first post-contrast pair by 1.33 points (−0.39 to 3.12), which is too small a difference for this study to detect. Replacing the clinical pair with the best pair on average was also inconclusive (+1.56). All further findings are exploratory. Averaged across the three pairs of each kind, post-contrast pairs exceeded pairs containing the pre-contrast volume by 2.52 points (1.26–3.91), but this concerned overlap, not boundary accuracy, was not significant in a second cohort, reversed in point estimate when the ISPY2 models were applied there without retraining, and could be inflated by how the reference segmentations were generated. No pair suited every patient: the clinical pair was best for 30% of patients and the best-on-average pair for only 12%, so the average does not identify the best pair for an individual. The interval from pre-contrast to first post-contrast volume varied 2.25-fold across sites (87–196 s), so phase index cannot be assumed to mark a fixed point on the enhancement curve. For data already acquired, phase choice is cost-free, but its effect is small, patient-specific and unconfirmed beyond one cohort; it does not warrant protocol changes. Recording injection time would allow phase timing to be compared across sites.

Objective. EEG stress classifiers are usually trained and tested on a single dataset, and accuracies above 90% are common in the literature. We asked how much of this performance holds for participants the model has never seen, and for a dataset recorded with other hardware and another stress protocol.Approach. Two open datasets were processed with one pipeline: SAM40 (40 participants; Stroop, arithmetic and mirror-image tasks versus relaxation; 32-channel Emotiv, 128 Hz) and EEGMAT (36 participants; serial subtraction versus rest; 19-channel clinical system, 500 Hz). Four compact classifiers(logistic regression, random forest and RBF-SVM on band power, and a Riemannian tangent-space classifier) were evaluated with leave-one-subject-out (LOSO) validation, with cross-corpus transfer in both directions on ten shared electrodes, and with two unsupervised alignment methods, Riemannian re-centring and CORAL. Main results. Within SAM40, the best LOSO macro-F1 for one stressor against relaxation was 70.3% (best AUROC 0.804). Cross-corpus transfer lowered macro-F1 significantly in 22 of 24 model–direction–stressor settings (pairedWilcoxon test, ???? < 0.01), by up to 37.6 percentage points. Transfer was asymmetric: classifiers trained on EEGMAT reached an AUROC below 0.5 on SAM40 in 9 of 12 settings, so the learned decision direction was reversed. Re-centring improved within-corpus performance (EEGMAT macro-F1 from 0.690 to 0.728) but not transfer, whereas CORAL raised the EEGMAT→SAM40 AUROC of band-power models by 0.04–0.19. Removing ocular components withICA reduced band-power macro-F1 by only 1.1–3.0 points. Significance. Stress labels defined by the task carry protocol-specific information that does not survive a change of dataset. Cross-corpus testing should come before any claim that an EEG stress detector is ready for everyday or biofeedback use.

Museum and archive storage facilities worldwide face persistent difficulties in maintaining stable microclimatic conditions. This study presents a Smart Archival Box, a modular hardware–software system for preventive conservation, together with a model-based evaluation of its microclimate control. The numerical domain has a nominal volume of 0.125 m³ (0.5 m × 0.5 m × 0.5 m) and combines a 1 kg ArtSorb cassette, four TEC1-12706 thermoelectric modules, particulate/activated-carbon filtration, distributed monitoring and RFID-based access control. Temperature is regulated by a single proportional–integral–derivative (PID) controller with ambient-load feedforward, back-calculation anti-windup, hot-side derating and a dew-point safety limiter. Three scenarios—hybrid (H), active (A) and passive (P)—were evaluated over 10,173 one-second intervals (10,174 records; 2.8258 h) while the prescribed ambient temperature rose from 16 to 32 °C, and ambient relative humidity fell from 67% to 27%. H and A remained within the broader 16–20 °C preservation band for 100% of the run and within the tighter 17.5–18.5 °C control band for 79.7% and 80.5%, respectively, finishing at 18.53 °C and 18.56 °C. P reached 21.20 °C and remained within the tighter band for 47.7% of the run. Calculated electrical consumption was 0.05177 kWh for H, 0.05156 kWh for A and 0.00565 kWh for P. Relative humidity remained within 45–55% in all scenarios. The filtered H and A scenarios reduced mean PM2.5 and total volatile organic compound (TVOC) exposure by 98.4% and 94.4%, respectively, relative to the prescribed ambient profiles. The graphs for document prototype construction and functional bench observations and model results that have not yet been validated by enclosure tests are presented in Results Section. The close H/A performance over this short disturbance does not establish an advantage of ArtSorb and requires repeated humidity-disturbance experiments.

Background/Objectives: Parkinson’s disease (PD) is a rapidly growing cause of neurological disability, yet diagnosis is often delayed in settings with limited access to movement-disorder specialists, including Kazakhstan and other Central Asian countries. This study aimed to develop a brief, self-administered, two-stage machine-learning-derived questionnaire based on binary symptom responses to identify individuals with possible clinically manifest PD who may benefit from specialist neurological assessment, and to evaluate its performance. Methods: From outpatient records of 402 patients with confirmed PD, a preliminary 100-item yes/no questionnaire was compiled. In a development cohort of 389 participants aged 60 to 90 years (146 with PD, 243 controls) recruited in Almaty, Kazakhstan, we used ElasticNet logistic regression, random forest, and gradient boosting to select features, reducing the instrument to 35 items (15 in Stage 1, 20 in Stage 2). We assessed performance on a held-out test set (n=78) and in an interim prospective external-validation cohort (n=290) recruited independently, with neurologists blinded to questionnaire results. Results: On the held-out test set, sensitivity was 86.2%, specificity 89.8%, accuracy 88.5%, and Stage 1 AUC 0.9557. In the interim external validation, sensitivity was 95.1%, specificity 94.4%, accuracy 94.5%, NPV 99.2%, and Stage 1 AUC 0.9962; only 12.1% of participants required Stage 2. Conclusion: The two-stage questionnaire showed strong case-finding ability in identifying individuals who warrant neurological evaluation. Findings are interim; completion of external validation, larger geographically diverse cohorts, and comparison against existing instruments are needed before wider implementation.

Reliable control of upper-limb prostheses remains challenging because muscle signals can change between sessions and wearable EMG sensors may suffer from poor electrode contact. This study presents a real-time prosthetic hand control platform based on an OYMotion gForce Pro armband and a five-finger Dynamixel-actuated prosthetic hand. The system is designed to work without an external dataset, using only a short user-specific recording session and brief calibration before real-time use — specifically, five main training repetitions plus four supplementary calibration repetitions per gesture — intentionally minimising the calibration burden to enable practical single-session deployment. Five hand gestures—power grasp, open hand, OK gesture, index pointing, and thumbs-up—were classified using several machine-learning and ensemble-based methods. All models were evaluated under the same live data stream to ensure a fair comparison. Six recording sessions collected across five able-bodied participants on different recording days were included in the final statistical analysis. For every session, the main and supplementary training samples were combined to train session-specific models before live evaluation. Across six independent recording sessions from one able-bodied participant, Linear Discriminant Analysis achieved the best average live accuracy of 85.9%, followed closely by majority voting and adaptive ensemble methods. In contrast, the LSTM model showed lower performance, reaching 70.2% accuracy. The results indicate that, for low-data real-time prosthetic control, ensemble-based methods preserved near-best average accuracy while reducing reliance on any single classifier assumption, making them a more robust deployment choice than individual classifiers. Statistical analysis confirmed that session-to-session variability was the dominant factor in system performance (Kendall's W = 0.249) rather than classifier choice (W = 0.043), suggesting that practical reliability depends more on robust calibration and signal monitoring than on model selection.

Screening mammography is read comparatively: radiologists hold one breast against the other, and several finding categories have no meaning outside that comparison. Most automated pipelines nonetheless encode a single breast at a time. We describe Bilateral MV-CLIP, a contrastive image-text framework in which all four projections of a patient enter one forward pass and the two lateralities exchange information through a Cross-Lateral Attention block, with a learned gate and a placeholder token preserving behaviour when a projection or an entire side is unavailable. Training combines image-text and image-image objectives, graded similarity targets, and a symmetry term tying embedding geometry to recorded asymmetry status. We train on VinDr-Mammo with reports synthesised from structured annotations and evaluate by linear probing of frozen features, cross-modal retrieval and report decoding, against an independently trained single-breast model matched in backbone, optimiser, schedule and epoch budget, with checkpoints selected on a held-out validation partition. On a pre-specified composite endpoint pooling the three comparison-defined categories, fusion did not improve on that baseline: area under the curve 0.781 against 0.792, a difference of -1.1 points with a 95% confidence interval from -9.6 to +7.5. Per-category effects follow the direction the design predicts, but the partition holds 49 positive comparison-defined breasts and no difference survived correction. Retrieval is limited by weak instance-level image-text alignment rather than by report redundancy, with a pool-size-independent ranking accuracy of 0.691. On an external cohort of 11,913 patients labelled by biopsy the representations transfer at reduced discrimination, reaching 0.65 for any abnormality. We therefore report explicit bilateral comparison as plausible but, on this dataset, unconfirmed, and give confidence intervals throughout.

Tendon-driven robotic fingers are underactuated and strongly nonlinear, and their unidirectional, friction-attenuated tendon tension imposes actuation constraints that any practical controller must respect. A previous study established that a deterministic learning controller (DDPG) can track a grasp trajectory for such a finger and compared it against model-based schemes. The present work extends that line of investigation by asking a narrower but unresolved question: for this same single-actuator, three-link tendon-driven architecture, how does a deterministic actor-critic policy (DDPG) compare with a stochastic, maximum-entropy policy (SAC) once a tension-limit penalty is built directly into the learning objective? Using the previously validated physics-based simulation model, both agents are trained under identical conditions with a reward that jointly penalizes tracking error and tension-constraint violation, and are evaluated across random, sinusoidal, and step disturbances using steady-state error, RMS, ISE, and IAE. The results show a consistent and interpretable trade-off: the stochastic SAC policy attains lower tracking error and better disturbance rejection, especially under time-varying and abrupt loads, at the cost of noisier control and longer training, whereas the deterministic DDPG policy converges faster and produces smoother torque and tension profiles but adapts less well to coupled disturbances. The study isolates the role of the exploration mechanism and the tension penalty in shaping these behaviors and offers practical guidance on selecting between deterministic and stochastic actor-critic methods for physically constrained, underactuated tendon-driven systems.

Emotion recognition from electroencephalography (EEG) signals is a central problem in affective computing. This study benchmarks seven classical machine learning models—XGBoost, LightGBM, Random Forest, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, and Logistic Regression—for multiclass EEG emotion classification on the FACED dataset, which provides fine-grained labels across nine emotion categories (amusement, inspiration, joy, tenderness, anger, fear, disgust, sadness, and neutral) from 123 participants. We extract Differential Entropy (DE) features across five frequency bands and, using an identical feature-and-preprocessing pipeline, evaluate all models under three distinct protocols that differ only in how the data are partitioned: (i) an intra-subject protocol, in which a separate model is trained and tested within each participant and the per-participant scores are averaged; (ii) a pooled segment-level protocol, in which segments from all participants are mixed so that segments from the same participant may fall in both training and test sets; and (iii) a strict subject-independent protocol, in which all segments from any given participant are kept together on one side of the training and test split, so that test participants are never seen during training. Under the intra-subject protocol, the best model (LightGBM) reaches 85.4% accuracy; under the pooled segment-level protocol the best model reaches 56.4%; and under the subject-independent protocol the best model (SVM with an RBF kernel) achieves only 18.2% accuracy—above the 11.1% chance level but far below the other two settings. We provide reproducible baselines under all three protocols and argue that subject-independent evaluation is essential for meaningful benchmarking on FACED.