The presented scientific work describes the results of the development and evaluation of two deep learning algorithms: long short-term memory with a convolutional neural network (LSTM+CNN) and long short-term memory with an autoencoder (LSTM+AE), designed for the diagnosis of autism spectrum disorders. The study focuses on the use of eye tracking technology to collect data on participants' eye movements while interacting with animated objects. These data were saved in NumPy array format (.npy) for ease of later analysis. The algorithms were evaluated in terms of their accuracy, generalization ability, and training time, which was confirmed by experts. The main goal of the study is to improve the diagnosis of autism, making it more accurate and effective. The convolutional neural network long short-term memory and autoencoder-long short-term memory models have shown promise as tools for achieving this goal, with the autoencoder model standing out for its ability to identify internal relationships in data. The article also discusses potential clinical applications of these algorithms and directions for future research.

This paper proposes a method for detecting autism spectrum disorders (ASD) through the analysis of facial and motor features using machine learning. The aim is to develop an algorithm for automatic ASD diagnosis based on spatiotemporal behavioral patterns. Traditional diagnostic methods rely on subjective expert observations, often delaying intervention. To address this, a hybrid convolutional neural network and long short-term memory (CNN+LSTM) model was employed. Convolutional layers extracted spatial features from video frames, while recurrent layers tracked temporal dynamics. Using MediaPipe face mesh, pose, and hands models, 1,639 parameters were obtained, including facial and pose coordinates, hand landmarks, mouth aspect ratio (MAR), and motion energy. The dataset comprised 100 children, aged 5–9 years (50 with ASD, 50 typically developing (TD)). Stratified cross-validation was applied to ensure subject-independent evaluation. Results showed 90% accuracy on the training set, 85–90% on validation, and an area under the curve (AUC) greater than 0.90, confirming model stability. Data visualization highlighted significant differences in motor activity and emotional expression between groups. The proposed approach demonstrates the potential for robust and objective ASD detection. It can be applied in clinical and educational contexts to improve early diagnosis and timely intervention.

This study presents NORIA, an availability-aware multitask learning framework for clinical prediction from incomplete electronic health records (EHRs). We analyzed a public dataset of 77,724 patients and 107 numerical features. To prevent temporal leakage, predictors were grouped by availability, and training-only feature selection identified three baseline and two first-year variables. NORIA jointly predicts baseline and updated glycemic control, an HbA1c change index, and 21 medication labels using shared representations, masked training, and stage-specific prediction updates. Comparators included classical machine learning, gradient boosting, deep tabular models, and multitask architectures (MMoE and PLE). On the internal test set, NORIA achieved AUROCs of 0.8554 and 0.8888 for baseline and updated glycemic control, respectively, and medication macro-AP of 0.1854 versus 0.1879 for CatBoost. The predefined medication improvement criterion was not met. With 20% of baseline features artificially masked, updated-control AUROC declined by 0.0271 for NORIA, versus 0.0510 for CatBoost and 0.0542 for FT-Transformer. Ablation analyses indicated that masked training provided the clearest robustness benefit. NORIA thus demonstrated competitive accuracy and robustness in selected simulated missing-data scenarios, without establishing overall superiority.

Objective and reproducible diagnosis of attention-deficit/hyperactivity disorder (ADHD) remains challenging because of the limited availability of reliable electroencephalography (EEG) biomarkers and the high variability of neural signals. This study proposes a computational framework for ADHD detection based on dynamic functional connectivity and network-index modeling. Multichannel EEG recordings were transformed into temporal connectivity graphs using sliding-window correlations of band-limited amplitude envelopes. Several network-index models were evaluated, including linear, graph-based, recurrent, and hybrid spatio-temporal approaches. The proposed Hybrid Spatio-Temporal Graph Transformer demonstrated moderate, yet reproducible, subject-level classification performance. On the independent test set, the model achieved an accuracy of 63.16%, a balanced accuracy of 62.22%, a sensitivity of 80.00%, a specificity of 44.44%, an F1-score of 69.57%, and an AUC-ROC of 0.7444. Additional analysis of the derived network index demonstrated moderate intergroup separability, with a mean index shift of 1.16, Cohen’s d = 0.73, Pearson’s r = 0.36, and distribution overlap = 0.72. These findings suggest that the proposed framework captures informative spatio-temporal EEG connectivity patterns associated with ADHD; however, the model’s diagnostic applicability should be considered preliminary and requires validation in larger independent cohorts.

Early and objective diagnosis of autism spectrum disorders (ASD) remains a pressing issue in modern medical and cognitive informatics, where motor behavior analysis is considered a promising source of digital biomarkers. In particular, kinematic gait characteristics obtained with contactless sensors enable the identification of subtle motor impairments that may not be apparent on visual assessment. In this paper, we propose a hybrid multi-branch deep learning model, HMB-TSC, for the automatic classification of children with ASD based on spatiotemporal gait data extracted from the Kinect v2 camera. An open dataset containing gait recordings of children with ASD and typically developing (TD) peers was used for the experiments. Data preprocessing included outlier filtering, feature normalization, feature expansion, and time-series segmentation using a sliding-window method. The HMB-TSC architecture combines three complementary branches: a Transformer for modeling global temporal dependencies, a CNN-BiLSTM for extracting local spatiotemporal patterns, and a CNN-BiGRU for efficient sequential analysis. The performance evaluation was conducted using the holdout, 5-fold, and 5 × 3 nested validation protocols. In a reproducible released run, HMB-TSC achieved accuracy = 0.850 and AUROC = 0.948 in holdout validation, while 5-fold and nested validation yielded Accuracy = 0.850 ± 0.035, AUROC = 0.940 ± 0.029, and F1 = 0.845 ± 0.036. Compared with the baseline CNN, LSTM, and Transformer models, the proposed model demonstrated the best overall performance, especially in terms of AUROC and balanced F1-score.

Early neurodevelopmental assessment in children plays an important role in identifying factors associated with neurodevelopmental disorders, including autism spectrum disorders (ASD). Brain ultrasound imaging is commonly used in pediatric practice, although the interpretation of images may differ between specialists. In this work, a prototype system for automated analysis of pediatric brain ultrasound images was developed using a hybrid approach that combines U-Net and CNN for segmentation and classification tasks. The obtained results indicate that such an approach is technically feasible and may serve as a basis for further studies focused on neurodevelopmental risk factors.