Predicting progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) requires models that represent both regional brain abnormalities and their evolution across repeated examinations. We developed a longitudinal graph neural network that integrates structural magnetic resonance imaging, FDG-PET, regional imaging biomarkers, and clinical covariates across irregular follow-up visits. The study included 614 participants with baseline MCI from the Alzheimer’s Disease Neuroimaging Initiative: 218 converters to AD within five years and 396 non-converters, with 2438 eligible longitudinal visits. Each visit was represented as an 82-node brain graph based on the Desikan–Killiany atlas. Node features combined a 128-dimensional multimodal convolutional embedding with four regional biomarkers. Graph-attention layers modelled spatial dependencies, a node-wise gated recurrent unit modelled longitudinal dependencies, and masked temporal self-attention aggregated variable-length visit sequences. Participants were divided at the subject level into development and held-out test sets, and hyperparameters were selected by five-fold cross-validation within the development set. On the held-out test set of 123 participants, the model reached an area under the receiver operating characteristic curve of 0.859 (95% CI 0.795–0.915), balanced accuracy of 0.805 (95% CI 0.736–0.862), sensitivity of 0.781, and specificity of 0.832. The AUC was numerically higher than that of the strongest baseline, a CNN–GRU sequence model, which reached 0.832 (95% CI 0.758–0.894); the paired AUC difference was 0.027 (95% CI 0.009–0.098), the unadjusted DeLong p-value was 0.026, and the Holm-adjusted p-value was 0.052, which was not significant at the conventional 0.05 threshold after correction for multiple comparisons. In ablation experiments, removing temporal modelling reduced the AUC to 0.818, and removing the spatial graph structure reduced it to 0.808, the two largest reductions observed. Integrated-gradient analysis placed the highest importance on hippocampal and entorhinal regions. Combining graph-based spatial modelling with recurrent longitudinal reasoning was associated with higher discrimination than sequence modelling alone, though this difference was not statistically significant after correction for multiple comparisons. Validation was restricted to a single research cohort (ADNI), and no independent external dataset was used; prospective external validation on an independent cohort is required before the model can be considered for clinical use Because FDG-PET was unavailable for 19.3% of visits, we report the headline result separately from a sensitivity analysis restricted to participants with complete FDG-PET at every visit (development set cross-validated AUC 0.891 vs. 0.874 for the full cohort with masked missing FDG-PET); multimodal performance should be read as cohort-dependent rather than as a single unconditional figure.

Sign Language Recognition (SLR) in low-resource sign languages is still limited due to the lack of annotated data, high variability of signers, and lack of benchmarks; Kazakh Sign Language (KzSL) is a representative example of such a low-resource language. The proposed research introduces a novel method for isolated KzSL recognition based on a purpose-collected dataset of RGB video streams of five signers, represented by 1662-dimensional per-frame pose, facial, and hand landmark vectors extracted using MediaPipe Holistic. The final dataset consists of 39,979 gesture samples across 187 output classes (186 gesture classes and 1 no_event class), each of which is resampled into a temporal landmark tensor 56 × 1662. A lightweight model, Transformer model (SignTransformerModel, 3,529,272 trainable parameters), is proposed with two parts: (1) Each landmark is projected through a separate branch of a MultiBranchStem, a branch representation is obtained for each landmark and fused into a unified representation of 256 dimension; (2) the fused sequence is processed by a four-layer Transformer encoder with eight attention heads, aggregated into a single sequence representation by a TemporalAttentionPool, which combines attention and max pooling, and then passed to a classification layer. The model was compared against three baseline architectures—a Standard Transformer, a Temporal Convolutional Network, and a Bidirectional GRU—and obtained a mean accuracy of 0.6698 ± 0.0376, a mean macro F1 of 0.4207 ± 0.0420, and a mean weighted F1 of 0.6546 ± 0.0375 over five outer test partitions of a nested, signer-independent Leave-One-Signer-Out cross-validation, outperforming all three baselines in macro F1 on every partition. The results show that compact landmark-based Transformer models are feasible for the recognition of KzSL, though aggregate accuracy is substantially inflated by the dominant no_event class (gesture-only accuracy: 0.4490), and that inter-signer variability and incomplete class coverage by signers as well as recurrent class confusions exist and present clear directions for expanding the dataset and fine-tuning the model with expert guidance.