What problem does it solve? Modeling ordered repeated measurements such as longitudinal visits, ROI time series, and dynamic connectivity requires sequence encoders that handle variable-length inputs without leaking information across subjects. This Skill trains temporal deep learning models on neuroimaging sequence data with subject-level cross-validation and training-only normalization. ## Core Features & Use Cases - Four Sequence Encoders: Supports LSTM, GRU, temporal convolutional networks (TCN), and masked temporal Transformers for both classification and regression tasks. - Variable-Length Handling: Uses an optional lengths array for sequence masking, with normalization statistics estimated from valid training frames only. - Subject-Level Cross-Validation: Produces fold assignments, per-fold checkpoints, predictions, metrics, and provenance manifests while keeping all visits from one subject in the same fold. - Use Case: Given an NPZ bundle of ROI time series from patients scanned at multiple visits, train a GRU classifier to predict diagnosis while ensuring no subject's visits are split across training and test sets. ## Quick Start Train a GRU classification model on my sequences.npz file with 5-fold subject-level cross-validation and save the outputs to a run directory.
