Pose-estimated 3D data of infant spontaneous activity from Helsinki and Pisa

This dataset contains pose-estimated 3D (and 2D proxy) data as well as trained models for generating infant Kinetic Age, collected from research trials in Helsinki and Pisa. The dataset is organized into separate archives for metadata, data streams, trained models, and predictions. Below is a detailed breakdown of the dataset contents:

Metadata

metadata/combined.csv  - test_id: Unique identifier for each infant.  - corrected_age: Corrected age of the infant in days.  - outcome: Neurodevelopmental outcome labels (0 typical, 1 weak impairment, 2 MNI).

Data

data/features.csv  - Handcrafted movement features computed for each, by experts annotated, useful video segment.

data/streams/combined/*.feather  - 3D recording segment, with 18 J, B, V, A streams over 600 time steps.

data/streams_2d/combined/*.feather   - 2D recording segment, with 18 J, B, V, A streams over 600 time steps.

Results

results/model/fold_n/...  - train_predictions.npy: Predictions made on the training set.  - val_predictions.npy: Predictions made on the validation set.  - best_model.ckpt: Checkpoint file containing the saved model weights.  - metadata.json: Metadata describing the fold, including training parameters and validation segments.  - scatter_*.png: Regression results.

Predictions

predictions/jb-aagcn-coord-xy_predictions.csv  - Model predictions for the 2D model on MNI samples.

predictions/jb-aagcn-coord_predictions.csv  - Model predictions for the 3D model on MNI samples.