MaleCNS connectome topology benchmarks for physiological time-series forecasting: processed data, graph controls, models and code

Published: 20 October 2026| Version 1 | DOI: 10.17632/bb65tgg5gc.1
Contributors:
,
,
,

Description

This dataset supports evaluation of MaleCNS v1.0 Drosophila connectivity as a fixed edge pattern in trainable sparse recurrent networks. It combines derived physiological forecasting data from VitalDB with selected connectome subgraphs, matched graph controls, trained models, and saved predictions. The processed physiological dataset contains 476,544 windows from 3,357 surgical cases and 3,253 subjects. Inputs are 60-s histories of invasive mean arterial pressure, heart rate, oxygen saturation, and end-tidal carbon dioxide, sampled on a 2-s grid. The target is mean arterial pressure averaged over a 60-s future interval beginning 300 s after the history ends. The subject-disjoint test set contains 74,187 windows from 484 subjects. The graph collection contains biological subgraphs of 256, 512, and 1,024 nodes at synaptic-count thresholds of 1 and 5, together with 30 null graphs. Node tables map model indices to source neuron body identifiers. Degree-preserving controls retain each node's incoming and outgoing degree; edge-matched random controls retain node and edge counts. The main trained biological network uses the binary threshold-5 topology with 1,024 nodes and 27,789 recurrent edges. The package includes 65 model checkpoints and prediction files, per-run metrics and training histories, cohort/QC tables, configuration and source code, node manifests, body-mapped edge exports, graph audits, and reproduction utilities. The main comparison uses five training seeds with fixed null graphs. Additional graph-realization, low-data, and capacity runs are exploratory. The deposited observations and results come from the supplied source datasets and experiment outputs. Release verification reproduces the saved metrics and the pointwise subject-and-seed bootstrap intervals. It also checks subject separation, normalization, graph mappings, and degree preservation. An optional CPU check replays each checkpoint on a deterministic sample. No models were trained while preparing the deposit.

Files

Institutions

Categories

Computational Neuroscience, Machine Learning, Time Series Forecasting

Licence