GPU-Accelerated Benchmarking of Transformer and Classical Models for Fetal Monitoring Time Series (CTG)

Dataset Description This dataset accompanies a comprehensive GPU-based benchmarking study of classical and transformer-based time-series models for fetal cardiotocography (CTG) signal analysis. The dataset was derived from clinical recordings exported from the MFM-CNS fetal monitoring system and processed to enable fair model comparison across deep learning and transformer architectures. The dataset is structured as follows: 01_ExamData.csv – The original raw data exported from the hospital information system. Each record may contain multiple patient sessions, as collected directly from the obstetric monitoring system. 02_SplitData.ipynb – The preprocessing notebook used to segment the raw data into individual patient-level sequences. This notebook performs filtering, cleaning, and restructuring of the input file into uniform arrays. _scored_output.xlsx – The finalized and de-identified dataset used in all machine learning (ML) and deep learning (DL) analyses. Each row represents one distinct patient recording with synchronized FHR (fetal heart rate), UA (uterine activity), and AFM (automatic fetal movement) channels, together with expert clinical scores. Additional files include model-level outputs, evaluation metrics, and scripts for full reproducibility: ctg_gpu_benchmark_v3_full.py – The main Python script containing all model definitions, GPU training routines, and metric computations. 19_GPU.ipynb – The Colab-ready notebook for running the full benchmarking pipeline. metrics_summary.csv, threshold_metrics.csv, model_complexity.csv, stability_mae.csv, wilcoxon_BH.csv, predictions_test.csv – Summary tables reporting accuracy, mean absolute error (MAE), model complexity, stability tests, and statistical comparisons. TRIPOD_report.txt – Transparent reporting checklist for model development and validation consistency. How to Reproduce Launch a Google Colab environment with GPU runtime enabled. Upload ctg_gpu_benchmark_v3_full.py and _scored_output.xlsx into the working directory. Open and execute 19_GPU.ipynb sequentially to reproduce all experiments, figures, and metric tables. This setup ensures full reproducibility of the model comparison results under identical preprocessing and data splits. Benchmarked Models Ten supervised models were trained and evaluated using identical data splits and hyperparameter budgets: Classical Deep Time-Series Models ResNet1D – Residual 1D convolutional architecture for hierarchical temporal abstraction. InceptionTime – Multi-scale inference via parallel convolutions with varying kernel sizes. TCN (Temporal Convolutional Network) – Dilated causal convolutions for long-range temporal dependency learning. LSTM-FCN – Hybrid model combining convolutional feature extraction with bidirectional LSTM layers. TimesNet-Lite – Frequency-domain representation of temporal patches for efficient learning. Transformer and Mixer-Based Models Vanilla Transformer – Multi-head self-attention with sinusoidal positional encodings. TSMixer – Alternating temporal- and feature-mixing operations for parameter-efficient representation learning. PatchTST – Patch-wise segmentation of sequences to capture long-term dependencies with transformer attention. iTransformer – Channel-transposed attention mechanism directly modeling inter-variable relationships. TST (Time Series Transformer) – Transformer variant optimized for regression tasks on multivariate physiological signals. Intended Use This dataset provides a reproducible foundation for fetal monitoring research, benchmarking new architectures in physiological time-series analysis, and validating transformer-based temporal models in clinical settings. All data are anonymized and derived from ethically approved retrospective sources.