longCOVID-fNIRS-dataset

This dataset supports the research article titled: "Exploring New Horizons: fNIRS and Machine Learning in Understanding postCOVID-19" This repository contains the expanded version of the functional Near-Infrared Spectroscopy (fNIRS) dataset and the complete Python scripts used for analyzing the neural correlates of postCOVID-19. The dataset consists of 29,737 time-series samples from 37 participants (9 postCOVID-19, 28 controls). We applied several machine learning algorithms with and without dimensionality reduction to evaluate their performance in classifying long COVID patients. In addition, the repository includes feature-level interpretability analyses based on subject-level statistical summaries (mean, standard deviation, minimum, and maximum per fNIRS channel and hemoglobin signal), performed through both Random Forest feature importance and PCA loading inspection to identify the most informative patterns distinguishing post-COVID and control groups. Contents Long_COVID_fNIRS_Raw_Data/: Raw fNIRS recordings (.tsv) from 37 participants (9 post-COVID, 28 controls). data_processing/: Python scripts implementing subject-aware preprocessing, feature extraction, dimensionality reduction, and classification using six machine learning models, and feature-level interpretability analyses. Methodology Acquisition: Brite MKII (Artinis), bilateral motor cortex, finger-tapping task. Preprocessing: standardization of HbO time-series signals; PCA where applicable. Machine Learning Strategies: Time-series: full-resolution HbO signals from 132 features per sample. PCA-based: reduced to 35 components retaining 95% of variance. Statistical features: 528 subject-level descriptors (mean, std, min, max). Hybrid: time-series samples augmented with per-subject statistics. Validation: strict subject-aware cross-validation using StratifiedGroupKFold (5-fold), avoiding data leakage. Metrics: Accuracy, Sensitivity, Specificity, PPV, NPV, and AUC-ROC. Feature-Level Interpretation: conducted on subject-level summaries using two complementary approaches: - Random Forest feature importance (supervised, 1,000 trees) to identify the most informative HbO/HbR/HbT descriptors under stratified subject-level cross-validation. - PCA loadings (unsupervised) weighted by explained variance ratio (95% retained) to reveal key signal components contributing to group discrimination. Key Findings The hybrid representation combining time-series and per-subject statistics achieved the most robust classification, with SVM yielding AUC = 0.91. PCA-based models also showed strong performance, while statistical-only models were limited (sensitivity < 0.30). Results emphasize the importance of preserving temporal structure and preventing data leakage across subjects. All scripts rely on standard Python libraries: scikit-learn, xgboost, matplotlib, seaborn, numpy, pandas. Feature-level analyses (Random Forest and PCA) revealed a consistent physiological signature dominated by extrema and variability of HbR/HbT, with supportive mean HbO descriptors, providing interpretable evidence of neural differences between post-COVID and control participants. Purpose This project aims to promote reproducibility and encourage further research into non-invasive, portable neuroimaging tools for the diagnosis and monitoring of long COVID. The results support the development of robust biomarkers using fNIRS and machine learning. How to Use Clone the Repository git clone https://github.com/usuario/longCOVID-fNIRS-dataset.git cd longCOVID-fNIRS-dataset Install Python Dependencies 2. Run Scripts 1_subject_aware_cv.py: time-series classification using group-stratified cross-validation 2_subject_aware_pca_cv.py: PCA-based dimensionality reduction and classification 3_statistical_features_cv.py: analysis based on statistical features per subject 4_hybrid_features_cv.py: hybrid model combining time-series and statistical features 5_feature_importance_RF.py: asupervised feature-level interpretability using Random Forest feature importance under subject-level cross-validation 6_pca_loadings_analysis.py: unsupervised feature-level interpretability through PCA loadings weighted by explained variance ratio 3. Explore Results Performance metrics are stored in the `results/` folder. Visualizations include ROC curves, confusion matrices, and boxplots. Feature-level interpretability outputs include ranked feature lists, barplots, and contribution tables for Random Forest and PCA analyses. Dataset Summary Participants: 37 (9 post-COVID, 28 control) Ethics Statement The study was approved by the State Research Ethics Committee in Health of San Luis Potosí, Mexico (SLP/08-2020). All participants provided informed consent in accordance with the Declaration of Helsinki. Authors' Contributions Edgar Guevara and Aaron López-Cano contributed to data acquisition. Antony Morales-Cervantes led data analysis and software development. All authors participated in the design, interpretation, and manuscript writing. All authors reviewed and approved this version.