Breath-Source Localization in Air-Based Molecular Communication: Experimental Data and Reproducibility Package

Dataset description This record contains the experimental data, one Jupyter analysis notebook, extracted features, and numerical output files associated with the paper "Breath-Source Localization in Air-Based Molecular Communication: A Learning-Driven Proof-of-Concept." The dataset comprises 1,021 controlled humidity–temperature recordings collected from three pseudonymized healthy adult participants using a DHT22 sensor. It includes 512 mouth-breathing and 509 nose-breathing recordings across Eupnea, Bradypnea, and Tachypnea. Measurements were acquired at approximately 2-second sampling intervals under controlled indoor conditions. The package also includes the raw `.dat` recordings underlying the system impulse response analysis in Figure 3. The accompanying Jupyter notebook loads and audits the raw recordings, performs signal preprocessing, extracts 76 handcrafted features from the humidity and temperature signals and their first-order differences, and evaluates logistic regression and CatBoost using a subject-independent leave-one-subject-out (LOSO) protocol. Missing-value handling, feature standardization where applicable, and decision-threshold selection are performed using only the training subjects within each fold. The generated outputs include the extracted-feature table, held-out predictions, aggregate and subject-wise performance metrics, normalized confusion-matrix data, and subject-wise ROC-curve coordinates. The package supports reproduction of the feature-extraction and classification analysis, comparison with alternative ML methods, and further research on respiratory sensing, breath-source localization, and air-based molecular communication. This package is intended to support: reproduction of the feature-extraction and subject-independent classification results reported in the paper comparison with alternative feature-extraction and classification methods research on respiratory sensing and mouth-nose breath-source localization studies in air-based molecular communication (ABMC) and the Internet of Bio-Nano Things (IoBNT) exploration of low-cost sensing approaches for future wearable or ambient respiratory-monitoring systems Data and Code Experimental data, metadata, and numerical outputs are licensed under the Creative Commons Attribution 4.0 International license. The Jupyter notebook and source code are licensed under the MIT License. Short description of the paper This paper presents a learning-driven proof of concept for localizing whether an exhaled-breath signal originates from the mouth or the nose using a single low-cost humidity-temperature sensor. Mouth and nose breathing are modeled as distinct molecular sources within an air-based molecular communication framework. Temporal measurements collected during Eupnea, Bradypnea, and Tachypnea are represented using handcrafted signal features and evaluated with logistic regression and CatBoost under a leave-one-subject-out (LOSO) protocol. The nonlinear CatBoost model achieved a mean balanced accuracy of approximately 76%, with performance reaching 87% for one held-out participant. The results show that humidity and temperature dynamics contain useful source-dependent information, even with minimal sensing hardware. The paper also interprets the classifier as a binary asymmetric communication channel and estimates an effective capacity of approximately 0.217 bits per breath.