Spinal cord injuries (SCI) commonly result in varying degrees of respiratory motor compromise. Electromyography (EMG) is a primary tool used in both pre-clinical and clinical settings to measure deficits in neural output of the diaphragm, the primary inspiratory muscle. Despite its widespread use, objective characterization of EMG activity across physiological states remains challenging, and basic quantitative parameters such as burst amplitude or area under the curve are often insufficient to capture physiologically meaningful differences across conditions or treatment paradigms. We propose leveraging Convolutional Neural Networks (CNNs) to address this limitation, as they are well-suited for extracting hierarchical features from noisy, variable time-series data, and can be used to robustly and objectively categorize diaphragm EMG activity across various physiological states. In the present study, we developed a supervised deep-learning platform utilizing CNNs for quick and reliable segmentation of diaphragm EMG waveforms. This framework automatically classifies respiratory motor patterns and identifies meaningful changes in EMG activity across experimental conditions. To standardize analysis across preparations, we established a quantitative signal-processing pipeline that includes burst segmentation, filtering, and extraction of key features. These features include burst amplitude, burst duration, tonicity, and area under the curve, along with additional morphological and temporal metrics. Experimentally, adult Sprague Dawley rats (males and females, 3–6 months old, n=17) were implanted with indwelling bilateral diaphragm EMG electrodes. Animals were monitored across a range of controlled conditions: awake versus anesthetized states (isoflurane), pre- and post-cervical hemisection, and periods of electrical stimulation (3mA cathodal direct current vs. Sham). EMG bursts were manually annotated according to breathing type, experimental condition, and the quantitative metrics derived from our processing pipeline. Models indicate that the CNNs can reliably distinguish eupneic breathing from non-eupneic respiratory or movement-related bursts. Furthermore, the models successfully detect spinal stimulation-induced changes in burst morphology. This approach offers a scalable and objective method for characterizing respiratory motor output in SCI models. It significantly reduces manual labeling overhead, improves data reproducibility, and provides a robust foundation for future real-time assessment or closed-loop stimulation strategies. Funding: 1K99NS133388-01A1 (SR), Parker B Francis Foundation Fellowship (SR) This abstract was presented at the American Physiology Summit 2026 and is only available in HTML format. There is no downloadable file or PDF version. The Physiology editorial board was not involved in the peer review process.
Shastry et al. (Fri,) studied this question.