Detection of epileptic seizures from audio recordings

The health risks associated with epileptic seizures make early detection and effective intervention essential. Seizures detection devices are a useful tool for epilepsy monitoring. Accurate detectors can alert caregivers and reduce the distress of individuals experiencing nocturnal seizures. For this reason, there is an urgent need for non-invasive, accurate, affordable, and easily wearable technologies that can be used both in clinical and home monitoring settings. Thanks to recent advancements in machine learning and audio technologies, audio recording systems are considered a valuable option for remote monitoring. This research project presents a preliminary study on detection of epileptic seizures from audio recordings acquired in the patients’ room at the Epilepsy Unit of La Pitié-Salpêtrière hospital (Paris, France). Here, we analysed more than seven hours of audio data extracted during both seizure (ictal) and seizure-free (interictal) epochs of 5 hospitalized patients. Every audio-video recording was hand-labelled, to create a carefully selected dataset. The manual labelling process has enhanced the availability of high-quality datasets, a vital resource that is frequently lacking, benefiting not only the study's immediate objectives, but also the larger epilepsy research community. After the video recordings have been carefully labelled, the audio data were segmented in overlapped 10-s windows and pre-processed to extract relevant features. The proposed approach integrated standard features from: i) the time-frequency domain (on Mel’s scale) such as the spectral centroid, spectral entropy, spectral roll-off point, the cepstral coefficients; ii) the time domain, which include, for example, zero-crossing rate and energy. The extracted features served as the foundation for the application of different statistical models. Firstly, I used the Mahalanobis distance to measure the statistical dissimilarity between features vector in a current audio segment from those obtained during some seizure-free periods. By considering possible feature correlations, the Mahalanobis distance offered a reliable measure of dissimilarity that is essential to detect significant deviations from the reference class (here the seizure-free epochs). In a second part, we applied different machine learning models (decision trees, SVMs, k-NN and neural networks), which were trained on the manually labelled dataset and evaluated with a k-fold cross-validation to distinguish ictal and interictal segments. Performance metrics such as accuracy, precision, recall and F1-score (a better adapted metric for imbalanced datasets) resulting from a k-fold cross-validation were used to assess the performance of the detector. To address the class imbalance of the dataset, we also applied a statistical technique (SMOTE) to over-sample the minority class. Results showed that class imbalance yield overoptimistic results in general. When the imbalanced was corrected, results indicated that some methods (Mahalanobis-based detector, k-NN and decision trees) performed well and allowed to accurately identify the epoch with seizures. Our results suggest that the proposed method for detecting seizures offers a promising starting point for further investigation and refinement. The study not only contributes to the development of reliable seizure detection systems but also addresses critical considerations for practical implementation. Beyond the algorithmic framework, the research offers insights into the potential applications of audio-based detection in real-time seizure monitoring, presenting a non-intrusive and cost-effective alternative for long-term patient care. If these preliminary results will be confirmed in larger datasets, the proposed approach may be a useful contribution to healthcare technology, especially in neurological research.