Embedded Files
Obstructive Sleep Apnea (OSA) Detection using Neural Network and Machine Learning
Obstructive Sleep Apnea (OSA) Detection using Neural Network and Machine Learning
Motivation
Motivation
I developed a keen interest in the application of Artificial Intelligence (AI), Machine Learning, and Deep Learning in the realm of physiological signal processing for disease detection. My motivation for focusing on sleep-related disorders stems from a personal connection – my father's recurrent experiences with Sleep Paralysis. This condition, characterized by a temporary inability to move or speak while falling asleep, has inspired my dedication to exploring innovative technological solutions for the identification and understanding of sleep-related ailments.
What is Sleep Apnea?
What is Sleep Apnea?
Sleep apnea stands as a prevalent manifestation among sleep disorders, marked by respiratory disturbances during the sleep cycle. These disturbances, referred to as apneas, manifest as interruptions in breathing lasting for durations ranging from seconds to minutes and may transpire repeatedly over the course of the night. There are three main types of sleep apnea:
Obstructive Sleep Apnea (OSA)
Central Sleep Apnea (CSA)
Mixed Sleep Apnea
Obstructive Sleep Apnea (OSA) stands out as the predominant type of sleep apnea, particularly prevalent among the elderly demographic. This condition manifests when the muscles located at the posterior of the throat excessively relax during sleep, causing a partial or complete obstruction of the upper airway. Despite the affected individual making efforts to breathe, the airway's blockage induces audible phenomena such as snoring, gasping, or a momentary cessation of breathing. Subsequently, the brain prompts a slight awakening, facilitating the resumption of normal breathing.
Proposed Methodology
Proposed Methodology
The dataset employed in this research project consists of ECG signals and is publicly accessible through the Physionet Web platform. Specifically, the Apnea-ECG Database comprises 70 recordings, each featuring a singular ECG signal with varying durations ranging from slightly under 7 hours to nearly 10 hours. Each recording includes a digitized ECG signal, a collection of apnea annotations determined by expert human annotators, and a set of QRS annotations generated by machine algorithms. The ECG signals were acquired at a sampling frequency of 100 Hz. ECG, recognized as a highly effective feature for detecting sleep disorders, reveals cyclic variations in the duration of heartbeats, denoted as R-R intervals (the time interval between successive R waves), which have been reported to be indicative of sleep apnea episodes.
For the purpose of data selection, ECG records were chosen based on the criterion of having continuous apnea data for a specific timeframe. Given that apnea annotations were provided on a per-minute basis, the ECG signals were segmented into 1-minute intervals. Notably, the apnea annotation relied solely on the presence or absence of apneic events at the commencement of the respective minute. Recognizing the potential for ambiguity in such an annotation scheme, a mitigation strategy was implemented. Specifically, only the initial 30 seconds of each 1-minute segment were taken into consideration, while the remaining duration was disregarded.
To differentiate the R waves from other waves in the ECG signal, the identification of R-peaks relied on the satisfaction of two specific conditions. An R peak was deemed identified only if both of the following conditions were met:
The R peak had to exhibit characteristics of being a local maximum, discerned through the application of a local maximum function within a window duration of 150 milliseconds.
The local maxima peaks were required to surpass a threshold of at least 2 standard deviations above the mean.
Upon the successful determination of the R-peak based on these conditions, subsequent computations involved the derivation of R-R intervals.
The ECG signal underwent feature extraction, yielding the following derived metrics:
NN50 Measure: This metric is articulated as the count of pairs of successive R-R intervals wherein the second R-R interval exceeds the first one by more than 50 milliseconds.
SDSD Measures: These measures encompass the standard deviation of the differences between consecutive R-R intervals.
RMSSD Measures: These measures are defined as the square root of the mean of the sum of squares of differences between adjacent R-R intervals.
Inter-quartile Range: This measure represents the disparity between the 75th and 25th percentiles within the distribution of R-R interval values.
The decision-making model underwent testing employing both Convolutional Neural Network and Support Vector Machine algorithms. The classification scheme involved labeling instances as "1" to denote apnea and "0" to signify a normal condition.
Results
Results
In evaluating model performance, metrics such as accuracy, precision, recall, and F1-score were considered. Throughout all assessments, the Convolutional Neural Network (CNN) consistently demonstrated superior performance in apnea detection.
Future Work
Future Work
In the upcoming days, I aspire to undertake the design of a low-power integrated circuit, with the overarching objective of implementing and advancing existing research. This endeavor entails the incorporation of alternative neural network architectures, including Bi-directional LSTM, transformer models, and Graph Neural Networks (GNN). Additionally, I aspire to broaden the scope of the detection system by incorporating other physiological signals such as EEG, EOG, EMG, and SpO2.
Page updated
Google Sites
Report abuse