CHAPTER 1

INTRODUCTION

1.1 INTRODUCTION
Sleep Apnea is a serious condition that occurs when a person's breathing halts and restarts
several times during the night, therefore the disorder causes the person not to get a good rest, and the
person does not feel fresh in the morning and lacks the energy and mental clarity that would
normally allow him to be at the top of his game. The most common form of apnea is Obstructive
Sleep Apnea (OSA), which is when the throat muscles relax and thus the air flow is partially or
completely blocked. The condition, which is merely an annoyance that interrupts sleep, has been
found to elevate the risk of different health problems such as high blood pressure, cardiovascular
disease, stroke, type 2 diabetes and cognitive dysfunction.

Polysomnography (PSG) is the reference standard to sleep apnea diagnosis. It is a

procedure that records brain activity (EEG), eye movement, muscle activity, heart rhythm (ECG),
blood- oxygen levels, and breathing continuously during sleep. Although PSG maintains excellent
diagnostic sensitivity, it is expensive, time-consuming and not easily accepted on a daily basis as
screening tools at home or community facilitation. These obstacles result in a delay in diagnosis
and treatment, particularly in rural or resource-poor environments.

Recent developments in Artificial Intelligence (AI) and Deep Learning have created new
opportunities for producing automated, non-invasive and cost-effective diagnostic tools. Snore sound
analysis is one feature that has gained prominence as an accessible and informative biomarker to
identify OSA. Snoring, commonly denoted as non-threatening or benign, does have unique
acoustic features that may be related to an apnea event, thus an automated system can differentiate
the pattern of normal and apnea snoring.

The objective of this project, "Sleep Apnea Diagnosis from Snore Sounds Using CNN-
LSTM Model," is to develop a smart, non-invasive, and inexpensive system for sleep apnea
diagnosis through the analysis of snore sounds. The model utilizes Convolutional Neural Networks
(CNN) for feature extraction and Long Short-Term Memory (LSTM) networks for learning
temporal sequences to utilize both spatial and temporal characteristics of a snore audio.

1
1.2 NEED FOR SNORE SOUND–BASED SLEEP APNEA DETECTION

Diagnostic apnea capabilities during sleep are limited by traditional detection methods. The
usual polysomnography (PSG) which is a hardware-intensive, staffing intensive and an overnight
hospital stay procedure is both laborious and expensive. As a solution, home sleep apnea tests
(HSATs) come in as an alternative; nevertheless, HSATs still have multiple wired components and
the limitations of patient adherence. The PSG and HSAT testing procedures are also unpleasant,
which may affect the participants' normal sleep patterns and thus, the results may be inaccurate.

Speaking volume correction snore proposition has another alternative: the sound from a
microphone in a store or a wearable device can be used to determine the Sleep stage behavior
patterns of the participants. Snoring is one of the leading physical signs of obstructive sleep apnea
(OSA), and its acoustic features such as frequency, intensity, and temporal patterns may reflect the
upper airway obstruction behavior. Machine learning and advanced sound signal processing can be
used to analyze snore sounds for apnea-related disruptions.

Polysomnography requires the patient to spend the night in a specially equipped laboratory,
with wires attached to sensors for continuous recording of EEG, ECG, EMG, airflow, oxygen
saturations, and body movements. While polysomnography gathers comprehensive data, it has the
following shortcomings.

1. Expensive - The extremely high setup and test fees create a barrier to use for most patients,
especially in developing areas.
2. Time consuming - The entire evaluation process consists of overnight testing and at least
days of analysis following testing.
3. Uncomfortable - The additional sensors attached from head to toe along with being in a
hospital setting influences the patient’s typical sleep behaviors and validity of results.
4. Limited access - There exists a shortage of sleep laboratories and sleep specialists, resulting
in long waitlists and patients going without treatment.
Given these limitations, the clinical field would benefit from a simpler, more accessible, and non-
invasive screen that can be given frequently and even at home.

2
 Fig 1.1 Mel-Spectogram representation of snore

1.3 BENEFITS OF IMPLEMENTATION

Using a system that identifies sleep apnea based on CNN-LSTM offers numerous
advantages across medical, technological and social dimensions.
1. Patient Comfort and Accessibility

The system cuts down on the number of wired sensors necessary, thus the users can be
screened in their homes. The whole operation is easy and convenient with the microphone-based
setup, and to make sure the compliance and reliability are better, normal sleep patterns are kept.
2. Cost-Effectiveness

If the AI-based analysis of snore sounds is used instead of PSG, the cost of a diagnosis can
be lowered by a large margin. Consequently, it becomes feasible to carry out large-scale screenings
in villages or poor areas that lack the accessibility of a sleep lab.

3. Early Detection and Preventive Healthcare

Whether done continuously or intermittently, monitoring the acoustic of snoring can lead
to the very first recognition of apnea episodes and, consequently, suitable medical intervention. It
can stop the occurrence of other comorbidities like high blood pressure, stroke, and diabetes.

3
4. Telemedicine

The data can then be sent to healthcare providers either via mobile apps or cloud-based
platforms, facilitating remote diagnosis and real-time monitoring in support of the goals of digital
transformation in [Link] system can collect and analyze large sums of snoring data, which
ultimately provides opportunities to further advance existing research in sleep medicine as well as
enhance deep learning models in a larger scope of medical practices.

1.4 METHODOLOGY AND SYSTEM DESIGN

The Sleep Apnea Diagnosis from Snore Sounds using CNN–LSTM Model is the planned
system that takes a well-thought-out approach to ensure accurate data capturing, efficient feature
extraction, and reliable classification. The architecture of the system merges signal-processing and
deep-learning methods to create a smart, non-invasive diagnostic model.

1.4.1 Audio Signal Acquisition

Initially, snore sound recording of a sleeping subject is the focus. In order to catch the snore
signals without the physical discomfort, good-quality microphones or wearable acoustic sensors are
located near the nasal or oral region. The recordings are done in digital form (WAV format) with the
sampling rate usually being between 8 kHz and 16 kHz.

Each recording period includes normal snoring, apnea events, and silence. The correct
identification of these segments is done by medical experts thus creating a trustworthy dataset. The
audio database for training and testing consists of both positive and negative apnea cases, thus
ensuring a balanced representation of the problem.

1.4.2 Pre-Processing and Segmentation

The original audio files may have background noise, breathing sounds, and other noises from
nature. To get rid of such unwanted parts the following pre-processing means are used:

Noise filtering: The usage of band-pass or Wiener filters to get rid of the ambient and low-
frequency noise.

4
Silence removal: The detection and removal of long silent parts that lack diagnostic information.

Amplitude normalization: Signal levels are equalized for feature extraction to be consistent.

Framing and windowing: The continuous signal is broken into short overlapping frames (e.g., 25
ms frames with 10 ms overlap) using the Hamming window to keep temporal continuity.

Once pre-processing is completed every frame gets a label of either normal or apnea-related thus
forming structured input samples for the deep-learning model.

1.4.3 Feature Extraction Using CNN

The CNN part of the system is aimed at getting spatial and spectral features from snore
sound [Link] order for CNN to work on the data, each pre-processed audio frame has to be a
Mel- Spectrogram or Mel-Frequency Cepstral Coefficient (MFCC) representation that changes the
1-D signal to a 2-D image showing both frequency and [Link] CNN levels automatically get the
features that separate the groups such as frequency bands, energy changes, and harmonic structures
that characterize apnea events.

Normally CNN operations are:

Convolution layers: To find local frequency-domain patterns.

Pooling layers: To lower dimensionality and keep the most used information.

Activation functions (ReLU): To add non-linearity.

Batch normalization and dropout: To help generalization and stop overfitting.

The CNN output is a high-level feature map which is sent to the next step of temporal learning.

1.4.4 Temporal Pattern Learning Using LSTM

Though CNNs can find local spectral patterns, they cannot show time dependencies across
frames. So, LSTM layers are there to figure out the temporal relationships between consecutive
snore frames. LSTM keeps memory cells that can hold long dependencies and thus help in the
detection of apnea events occurring gradually over time.

Output vectors from CNN are fed to LSTM which handles them in the order of time thus it
can find breathing pauses, irregular rhythms, and recovery snores typical of obstructive events. Such
a combined architecture lets the system grasp not only the spatial but also the temporal aspects snore
sounds thereby increasing the accuracy of classification.

5
1.4.5 Model Training and Evaluation

After deciding on the architecture, the CNN–LSTM model is trained with labeled snore data.
The data is split into training, validation, and testing sets (usually 70 % / 15 % / 15 %). Training is
about minimizing a categorical cross-entropy loss function through use of optimization algorithms
such as Adam or RMSProp

Several methods are used to ensure the model's performance is up to the mark and they are:

 Data augmentation (pitch shifting, time-stretching) to broaden the dataset.

 Early stopping to avoid overfitting.
 Hyper-parameter tuning (batch size, learning rate, epochs) to get the best result.

They assess the model's performance using metrics such as accuracy, precision, recall, F1-score, and
confusion matrix. The findings are measured against those of the baseline models like standalone
CNN or LSTM to show the extent of improvement in classification ability.

1.4.6 System Integration and Implementation

The last step is about connecting the trained CNN–LSTM model with a simple-to-use
interface that can handle either real-time or offline analysis.

Where the system might be set up by:

Embedded boards (e.g., Raspberry Pi) for on-device inference, or Mobile/Cloud platforms for
remote monitoring and data storage.

In a real-time set-up, the microphone is constantly capturing snore sounds which the model
is processing to give an instant classification output that reads either 'normal snoring' or 'apnea
detected. The results can be viewed on a mobile app or web dashboard and, if desired, sent to
healthcare professionals through a telemedicine network. Such an integrated layout yields a portable,
inexpensive, and energy-efficient sleep apnea detection tool that is aligned with the preventive mode
of healthcare and can also be used for continuous home-based monitoring.

6
1.5 MOTIVATION OF THE PROJECT

Sleep apnea is among the most common and worst-diagnosed sleep disorders troubling
people around the world. Sleep apnea interrupts sleep and can create the risk of serious health
problems, including hypertension, stroke, heart attack, diabetes, obesity and depression. Most
people are unaware of their sleep apnea until it becomes persistent, even if they exhibit serious
symptoms, and therefore sleep apnea continues to have negative consequences on sleep and
short and long- term health. According to WHO, a significant proportion of people with sleep
apnea in developing countries are not identified because of limited access to diagnostic
services and costs associated with testing.

The conventional diagnostic procedure, Polysomnography (PSG), while medically

accurate, has disadvantages. The patient must spend one night at a sleep research center, with
multiple wired sensors connected to monitor EEG, ECG, EMG, airflow, and oxygen saturation.
This is an expensive and inconvenient process, and would not be feasible for mass screening or
long-term monitoring. Furthermore, the stress of sleeping in a clinical setting changes normal
sleep behavior and diminishes the accuracy of the data. Therefore, there is an urgent need for a
comfortable, inexpensive diagnostic alternative.

At the same time, the analysis of snoring sounds has emerged as a promising and non-
invasive method for the detection of sleep apnea. Snoring is one of the most prevalent and early
signs of obstructive sleep apnea. Each snore that relates to an apnea episode has unique
acoustic characteristics that are distinct from typical snoring — including amplitude of sound
interruptions, irregularities and changes in breaths, and fluctuations in frequency. These
characteristics of sound can be automatically analyzed by using a deep learning model to detect
episodes of apnea without having any physical sensors worn by the patient.

7
 Fig 1.2 Snore Sound Analysis

1.6 SCOPE OF THE PROJECT

The primary objective of the "Sleep Apnea Diagnosis from Snore Sounds using CNN-
LSTM Model" project is to create a clever, non-invasive, inexpensive way that can locally but
automatically detect the incidences of sleep apnea from the analysis of snore sounds. The scope of
the work is a multidisciplinary-circle going across anatomical fields - signal processing, deep
learning, embedded systems, and telemedicine - and giving a dataset and peripheral as well as an
implementable health- focused [Link] goal of this project is to reform the typical diagnosis
process as we know of sleep apnea into a portable, comfortable, and AI-driven diagnosis system
addressing every aspect from audio capture and data classification, as well as performance
assessments while integrating with an implementable aspect.
Specifics of scope includes:
1. Audio-Based Diagnosis System:
This project will be limited to snore sound auditory based analysis which alleviates the need
for additional invasive sensors such ECG and airlfow or SpO2 monitors. The method minimizes an
intrusive patient experience without uncomfortable variables for audio capture.
2. Deep Learning Model:
The scope addresses the building and train of connected nodes in hybrid design communicating
in partially absorbable computing, describing sound through spatial (frequency-like for snore
sound location -- e.g. probability increasing and equal spacing sound waves) and temporal (time-
based --

8
probabilities effect on time based factors for sleep stage example) meaning.

3. Signal Processing and Feature Engineering:

The project incorporates preprocessing activities will be needing capture captures noise,
filtering noise during the framing and extraction the appropriate Mel-Spectrograms or MFCC parts
of the sound, which will provide process ways to communicate meaning into feature models for
deep learning classifications.
4. Dataset Preparation and Evaluation:
We will work with snoring sound datasets that are available online for model training as well
as testing. The model is evaluated on the criteria of accuracy, precision, recall, F1-score, and
confusion matrix analysis to assess overall performance.
5. Prototype and System Integration:
The system will be exploring the capability of deploying the trained model onto an
embedded or mobile platform for continuous, offline monitoring. This would
potentially be in the form of a wearable device or smartphone-based application for
usability.
6. Telemedicine and Cloud Connectivity:
The project will investigate the integration of cloud services because this will allow
remote diagnosis and data sharing between patients and healthcare professionals, thus
improving real-time monitoring and giving a substantial contribution to the telemedicine
landscape and connected, digital healthcare.
7. Impact on Society and Research:

With affordable large-scale screening and increased awareness of sleep disorders, the
planned system will be of significant benefit to public health. Furthermore, the system
will serve as a research tool for further investigations in biomedical signal analysis and
AI- based health-care.

1.7 OBJECTIVES OF THE PROJECT

The major goal behind this "Sleep Apnea Diagnosis from Snore Sounds using CNN-
LSTM Model" initiative is to create and put into operation a smart, non-invasive, and
inexpensive deep learning system that can identify sleep apnea from snore sounds. The
project will address the limitations of traditional diagnostic methods such as
Polysomnography (PSG) by providing an automated, comfortable, portable alternative

9
that allows for early

10
 detection and in-home sleep tracking. The specific objectives are:
To develop a CNN-LSTM hybrid model to extract and classify audio features
from snore recordings.
 To apply audio preprocessing approaches such as noise filtering, and spectral
transformation.
 To provide high accuracy detection with a low computational cost.
 To provide a portable low-cost home-based solution for sleep monitoring.
 To provide telemedicine features with mobile or cloud-based access.

1.8 ORGANIZATION OF THE REPORT

Chapter 1: Introduction

This chapter gives a thorough overview of the project, beginning with an overall introduction
to sleep apnea and discussing the issues posed by traditional testing such as polysomnography-
based diagnostic approaches. It also justifies the need, benefits, motivation, scope, and rationale for
developing a non-invasive, deep learning–based snore sound analysis system. So it contains the
basis by setting out clearly the purpose of the work, its aims, and the significance of the project in
modern- day healthcare issues.

Chapter 2: Literature Review

This chapter provides a discussion around previously proposed studies, research papers, and
existing technologies around sleep apnea detection. It does a review of the various methods used
for analyzing respiratory sounds, ECG, and other signal processing techniques using AI and
machine learning to study complex physiological signals. A literature survey would help
understand research gaps, the limitations of current systems, and give justification for selecting the
CNN-LSTM model as the appropriate approach to the work.

Chapter 3: Methodology and System Design

This chapter elaborates on the intended methodology. The steps of data collection, audio
preprocessing, feature extraction, and model development are described. Furthermore, there are
diagrams and flowcharts present that describe the architecture of the proposed CNN- LSTM model
and its working principle. Hardware and software requirements are elaborated as well as the
dataset

11
details and training aspects associated with implementing the system.

Chapter 4: Results and Discussion

This chapter illustrates the experimental findings of the designed model. It includes the
assessment of system performance on metrics such as accuracy, precision, recall and F1-score with
the ensuing confusion matrices and graphical comparisons. The chapter assesses the extent to
which the CNN- LSTM model accurately differentiates between normal and apnea-associated
snore sounds while also comparing the performance assumptions of the model with other
traditional or AI- based 'systems'. The performance insight of the trial enables to validate the
efficiency of the proposed system.

Chapter 5: Conclusion and Future Work

The last chapter first summarizes the main results and overall accomplishments of the
project. It considers how far the project objectives were achieved, the advantages brought about by a
CNN- LSTM-based solution, and the resulting impact of the deployment of such a system in a real-
world healthcare setting. Besides that, it elaborates on the upgrade possibilities such as the local
real-time implementation, app development integration, and the addition of multimodal
physiological data to further enhance trustworthiness and diagnostic precision.

12
 CHAPTER 2

2.1 OVERVIEW

Sleep apnea ranks as one of the commonest sleep disorders that have not been
sufficiently recognized, which is characterized by repeate episodes of partial or complete
blockage of the upper airway during sleep resulting in oxygen desaturation and disrupted
sleep patterns. The most common type, Obstructive Sleep Apnea (OSA), may lead to chronic
fatigue, cardiovascular problems, and metabolic disorders after a long period of the condition
without treatment. Generally, the identification has been centered on Polysomnography
(PSG), which necessitates a night stay at a hospital or clinic with the patient being connected
to different sensors. Despite high precision, PSG is costly, takes a lot of time, and is
inconvenient, thus, it is less suitable for extensive or home-based screening.

During the last few years, several non-invasive and inexpensive sleep apnea detection
techniques procured through acoustic signal analysis combined with artificial intelligence (AI)
have attracted a lot of attention. Obstructive sleep apnea (OSA) is usually accompanied by
snoring, which has turned to be a highly informative biomarker for apnea events detection.
The changes in the acoustic properties of the snoring sounds (e.g., pitch, amplitude
modulation, and spectral patterns) caused by apnea-related episodes as compared to the ones
of normal breathing are quite significant, thus, they can be effectively employed as a basis for
AI-driven classification.

2.2 LITERATURE SURVEY

Title of the Paper Journal Name and Methodology Limitations

Year
Feature-Based Detection and Terisara Remote Needs high-quality
Classification of Sleep Apnea and Micaraseth; Shanop photoplethysmography and infrared imaging
Hypopnea Using Multispectral Shuangshoti; multispectral feature-based setup; real-world
Imaging Kantawatchr classification scalability untested
Chaiprabha; Natcha
Wanpiyarat;-2025
An Adaptive Sleep Apnea IEEE Journal of Multi-Cascaded Atrous High model
Detection Model Using Multi Biomedical and CNN optimized with complexity;
Cascaded Atrous-Based Deep Health Informatics, AHPBA algorithm optimization is
Learning Schemes With Hybrid March 2025 computationally
Artificial Humming Bird Pity expensive.
Beetle Algorithm

13
A Comparison of Signal IIEEE Access, Oct Deep learning with multi- Uses full PSG for
Combinations for Deep Learning- 2023 signal inputs (PPG, SpO₂, higher accuracy;
Based Simultaneous Sleep Staging EEG) limited suitability for
and Respiratory Event Detection lightweight home
systems
Cascade Windows-Based Multi- IEEE Access, Aug Multi-stream CNN with High memory and
Stream CNN Framework for Early 2023 cascading windows and processing
Detecting In-Sleep Stroke Using ensemble (EDIS- requirements; longer
Wristbands ResNet50) inference time
Noncontact Detection of Sleep IEEE Sensors FMCW radar with Accuracy depends on
Apnea Using Radar and Journal, Oct 2024 Expectation–Maximization radar positioning and
Expectation–Maximization (EM) algorithm environment; not
Algorithm ML-based
Sleep Apnea Events Recognition IEEE OJEMB, Dec Machine learning on Differentiation
Based on Polysomnographic 2024 multi- channel PSG between apnea types
Recordings: A Large-Scale Multi- signals less accurate
Channel Machine Learning (62.9%); requires full
Approach PSG setup
A Novel Multimodal Deep IEEE Access, 2022 Multimodal DL combining Limited to two
Learning Approach With Loss ECG and respiration with physiological signals;
Function for Detection of Sleep weighted focal loss not optimized for
Apnea Events real-time wearable
use
Feature-Based Detection and IEEE J-BHI, March Remote Needs high-quality
Classification of Sleep Apnea and 2025 photoplethysmography and infrared imaging
Hypopnea Using Multispectral multispectral feature-based setup; real-world
Imaging classification scalability untested
Explainable Non-Contact Sleep IEEE Access, 2022 Non-contact IR Uses full PSG for
Apnea Syndrome Detection Based thermography + higher accuracy;
on Comparison of Random Forests Explainable Random limited suitability for
Forest classification lightweight home
systems
Central Sleep Apnea Detection by IEEE Transactions Multi-kernel CNN + Bi- Only detects central
Means of Finger on Biomedical LSTM with attention and apnea, not
Photoplethysmography Engineering, 2019 decision fusion obstructive; limited
accuracy

14
 2.3 SUMMARY OF LITERATURE SURVEY

The chapters literature survey has traced the evolving trend of deep learning methods to
diagnose sleep apnea. The movement of diagnosis from standard PSG to AI-based acoustic
monitoring is a demonstration of the rising importance of non-invasive devices in the medical field.
Some of the major points from the research are:
 Definitely CNN architectures are the best to extract features in spatial and frequency-domain
from sound of snore
 The systems which are LSTM-based have the most accurate temporal apnea pattern to depict in
the consecutive breathing cycles
 The hybrid CNNLSTM models to a great extent outperform the classification by combining
both spatial and temporal learning capabilities
 The transfer will be much more accurate if scenarios will be limited but hardware will be increased.
 The attention mechanisms enable the interpretability of the model by allowing it to focus on
the critical apnea events
 The multimodal as well as the wearable systems increase the robustness of the detection,
however, the comfort and cost-efficiency may be compromised
 The existence of deep learning simple models is very important for local observation at real
time and low power in a home environment
 There are still great difficulties in solving the problem of insufficient data, noise and
understanding the model

15
 CHAPTER 3

PROPOSED MODEL

3.1 INTRODUCTION

This project delves into the utilization of deep learning for the purpose of sleep apnea
identification based on snoring sound recordings. By isolating the acoustic features of the sounds, the
system figures out the regularities that accompany apneic events. The model uses a hybrid
architecture combining Convolutional Neural Networks (CNNs) to extract short- term temporal and
frequency features and Long Short-Term Memory(LSTM) networks that capture sequential
dependencies across time.
The dataset that was used consists of audio samples with labels of normal and apneic snoring.
To make the recordings more suitable for the downstream tasks, they are trimmed, normalized, and
converted into mel-spectrograms, thus giving the model an opportunity to learn more distinctive acoustic
patterns. The goal of this method is to reduce the dependence on the standard clinical setup by
offering a non-invasive and easy-to-use diagnostic method for the earlydetectionofsleep apnea.
Besides being technically innovative, the system can be very helpful in the telemedicine field,
thus ensuring remote health monitoring and timely medical intervention. In short, the project is
an exemplification of how artificial intelligence can make healthcare more accessible and patient
outcomes better through intelligent, data-driven insights.
The training procedure starts with the determination of an appropriate loss function, usually
binary cross-entropy if the task is sleep apnea classification. The optimizer, which is normally Adam,
makes the network's weights update during backpropagation very quick and effective. The training is
done in a number of small batches ofpreprocessed mel-spectrograms which ensures the learning
process is stable.
The most important evaluation metrics such as accuracy, precision, recall, and F1-score serve as a good
measure of model performance. Checkpointing is a way to store the model weights that result in the
best performance at some point later inference.

16
3.1.1 System requirements

Software:

 Python 3.12

 Libraries: PyTorch, torchaudio, librosa, matplotlib, numpy, graphviz

 IDE: Visual Studio Code / Python IDLE

 Dataset: Snoring Dataset (0 – Normal, 1 – Apnea)

3.1 Dataset Description

1. The dataset contains snore audio clips categorized as:

1. 0 – Normal snoring

2. 1 – Apnea snoring

2. Each clip is a .wav file recorded under similar conditions.

3. Example folder

structure:

data/

├── 0/

└── 1/

3.1.2 Data processing

1. Loaded .wav files using [Link]().

2. Converted audio signals into Mel-spectrograms (frequency vs time).

3. Normalized the data and labeled them.

4. Split into training and testing sets.

17
Model Architecture :

 LSTM Layer: Captures temporal dependencies over time

 Fully Connected Layer: Outputs the final classification (Normal / Apnea)

 CNN Layers: Extract spatial (spectral) features from spectrograms

Training Process

• Batch size: 2
• Epochs: 5
• Optimizer: Adam
• Loss Function: CrossEntropyLoss
• Device: CPU (or GPU if available)
• The model was trained using [Link] on 1000 samples.

Testing and Results

• Model tested using [Link].

• Accuracyobtained: 51%
• Enter .wav filename:1.9
• Prediction: Normal

3.1.3 Algorithm and Mathematical Model

The CNN-LSTM model was operating by leveraging a carefully preprocessed snore sound
dataset which contained both normal and apnea-affected recordings. In the training period, the CNN
layers were able to capture the most significant spectral and temporal features of the audio inputs,
whereas the LSTM layers were figuring out the sequential dependencies and breathing rhythm
patterns in the snore signals.
The work was done in Python by means of the PyTorch and torchaudio libraries to promote
flexibility and efficient processing of audio data

18
A. Convolutional Layer

The CNN learns local spatial patterns from the Mel-spectrogram:

𝒀 = 𝒇(𝑾 ∗ 𝑿 + 𝒃)

Where:

• 𝑿= input spectrogram

• 𝑾= convolution kernel

• ∗= convolution operation

• 𝒃= bias term

• 𝒇= activation function (ReLU)

B. Pooling Layer

Reduces dimensionality and retains dominant features:

Y = max(Y 𝒊:𝒊 + k,𝒋:𝒋 + k)

where k is the pooling window size.

C. LSTM Layer

Captures sequential dependencies in time frames:

𝒊𝒕 = 𝝈(𝑾𝒊 ⋅ [𝒉𝒕−𝟏, 𝒙𝒕] + 𝒃𝒊)

𝒇𝒕 = 𝝈(𝑾𝒇 ⋅ [𝒉𝒕−𝟏, 𝒙𝒕] + 𝒃𝒇)
𝒐𝒕 = 𝝈(𝑾𝒐 ⋅ [𝒉𝒕−𝟏, 𝒙𝒕] + 𝒃𝒐)
𝒄𝒕 = 𝒇𝒕 ∗ 𝒄𝒕−𝟏 + 𝒊𝒕 ∗ 𝐭𝐚𝐧𝐡(𝑾𝒄 ⋅ [𝒉𝒕−𝟏, 𝒙𝒕] + 𝒃𝒄)
𝒉𝒕 = 𝒐𝒕 ∗ 𝐭𝐚𝐧𝐡(𝒄𝒕)

19
 where 𝒊𝒕, 𝒇𝒕, 𝒐𝒕are input, forget, and output gates respectively.

D. Fully Connected Layer

Performs final classification:

𝑦^= Softmax(𝑾𝒇𝒄 ⋅ 𝒉𝒕 + 𝒃𝒇𝒄)

3.1.4 System Architecture Diagram

3.1 Block Diagram of System Architecture

3.1.5 Code used:

1. [Link]

import os import
torch import [Link]
as nn

20
from [Link] import Dataset, DataLoader import
torchaudio from model import SimpleCNNLSTM
class SnoreSet(Dataset):
def init (self, root='data', sr=16000, dur=3): [Link], [Link] = [], []
[Link], [Link] = sr, dur
for i in ['0', '1']:
folder = [Link](root, i) if not [Link](folder):
print(f"Folder not found: {folder}") continue
wav_files = [f for f in [Link](folder) if [Link]().endswith('.wav')] for f in wav_files:
[Link]([Link](folder, f)) [Link](int(i))
print(f"Loaded {len([Link])} audio files from '{root}'")
def len (self): return len([Link])
def getitem (self, idx): file_path = [Link][idx]
waveform, sr = [Link](file_path)
if [Link][1] < self.target_len:
pad_len = self.target_len - [Link][1]
waveform = [Link](waveform, (0, pad_len)) else:
waveform = waveform[:, :self.target_len]
mel_spec = [Link](sample_rate=[Link], n_mels=64)
(waveform) mel_spec = [Link]()(mel_spec)
mel_spec = (mel_spec - mel_spec.mean()) / (mel_spec.std() + 1e-9)
label = [Link]([Link][idx], dtype=[Link]) return mel_spec, label
device = 'cuda' if [Link].is_available() else 'cpu'
print(f"Using device: {device}")
dataset = SnoreSet(root='data') if len(dataset) == 0:
raise ValueError("Dataset is empty! Make sure you have .wav files in 'data/0' and 'data/1'.")
loader = DataLoader(dataset, batch_size=8, shuffle=True) model =
SimpleCNNLSTM(n_mels=64).to(device)
opt = [Link]([Link](), lr=1e-4) loss_fn = [Link]()
for epoch in range(5): total_loss = 0
for xb, yb in loader:
xb, yb = [Link](device), [Link](device) out = model(xb)

21
loss = loss_fn(out, yb) opt.zero_grad() [Link]() [Link]()
total_loss += [Link]()
print(f"Epoch {epoch+1}/5 | Avg Loss: {total_loss/len(loader):.4f}")
[Link]('models', exist_ok=True) [Link](model.state_dict(), 'models/[Link]')
print("Training finished! Model saved to 'models/[Link]'")

2. [Link]

import torch import torchaudio

from model import SimpleCNNLSTM
device = 'cuda' if [Link].is_available() else 'cpu' print(f"Using device: {device}")
filename = "your_audio_file.wav" # replace with your file name
model =
SimpleCNNLSTM(n_mels=64).to(device)
model.load_state_dict([Link]("models/[Link]", map_location=device)) [Link]()
paths = [f"data/0/{filename}", f"data/1/{filename}"] file_path = None
for p in paths: try:
[Link](p) file_path = p
break except:
continue
if not file_path:
raise FileNotFoundError(f"File not found in 'data/0' or 'data/1': {filename}")
waveform, sr = [Link](file_path)
if sr != 16000:
waveform = [Link](sr, 16000)
(waveform if [Link][0] > 1:
waveform = [Link](dim=0,
keepdim=True) target_len = 16000 * 3
if [Link][1] < target_len:
pad_len = target_len -
[Link][1]

22
waveform = [Link](waveform, (0, pad_len)) else:
waveform = waveform[:, :target_len]
mel_spec = [Link](sample_rate=16000, n_mels=64)(waveform)
mel_spec = [Link]()(mel_spec)
mel_spec = (mel_spec - mel_spec.mean()) / (mel_spec.std() + 1e-9) mel_spec =
mel_spec.unsqueeze(0).to(device)
with torch.no_grad():
out = model(mel_spec)
pred = [Link](out, dim=1).item(
classes = {0: "Normal", 1: "Apnea"} print(f"Prediction: {classes[pred]}")

3. [Link]

import torch
import [Link] as nn
class SimpleCNNLSTM([Link]):
def init (self, n_classes=2, n_mels=64): super(SimpleCNNLSTM, self). init ()
[Link] = [Link](
nn.Conv2d(1, 16, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(16),
[Link](), nn.MaxPool2d(2, 2),
nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(32),
[Link](), nn.MaxPool2d(2, 2)
)
self.n_cnn_channels = 32
self.n_lstm_input = self.n_cnn_channels * (n_mels // 4)
[Link] = [Link]( input_size=self.n_lstm_input, hidden_size=128, batch_first=True,
bidirectional=True)
[Link] = [Link](128 * 2, n_classes)
def forward(self, x):
batch_size, _, _, time = [Link] x = [Link](x)
c, h, w = [Link][1], [Link][2], [Link][3]

23
 x = [Link](0, 3, 1, 2) # batch, width(time), channels, height x = [Link](batch_size, w, c *
h) out, _ = [Link](x) out = out[:, -1, :] out = [Link](out) return out

3.1.6 Future Research Scope

1. Collect larger, real-world snore datasets.

2. Explore transfer learning using pretrained models like VGGish or YAMNet.

3. Implement attention mechanisms in LSTM for temporal focus.

4. Deploy as a mobile app or IoT-based snore monitor.

5. Combine with physiological sensors (SpO₂, ECG) for multimodal diagnosis.

Fig 3.2 Output of the Proposed Model

24
 Fig 3.3 Snore Analysis using LSTM-CNN

The model training phase consisted of inputting the preprocessed snore sound dataset into
the CNN LSTM architecture that was created. As the training progressed, the Convolutional
Neural Network (CNN) layers were able to extract spectral and temporal features of the snore
audio signals in an automated manner, whereas the Long Short- Term Memory (LSTM) layers
were able to capture the sequential dependencies in the sound patterns. The dataset was composed
of audio samples labeled as normal and apnea-affected snores, thus enabling the model to learn
how to distinguish between the two effectively

Once trained, the model underwent testing with real snore audio samples to check for its
prediction accuracy. Basically, each input audio file in .wav format was run through the system to
get a classification output that shows whether it is a normal or apnea condition. Results were
obtained on the spot, thus showing that this method can be used in practice for a quick diagnosis of
sleep apnea at an early stage. The output files recorded the device, preprocessing, and final results,
thus serving as evidence that the CNNLSTM model was capable of snore pattern separation based
on the sound features.

3.1.6 Dataset Description

The dataset used for this project includes audio samples of snoring and non-snoring sounds,
which are representative of both normal breathing and apnea-affected breathing. The data was
arranged in a very neat manner into folders for training and testing, which were usually labeled as
"0" for normal snoring and "1" for apnea-related snoring, to ensure that there would be no issues.

25
1. Model Architecture Explanation
The CNN-LSTM hybrid model devised in this work combines the feature extraction power of
Convolutional Neural Networks (CNNs) with the time-sequence learning ability of Long Short-Term
Memory (LSTM) networks. The convolution layers capture localized spectral patterns from the snoring
sounds very effectively, thus the most relevant frequency components associated with apnea events are
highlighted. After that, the feature maps are reshaped and fed to the LSTM layers that study the time-
dependent evolution of the sounds in order to recognize the breathing normality or apnea episodes

2. Challenges and Limitations

The proposed system is a great idea to solve the problem, but during the development phase, it was
accompanied by various challenges and obstacles. One of the significant problems was the dataset
imbalance, in which the number of normal snore samples was twice as high as the apnea samples.
Moreover, the presence of the background noise, different quality of recordings, and small amount of
labeled data made it a hard task of achieving very high accuracies consistently. The limitations of the device
also hindered the model's complexity, thus the number of layers and the total learning capacity were
reduced.

3. Future Enhancements
Additionally, many enhancements of the present project can be considered in subsequent phases to
finally achieve higher effectiveness. The expansion of the dataset with real-world recordings of various
people will make the model more generalizable. By employing data augmentation methods, e.g., pitch
shifting, noise addition, or time-stretching, a problem of overfitting and a small number of samples can be
resolved. If the trained model is deployed on mobile or IoT-based devices, the monitoring of sleep apnea at
home will be possible in real-time and, thus, with cloud connectivity, the results can be sent directly to
healthcare professionals for their remote analysis. They may also consider exploring transformer-based
audio models or self- supervised learning approaches for improved performance and adaptability in
iterations to come.

26
Metrics Used:

 Accuracy: Percentage of the correct predictions made from the total samples.
 Loss: Is the measure of the models errors in prediction and misreadings.
 Confusion Matrix: it compares true with predicted classes overall.

Fig 3.4 Flowchart of the model

27
3.1.7 Future Research Scope

• Collect larger, real-world snore datasets.

• Explore transfer learning using pretrained models like VGGish or YAMNet.

• Implement attention mechanisms in LSTM for temporal focus.

• Deploy as a mobile app or IoT-based snore monitor.

• Combine with physiological sensors (SpO₂, ECG) for multimodal diagnosis.

3.2 CONCLUSION
• This project is a perfect example of how deep learning can change the face of medical diagnosis,
bymaking it less invasive and more accessible. The system, which utilizes a fusion of
Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, is
able to identify whether the snore audio signals are from a healthy individual or one with apnea
• The technique dispenses with the intricacies and high prices of traditional polysomnography
tests, thereby providing a cheap way to detect sleep disorders in their early stages. The CNN
layers are capable of extracting spatial and spectral features from the Mel-spectrogram
representation of the audio, whereas the LSTM layers are able to capture temporal dependencies
across time frames. Such a hybrid architecture allows the model to identify frequency-based and
sequential patterns in snoring, thus making the predictions more accurate and trustworthy
• To sum up, the CNNLSTM-based snore sound analysis system is a major milestone in non-
invasive, accessible, and cost-effective sleep apnea screening. The system not only reduces the
dependence on clinical equipment such as polysomnography but also makes use of deep learning
to detect minute changes in snore acoustics
• Such an innovative AI-driven audio diagnostic technique as this is proof of the possibility of
integrating it into health monitoring devices that are either mobile or wearable. By investing in
further refinement and using larger datasets that are more diverse, the models accuracy and
consistency will be at a level where it can be trusted for real-world healthcare applications, thus,
allowing early diagnosis for better sleep health management
• The research therefore becomes a stepping stone, not only in detecting sleep apnea, but also in
eventually employing similar deep-learning audio analysis techniques for the treatment of
respiratory or sleep-related disorders.

28
 3.3 REFERENCES

[1] A. Al-Kabbany, H. Elayan, and S. Shams, “Automatic snoring detection using hybrid
1D–2D convolutional models,” Scientific Reports, vol. 13, no. 1, p. 15670, 2023. [Online].
Available: [Link]

[2] T. Ahmed and Z. Chen, “Transfer learning for pediatric sleep apnea detection using
acoustic models,” arXiv preprint arXiv:2509.15008, 2025. [Online]. Available:
[Link]

[3] I. Almendros, J. Montserrat, and R. Farré, “Non-invasive detection of obstructive sleep

apnea using sound analysis,” Sleep Medicine Reviews, vol. 44, pp. 45–53, 2019.

[4] P. Ghosh and R. Das, “LSTM-based apnea detection from sleep sound using time–
frequency analysis,” Procedia Computer Science, vol. 216, pp. 857–865, 2022.

[5] S. Kim, J. Park, and Y. Lee, “Detection and classification of obstructive sleep apnea
using audio spectrograms with pretrained CNN and Bi-LSTM models,” Electronics, vol.
13, no. 13, p. 2567, 2023. [Online]. Available: [Link]
9292/13/13/2567

[6] X. Li, Y. Zhang, and H. Zhao, “Deep recurrent neural networks for apnea detection
from respiratory signals,” Sensors, vol. 20, no. 18, p. 5037, 2020. [Online]. Available:
[Link]

[7] T. Penzel and C. Schöbel, “Acoustic analysis of snoring sounds for sleep
apnea detection,” Sleep and Breathing, vol. 24, no. 4, pp. 1475–1484, 2020.

[8] M. Rahman, M. Islam, and F. Hossain, “Deep learning approach for sleep apnea
detection using snoring sound,” Biomedical Signal Processing and Control, vol. 68, p.
102709, 2021.

[9] F. Rofouei and Y. Zeng, “Audio feature extraction and deep learning for real-time
snore detection in mobile devices,” IEEE Transactions on Biomedical Engineering, vol.
68, no. 9,
pp. 2673–2684, 2021.

[10] A. Sharma and N. Verma, “Low-cost snoring sound analysis system for sleep apnea
detection using deep CNN,” IEEE Sensors Journal, vol. 23, no. 7, pp. 11259–11269,
2023.

[11] M. Uddin, R. Chowdhury, and M. Rahman, “Automatic detection of sleep apnea

events from snore sound using CNN-LSTM model,” IEEE Access, vol. 10, pp. 74582–
74591, 2022.

29
[12] J. Wang and K. Liu, “An intelligent framework for sleep apnea classification using
CNN with mel-spectrogram features,” Journal of Ambient Intelligence and Humanized
Computing, vol. 13, no. 5, pp. 2341–2355, 2022.

30