Description:FIELD OF THE INVENTION
The present invention broadly relates to the field of machine learning based health monitoring. More particularly, the present invention relates to a non-invasive portable system and method for detecting sleep disorders through advanced machine learning-based photoplethysmography (PPG) signal analysis technique, so that timely medical procedures are adapted to cure the disorders.

BACKGROUND OF THE INVENTION
Sleep is an essential component of human health, influencing both physical and mental well-being. It plays a critical role in memory consolidation, emotional regulation, immune function, and overall productivity. Despite its importance, many individuals experience sleep disturbances that can lead to chronic disorders. Sleep disorders such as insomnia, sleep apnea, and restless leg syndrome disrupt normal sleep patterns and have far-reaching consequences, including daytime fatigue, reduced cognitive performance, increased risk of cardiovascular diseases, and mental health issues like anxiety and depression. Understanding and addressing these disorders is vital to improve quality of life and reducing associated health risks.

Sleep disorders can arise from various factors, including stress, lifestyle habits, medical conditions, and environmental influences. For instance, insomnia, one of the most common sleep disorders, is often triggered by anxiety or irregular sleep schedules, while sleep apnea is typically caused by physical obstructions in the airway. If left untreated, these disorders can severely impact on daily functioning and long-term health. Therefore, the early diagnosis and prompt treatment is inevitable to ameliorate the sleeping condition of such people/patients.

To diagnose sleep disorders, clinicians traditionally rely on polysomnography (PSG), the gold standard method for sleep analysis. The PSG provides comprehensive data by monitoring various physiological signals, such as brain activity (EEG), muscle movements (EMG), and heart rate (ECG). Despite its accuracy, this method has significant limitations. It requires patients to spend a night in a specialized sleep lab, connected to numerous sensors and wires, which can be uncomfortable and affect natural sleep patterns. Additionally, the PSG setups are costly, time-consuming, and dependent on trained professionals, making them less accessible for routine or widespread use. The reliance on ECG signals introduces a level of invasiveness and complexity not suited for portable or wearable systems. The use of EEG signals requires more intrusive setups.
Further, the commercially available wearable devices (such fitness bands/rings/belts) primarily focus on tracking sleep stages, such as light, deep, and REM sleep, but they do not provide the detailed signal data needed to analyse and diagnose sleep disorders like insomnia or sleep apnea. Additionally, these devices can be quite expensive, making them inaccessible for common patients.

To address the limitations the conventional/existing approaches, there is a need exploring an alternative data source such as photoplethysmography (PPG) signals/data for detection of different type of sleep disorders. The PPG is a non-invasive optical technique that measures blood volume changes in the skin using a simple sensor. Unlike PSG, the PPG requires only a single-channel sensor, which is compact, cost-effective, and can be integrated into wearable IoT devices for easy use. This PPG signal analysis approach can further be fine-tuned with advanced machine learning techniques to design an accessible system for health monitoring, especially sleep disorder prediction with improved accuracy and high precision. Moreover, it is desired to develop a technically advanced non-invasive and portable system and method for detecting different types of sleep disorders, which includes all the advantages of the conventional/existing techniques/methodologies and overcomes the deficiencies of such techniques/methodologies.

OBJECT OF THE INVENTION
It is an object of the present invention to minimize usage of multiple expensive sensors or medical/diagnostic devices and the associated cost used in sleep disorder detection.

It is another object of the present invention to leverage potential of PPG signals to detect sleep disorders in a simple, effective, and non-invasive manner.

It is one more object of the present invention to develop a reliable machine learning based PPG signal/data processing technique.

It is a further object of the present invention to develop a system and method for detecting various sleep disorders with improved accuracy and high precision.

SUMMARY OF THE INVENTION
In one aspect, the present invention provides a system for detecting various sleep disorders. The system comprises a photoplethysmography (PPG) sensor coupled to a wearable device to capture real time PPG signals of patients; and a user device wirelessly connected with the wearable device to receive and the PPG signals. The user device has installed therein an application interface configured with a preprocessing module, a wavelet decomposition module, a statistical indicator module, a machine learning (ML) classifier model, and a report generation module. The wavelet decomposition module and the statistical indicator module jointly act as a feature extractor adapted to extract all meaningful (relevant) features associated with physiological changes encountered during various sleep disorder conditions which somehow affect the PPG signals. The preprocessing module segments the PPG signals into defined time interval segments. The feature extractor decomposes the segmented signals into different frequency sub-bands through computation of approximation (low-frequency components) and detail (high-frequency components) coefficients followed by down sampling over five levels in an Optimal Biorthogonal Wavelet Filter Bank (OBWFB), and then applies Hjorth parameters to extract meaningful features associated with various sleeping conditions from each of the frequency sub-bands. The ML classifier is trained to classify the extracted features in a healthy class or a sleep disorder class selected from insomnia, sleep-disordered breathing (SDB), periodic limb movements (PLM), narcolepsy, rapid eye movement behaviour disorder, nocturnal frontal lobe epilepsy (NFLE). Finally, the classification results are used by the report generation module to generate diagnostic reports in real-time.

Other aspects, advantages, and salient features of the present invention will become apparent to those skilled in the art from the following detailed description, which delineate the present invention in different embodiments.

BRIEF DESCRIPTION OF DRAWINGS
These and other features, aspects, and advantages of the present invention will become better understood when the following detailed description is read with reference to the accompanying figures.

Fig. 1 illustrates hardware setup of the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 2 illustrates application interface configuration for PPG signal processing/analysis used in the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 3 illustrates graphical presentations of wavelet function and scaling function of an Optimal Biorthogonal Wavelet Filter Bank (OBWFB) as employed in the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 4 illustrates Ensemble Bagged Decision Trees (EBDT architecture used for training of a machine learning classifier employed in the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 5 illustrates Support Vector Machines (SVM) architecture used for training of a machine learning classifier employed in the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 6 illustrates K-Nearest Neighbours (KNN) architecture used for training of a machine learning classifier employed in the sleep disorder detection system, in accordance with an embodiment of the present invention.

Fig. 7 illustrates method steps of detecting sleep disorder, in accordance with an embodiment of the present invention.

List of reference numerals
100 PPG sensor
200 wearable device
300 user device
400 application interface (mobile App)
402 preprocessing module
404 Optimal Biorthogonal Wavelet Filter Bank (OBWFB)
406 statistical indicator module (Hjorth parameters)
408 machine learning classifier model
410 diagnostic report generation module
500 Cloud server

DETAILED DESCRIPTION OF THE INVENTION
Various embodiments described herein are intended only for illustrative purposes and subject to many variations. It is understood that various omissions and substitutions of equivalents are contemplated as circumstances may suggest or render expedient, but are intended to cover the application or implementation without departing from the scope of the present invention. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting.

The use of terms “including,” “comprising,” or “having” and variations thereof herein are meant to encompass the items listed thereafter and equivalents thereof as well as additional items. Further, the terms, “an” and “a” herein do not denote a limitation of quantity, but rather denote the presence of at least one of the referenced items.

In accordance with an embodiment of the present invention, as shown in Fig. 1, the system for detecting various sleep disorders is depicted. The system comprises: a photoplethysmography (PPG) sensor (100), an IoT enabled wearable device (200), a user device (300), an application interface (400), and a cloud server (500). The PPG sensor (100) is coupled to a wearable device (200) adapted to capture bio-signals (PPG signals) associated with physiological changes of a patient, and wirelessly transmits the raw PPG signals to the user device (300). The application interface (400) is installed in the user device (300), and configured to process the PPG signals using advanced machine learning (ML) techniques to predict whether the patient is healthy or suffers from any type of sleep disorders. The user device (300) has a memory (for storing all processor executable instructions/codes), a processor (for processing the captured data in accordance with the instructions/codes), a display (for showing diagnosis results), and a power circuitry (including rechargeable battery). The cloud server (500) is used to store training datasets and patient history records. The user device (200) communicates with the cloud server (500) over internet.

In an example, the PPG sensor (100) captures real-time changes in blood volume by measuring how light interacts with the skin. The PPG sensor allows for non-invasive and continuous monitoring, eliminating the need for any intrusive procedures. This sensor is seamlessly integrated into the wearable device, ensuring patient comfort and portability while providing consistent and reliable data collection. The wearable device (200) is a smart watch or fitness band. The user device (300) is a smartphone. The application interface (400) is a mobile app (that can be operated in any smartphone operating system). Using Wi-Fi or Bluetooth connectivity, the smart watch securely transfers the PPG signal data to a user’s mobile device app. The mobile app then processes the data, performing the necessary analysis and providing insights. This setup eliminates the need for large or complex equipment, making the entire the diagnosis procedure simpler, more convenient, and user-friendly.

In accordance with an embodiment of the present invention, as shown in Fig. 2, the application interface (400) comprises one or more modules such as a preprocessing module (402), a wavelet decomposition module (404), a statistical indicator module (406), a machine learning (ML) classifier model (408), and a report generation module (410). All these modules/models are stored in the memory, and executed by the processor to carry out various data processing/analysis operations in specific sequences/capacities to achieve the diagnosis results reliably.

In accordance with an embodiment of the present invention, the preprocessing module (402) is configured to clean up the raw PPG signals by reducing noise to ensure better accuracy, and segment the PPG signals into defined time interval segments. In an example, a code is written in Python to break the raw PPG signal into smaller chunks, and each chunk lasts for 30 seconds, making it easier to process and analyse.

In accordance with an embodiment of the present invention, the wavelet decomposition module (404) includes an Optimal Biorthogonal Wavelet Filter Bank (OBWFB) configured to perform Wavelet Function and Scaling Function. The Wavelet Function captures the high-frequency components (details) in the signal, where the Scaling Function captures the low-frequency components (approximation) in the signal. The OBWFB is designed for signal decomposition and reconstruction. Unlike orthogonal filters (e.g., Daubechies), the biorthogonal filters have separate analysis and synthesis filters, allowing for more flexibility and adaptability. The wavelet decomposition involves iteratively applying analysis filters to break down a signal into low-frequency (approximation) and high-frequency (detail) components. Particularly, as shown in Fig. 3, the OBWFB utilizes an analysis low-pass filter (H_0) for approximation coefficient computation (i.e., performing scaling function), and an analysis high-pass filter 〖(H〗_1) for detail coefficient computation (i.e., performing wavelet function).
For the OBWFB, the frequency-response H(ω) of the filter h(n) can be presented as equation 1.

H(ω)=h(0)+2∑_(n=1)^(+N)▒〖h(n) cos⁡(ωn) 〗 equation 1
where h(n) represents the impulse response of the analysis filter h_0 (n).

To formulate optimization problems for designing optimal filters of the OBWFB, Objective function (φ) is defined as equation 2, assuming the energy of the filter h(n) to be unity.
φ=α_1 σ_n^2+α_2 σ_ω^2+α_3 E_p+α_4 E_s equation 2

where σ_n^2 and σ_ω^2 are time and frequency variances of the filter with unit energy and can be represented as equations 3-4.
σ_n^2=1/π ∫_0^π▒〖|(dH_0 (ω))/dω|^2 dω〗 equation 3
σ_ω^2=1/π ∫_0^π▒〖ω^2 |H_0 (ω)|^2 dω〗 equation 4
The Ep and Es are the passband and stopband error between the ideal and the desired frequency responses of the filter and can be as equations 5-6.
E_p=1/π ∫_0^(ω_p)▒|1-H_0 (ω)|^2 dω equation 5
E_s=1/π ∫_(ω_s)^π▒|H_0 (ω)|^2 dω equation 6
where, ω_p and ω_s are the passband edge and stopband edge frequencies, respectively.
By defining the following (N+1) × 1 vectors as equations 7-9.
a = [h(0) √2 h(1) … √2 h(N) ]^T equation 7
c(ω) = [1 √2 cos⁡〖(ω)〗 … √2 N cos⁡〖(Nω)〗 ]^T equation 8
f(ω) = [0 -√2 sin⁡〖(ω)〗 … -√2 N sin⁡〖(Nω)〗 ]^T equation 9

The time and frequency variances and passband and stopband errors can be expressed as equations 10-13,
σ_n^2= a^T {∫_0^π▒〖f(ω)f^T (ω)dω/π〗} a = a^T Ta equation 10
σ_ω^2= a^T {∫_0^π▒〖ω^2 c(ω)c^T (ω)dω/π〗} a = a^T Fa equation 11
E_s= a^T {∫_(ω_s)^π▒〖c(ω)c^T (ω)dω/π〗} a = a^T Sa equation 12
E_p= a^T {∫_0^(ω_p)▒〖[c(0) - c(ω)] [c(0) - c(ω)]^T dω/π〗} a = a^T Pa equation 13

Thus, the objective function can be expressed as equation 14.
φ=a^T {α_1 T+α_2 F+α_3 P+α_4 S} a = a^T Ra equation 14

Where, {α_i s.t. 0≤α_i≤1; ∑▒〖α_i=1 〗} are weighting or trade-off factors that provide the trade-off between time and frequency variances as well as pass band and stopband errors. The matrices, T,F,P,S and R∈R^((N+1)× (N+1)) are positive definite matrices.

The filter design problems are formulated as constrained optimization problems. The regularity, halfband and perfect reconstruction conditions are imposed as linear equality constraints in the design variables. In the design of a wavelet filter bank, regularity of the filters is an essential requirement, which is attained by imposing zeros at z = -1 in the case of two-band filter bank. To impose the regularity of order K (i.e. K zeros at z =-1) (vanishing moments), the frequency response H(ω) should satisfy the condition as expressed as equation 15.
((d^l H(ω))/(dω^l ))_(ω=π)=0, l=0,1,2,…,K-1 equation 15
Then, the filter is said to be K–regular filter. It is to note, that in case of type-1 filters, the regularity order must be even, i.e K = 2M. The regularity condition mentioned above, can be expressed as the group of linear equalities, Aa = 0. The entry corresponding to the 〖(k,l)〗^th element of R∈R^((N+1) × M) can be given by equation 16,
[A]_(k,l) = {█(1 k,l =0 @√2〖(l)〗^2k 〖(-1)〗^l otherwise)┤ equation 16
Here, the analysis lowpass filter is constrained to be a halfband filter. If the filter h(n) is a halfband filter then it must satisfy the following condition as expressed in equation 17.
h(2n)=0 except for n=0 equation 17
Then the halfband condition has been formulated as a collection of linear equalities, B a = 0. where, the matrix B∈R^((N+1) × ((N-1))⁄2) whose 〖(k,l)〗^th entry is given as equation 18:
[B]_(k,l) = {█(1⁄(√2) k = 0,1...N ,l =2,4,...(N-1) @0 otherwise )┤ equation 18
In the case of a halfband filter, N must be an odd integer.

Now, the parameters selected for both analysis filters are as follows: number of vanishing moments = 4, length = 15, ω_p = 0.3801π , ω_s = π - ω_p, and the weighting factors as α_1= 0, α_2= 0 and α_3= 0.5.

In accordance with an embodiment of the present invention, the OBWFB is configured to decompose the segmented signals into different frequency sub-bands through computation of the approximation and detail coefficients followed by down sampling over five levels. The approximation coefficients represent low-frequency components, where the details coefficients represent high-frequency components.

The wavelet and scaling functions to decompose a signal (x(n)) into the approximation coefficients (A[n]) (through the analysis low-pass filtering) and the detail coefficients (D[n]) (through analysis high-pass filtering) at five levels are expressed in equations 19-22.

Initial Decomposition: The input signal (x[n]) is passed through the analysis lowpass (H_0) and high pass (H_1) filters:
A_1 [n]=∑_k▒〖H_0 [k] 〗⋅x[n-2k] "(Approximation Coefficients) " equation 19
D_1 [n]=∑_k▒〖H_1 [k] 〗⋅x[n-2k] "(Detail Coefficients)" " " equation 20

Recursive Decomposition: The approximation coefficients (A_1 [n]) are recursively decomposed at each subsequent level.
A_(j+1) [n]=∑_k▒〖H_0 [k] 〗⋅A_j [n-2k] "(Approximation Coefficients) " equation 21
D_(j+1) [n]=∑_k▒〖H_1 [k] 〗⋅A_j [n-2k] "(Detail Coefficients)" equation 22

Where, j = 1,2,3,4, culminating in A and D_1,D_2,D_3,D_4,D_5

Both approximation and detail coefficients are down sampled by a factor of two to reduce (temporal and frequency) resolution. Down sampling removes redundant information, simplifying the signal while preserving essential features. This helps in analysing the signal at different levels of detail. Resolution in wavelet transforms involves a trade-off between time and frequency precision. The five-level OBWFB wavelet transform operation is presented in Table 1.
Table 1

As shown in Table 1, in the five-level OBWFB based wavelet transform, the input signal (x(t)) is repeatedly decomposed into approximations (A) and details (D1-D5) over five levels corresponding to five frequency sub-bands such as 0-2Hz, 2-4Hz, 4-8Hz, 8-16Hz, 16-32Hz, and 32-64Hz. At the final level, the signal is represented as equation 23.

x(t)=A (for 0-2Hz)+D5 (for2-4Hz)+D4(for 4-8Hz)+D3(for 8-16Hz)+D2(for 16-32Hz)+D1(for 32-64Hz) equation 23

In accordance with an embodiment of the present invention, the statistical indicator module (406) includes Hjorth parameters configured to extract meaningful features associated with various sleeping conditions from each of the frequency sub-bands. These parameters give us three important values associated with activity, mobility, and complexity of each decomposed frequency sub-bands.

The activity tells how strong or powerful the signal is. It is calculated using the spread of the signal values, also known as variance. In other words, it shows how much the signal varies over time. The activity is defined as a variance of the signal, as mathematically expressed in equation 24.

"Activity"=1/N ∑_(i=1)^N▒(x_i-x ̅ )^2 equation 24
Where x_i is each data point in the signal, x ̅ is the average of the signal, N is the total number of data points.

The mobility gives an idea of the signal's overall frequency. It is calculated by comparing how quickly the signal changes (using its derivative) to the total changes in the signal. This helps us understand how smooth or sharp the signal is. The mobility is defined as a square root of the ratio of the variance of the signal’s first derivative to its original variance, as mathematically expressed in equation 25.
"Mobility"=√((σ_(x^')^2)/(σ_x^2 )) equation 25
Where x^' is the derivative of the signal (how fast the signal is changing), σ_(x^')^2 is the variance of the derivative, σ_x^2 is the variance of the original signal.

The complexity measures how the frequency of the signal changes over time. It shows whether the signal is more regular or has a lot of sudden variations. The complexity is defined as a ratio of the mobility of the first derivative (x^') to the mobility of the original signal (x), as mathematically expressed in equation 26.

"Complexity"=〖"Mobility" 〗_(x^' )/〖"Mobility" 〗_x equation 26

By applying these Hjorth parameters to each sub-band, the detailed features about how the signal behaves over time and across different frequencies are extracted. These features are crucial for identifying patterns that could indicate different types of sleep disorders.

In accordance with an embodiment of the present invention, the machine learning classifier model (408) is trained to classify the extracted features in a healthy class or a sleep disorder class selected from insomnia, sleep-disordered breathing (SDB), periodic limb movements (PLM), narcolepsy, rapid eye movement behaviour disorder (REMBD), nocturnal frontal lobe epilepsy (NFLE).

In accordance with an embodiment of the present invention, the report generation module (410) generates final diagnostic reports indicating whether the patient is heathy or has any specific disorder based on the classification outputs. The reports (classification results) are instantly shown in the display of the user device (mobile app). Such report provides a clear report on an individual’s sleep health, and indicates whether the person is diagnosed with a sleep disorder or not. The best part is that this report is sent directly to the user’s mobile phone, making it easy to access and understand. Additionally, all the collected data, along with the classification results, are securely stored in the cloud server (500). This allows doctors to remotely monitor an individual’s sleep health and the quality of their sleep without needing the patient to visit the clinic frequently. The data stored on the cloud can also be used for further analysis and improvements. Researchers can study this data to gain deeper insights into sleep disorders and develop better methods for diagnosis and treatment, ensuring more efficient and advanced solutions.

SDB includes conditions like obstructive sleep apnea, where breathing stops and starts during sleep due to airway obstruction. Symptoms include loud snoring, gasping for air, and excessive daytime sleepiness. SDB disrupts normal sleep cycles, reducing overall sleep quality.

PLM causes involuntary leg movements during sleep, often resulting in arousals and fragmented sleep. Symptoms include repetitive leg twitching or jerking, which can make you feel unrefreshed in the morning.

Narcolepsy is a neurological disorder that causes excessive daytime sleepiness, sudden muscle weakness (cataplexy), vivid dreams, and sleep paralysis.

REMBD is a condition where people physically act out vivid dreams, often with violent or sudden movements, during Rapid Eye Movement (REM) sleep. Unlike normal REM sleep, where the body is paralyzed, individuals with RBD retain muscle activity, leading to these behaviours.

NFLE involves seizures originating in the frontal lobe during sleep, often causing unusual movements, vocalizations, or behaviours.

In accordance with an embodiment of the present invention, these disorders are interconnected because they affect the way human brain and body behave during sleep. For example, poor breathing during SDB can interrupt sleep cycles, leading to symptoms of insomnia or excessive leg movements. Similarly, issues in the brain's regulation of sleep stages, as in narcolepsy, can lead to vivid dreams or movements during REM sleep, which may resemble symptoms of REMBD or NFLE. The present invention is focused on using the PPG signals to directly classify into specific sleep disorders like insomnia, SDB, PLM, narcolepsy, REMBD, or NFLE. The combined approach of OBWFB (wavelet transform) with Hjorth parameters and ML model not only help in identifying sleep stages but also detect patterns related to these specific sleep disorders using the same PPG signals. This approach can provide deeper insights into these conditions and make sleep disorder detection more efficient and accessible, enhancing how the medical experts understand and treat these disorders.

In accordance with an embodiment of the present invention, the machine learning classifier (408) is trained through any of Ensemble Bagged Decision Trees (EBDT), Support Vector Machines (SVM), and K-Nearest Neighbours architectures with ten-fold cross-validation. These ML classifiers analyse specific patterns from the PPG signal and decide whether the person is healthy or has a specific sleep disorder. This combined approach makes the system more flexible and reliable, as it can work well with various types of signals.

The EBDT classifier is an ensemble learning method that combines multiple decision trees to improve accuracy and robustness. The ensemble structure consists of a collection of decision trees, where each tree is trained on a unique subset of the training data, created through bootstrap sampling (random sampling with replacement). This ensures diversity among the trees, which is essential for the ensemble's effectiveness. As shown in Fig. 4, the architecture of EBDT classifier illustrates the flow of data through the ensemble. The training dataset is divided into multiple bootstrap samples, each used to train a separate decision tree. Once the trees are trained, they work collaboratively to make predictions. For classification tasks, the outputs of all trees are aggregated through a majority voting mechanism to determine the final class label. This structured approach leverages the strengths of individual trees while mitigating their weaknesses, resulting in a more accurate and reliable model.

The Support Vector Machine (SVM) is a supervised learning algorithm primarily used for classification tasks. The key objective of SVM is to find the optimal hyperplane that best separates the data points of different classes in the feature space. For linearly separable data, the algorithm determines a hyperplane that maximizes the margin between the closest data points of each class, known as support vectors. These support vectors are crucial as they define the boundary and influence the model’s decision-making. As shown in Fig. 5, the architecture of SVM illustrates a feature space containing data distributed among classes. The optimal hyperplane is depicted as a straight line (or a plane in higher dimensions) that divides the classes with the maximum margin. For cases where the data is not linearly separable, SVM employs kernel functions to transform the data into a higher-dimensional space where a linear separation becomes possible. The classification process in SVM is robust, particularly for high-dimensional datasets or when there are clear class boundaries. Its reliance on support vectors ensures that the model focuses on the most critical data points, making it efficient and effective for classification tasks.

The K-Nearest Neighbours (KNN) classifier is a simple and effective algorithm used for binary classification. It classifies a new data point by finding the K-nearest neighbours in the training dataset based on a distance metric, such as Euclidean distance. The class of the new data point is determined by the majority class of its K nearest neighbours. As shown in Fig. 6, the architecture of KNN involves calculating distances between the new point and every point in the training set, identifying the closest neighbours, and assigning the class label based on the majority vote among those neighbours. The value of K (the number of neighbours) and the choice of distance metric play a crucial role in the classifier's performance. KNN is highly useful for binary classification tasks where the data points of the two classes are well- separated and the relationship between features is important. It is simple, flexible, and effective for small to medium-sized datasets, but can become computationally expensive as the dataset size increases.

In accordance with an embodiment of the present invention, as shown in Fig. 7, the method for detecting sleep disorders is depicted. The method comprises steps of: capturing (S1) signals associated with physiological changes of the patient in sleeping stage using the PPG sensor (100) based wearable device (200); transmitting (S2) the PPG signals wirelessly from the wearable device (200) to the user device (300) for their processing and analysis through the application interface (400) installed therein; segmenting (S3) the PPG signals into defined time interval segments; decomposing (S4) the segmented signals into different frequency sub-bands through computation of approximation coefficients (low-frequency components) and detail coefficients (high-frequency components) followed by down sampling over five levels using the OBWFB; applying (S5) Hjorth parameters to extract meaningful features associated with various sleeping conditions from each of the frequency sub-bands; training (S6) the machine learning classifier to classify the extracted features in a healthy class or a sleep disorder class selected from insomnia, sleep-disordered breathing (SDB), periodic limb movements (PLM), narcolepsy, rapid eye movement behaviour disorder, nocturnal frontal lobe epilepsy (NFLE); and displaying (S7) the classification results in real-time in the user device display.

For simulation studies of the proposed ML model, the CAP (cyclic alternating pattern) Sleep Database, a publicly available collection of sleep recordings from a hospital in Italy, is accessed on the Physionet website. This database has detailed overnight sleep data from 108 individuals, including 16 healthy people and 92 patients with sleep disorders such as insomnia, epilepsy, REM behaviour disorder (RBD), periodic limb movements (PLM), and others. The recordings include various signals like brain activity (EEG), eye movements (EOG), muscle activity (EMG), heart signals (ECG), and breathing data, collected using advanced medical equipment. Each person’s sleep is recorded for 9-10 hours, capturing a full night. For the ML model development, the PPG signals are taken into consideration, which measure blood flow and heart rate, recorded from the plethysmography (PLETH) channel at 128 Hz. PPG acquisition is a simple, non-invasive approach, making it ideal for practical use. The PPG data from 7 patients with insomnia and 4 healthy individuals from the CAP database are selected for analysis/testing. Hyperparameters used in the ML models are shown in Table 2.
Table 2

Classifiers Hyperparameters Details
EBDT Ensemble Method: Bag, Learner Type: Decision Tree, No. of Splits = 20409, No. of Learners = 70 Ensemble Method: Bagging (Bootstrap Aggregating) trains multiple models using random subsets of the dataset. Reduces overfitting and improves stability and accuracy by averaging predictions.
Learner Type: Decision trees are used as base learners due to their simplicity and high variance. Bagging helps mitigate the high variance of decision trees, making the ensemble more robust.
Number of Splits (20409): Specifies the maximum number of decision nodes in each tree, determining its depth. Deeper trees can capture complex patterns but risk overfitting, mitigated by bagging.
Number of Learners (70): Refers to the total number of decision trees in the ensemble. More trees enhance robustness and accuracy but have diminishing returns after a certain point.
SVM Kernel Function: Gaussian, Kernel Scale: 1.1, Box constraint Level: 1, Multiclass Method: One vs One, Standardize Data: True Kernel Function (Gaussian): Transforms data into a higher-dimensional space using a Gaussian (RBF) kernel. Helps capture non-linear relationships in data.
Kernel Scale (1.1): Controls the spread of the kernel function. It affects the smoothness and flexibility of the decision boundary.
Box Constraint Level (1): Regularization parameter for controlling model complexity. A lower value allows more flexibility, while higher values enforce stricter margins.
Multiclass Method (One-vs-One): Breaks down multiclass classification problems into binary classification tasks. Compares pairs of classes for prediction, suitable for a small number of classes.
Standardize Data (True): Ensures all features have a mean of 0 and a standard deviation of 1. Improves convergence and performance of distance-based and kernel-based methods.
KNN No. of Neighbors: 20, Distance Matric: Jaccard, Distance Weight: Square Inverse, Optimizer: Bayesian Optimizer Number of Neighbors (20): Determines the number of neighbors to consider in K-Nearest Neighbors (KNN) or related algorithms. A higher value smoothens predictions but may reduce sensitivity to local patterns.
Distance Metric (Jaccard): Measures similarity/dissimilarity between samples based on the Jaccard index. Suitable for binary and categorical data.
Distance Weight (Square Inverse): Assigns higher weights to closer neighbors using the inverse square of the distance. Ensures closer points have more influence on predictions.
Optimizer (Bayesian Optimizer): An optimization method used to tune hyperparameters by balancing exploration and exploitation. Efficient for finding optimal parameter values in complex models.

To improve prediction accuracy, the present invention employs 10-fold cross-validation with the ML classification model. The entire dataset is divided into 10 equal (or nearly equal) parts or subsets, called "folds". In each iteration, one-fold is used as the test set, and the remaining 9 folds are combined to form the training set. The model is trained on the training set and then evaluated on the test set. This process is repeated 10 times, with each fold serving as the test set exactly once. After all, 10 iterations, the evaluation metrics (e.g., accuracy, precision, recall, etc.) from each iteration are averaged to obtain a single performance estimate. The 10-fold cross validation is used to ensure that every data point is used for both training and testing, reducing bias and variance in performance evaluation. Data is shuffled before splitting into folds to ensure representative distribution of data in each fold. The 10-fold cross-validation also helps to mitigate the risk of overfitting by ensuring the model is trained and tested on different subsets of the data, allowing for a more robust evaluation of its performance. By adopting these approaches, the proposed model’s prediction becomes accurate, reliable, and ready to help doctors detect sleep disorders more easily and effectively.

The performance metrics (Accuracy, Precision, Recall, F1 Score,) of the proposed PPG analysis ML model for sleep disorder classification are compared with that of the conventional (ECG/EEG/EOH/EMG) models, as shown in Table 3.

Table 3

Author Channel
(Signals) ML Model Results
Accuracy (%) Precision (%) Recall (%) F1 Score (%)
Sharma et al., 2021 ECG ML with KNN 97.87 98.15 96.65 97.39
Siddiqui et al., 2016 EEG PSD Features - - - -
Sharma et al., 2022 EOG + EMG ML WITH EBTC 95.60 96.21 - 96.45
Present Invention PPG OBWFB + Hjorth + ML EBDT 96.88 99.05 94.67 96.80
KNN 96.37 93.23 100.00 96.50
SVM 95.34 95.42 95.25 95.33

It is observed that the proposed PPG ML model demonstrates superior performance in terms of improved accuracy with a good balance between correctly identifying positive cases and minimizing false positives/negatives.

The foregoing descriptions of exemplary embodiments of the present invention have been presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed, and obviously many modifications and variations are possible in light of the above teaching. The exemplary embodiment was chosen and described in order to best explain the principles of the invention and its practical application, to thereby enable the persons skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. It is understood that various omissions, substitutions of equivalents are contemplated as circumstance may suggest or render expedient, but is intended to cover the application or implementation without departing from the scope of the claims of the present invention. , Claims:We claim:

1. A system for detecting sleep disorders, the system comprises:
a photoplethysmography (PPG) sensor (100) coupled to a wearable device (200) adapted to capture signals associated with physiological changes of a patient in sleeping stage; and
a user device (300) wirelessly connected with the wearable device (200) to receive the PPG signals therefrom, wherein the user device (300) has installed therein an application interface (400) configured to:
segment the PPG signals into defined time interval segments;
decompose the segmented signals into different frequency sub-bands through computation of approximation and detail coefficients followed by down sampling over five levels in an Optimal Biorthogonal Wavelet Filter Bank (OBWFB); wherein the approximation coefficients represent low-frequency components, and the details coefficients represent high-frequency components;
apply Hjorth parameters to extract meaningful features associated with various sleeping conditions from each of the frequency sub-bands;
train a machine learning classifier to classify the extracted features in a healthy class or a sleep disorder class selected from insomnia, sleep-disordered breathing (SDB), periodic limb movements (PLM), narcolepsy, rapid eye movement behaviour disorder, nocturnal frontal lobe epilepsy (NFLE); and
display the classification results in real-time.

2. The system as claimed in claim 1, wherein the OBWFB includes an analysis high-pass filter for computing the detail coefficients.

3. The system as claimed in claim 1, wherein the OBWFB includes an analysis low-pass filter for computing the approximation coefficients.

4. The system as claimed in claim 1, wherein the OBWFB performs down sampling of the approximation and detail coefficients by a factor of two to reduce temporal and frequency resolution.

5. The system as claimed in claim 1, wherein the Hjorth parameters involve computation of an activity value, a mobility value, and a complexity value for each of the of the frequency sub-bands.

6. The system as claimed in claim 1, wherein the machine learning classifier (408) is trained through any of Ensemble Bagged Decision Trees (EBDT), Support Vector Machines (SVM), and K-Nearest Neighbours architectures with ten-fold cross-validation.

7. The system as claimed in claim 1, wherein the machine learning classifier employs a training database hosted in a cloud server (500).

8. The system as claimed in claim 1, wherein the wearable device (200) is a smart watch or fitness band.

9. The system as claimed in claim 1, wherein the user device (300) is a smartphone.

10. A method for detecting sleep disorders, the method comprises steps of:
capturing (S1) signals associated with physiological changes of a patient in sleeping stage through a photoplethysmography (PPG) sensor (100) coupled to a wearable device (200);
transmitting (S2) the PPG signals wirelessly from the wearable device (200) to a user device (300) for their processing and analysis through an application interface (400) installed therein;
segmenting (S3) the PPG signals into defined time interval segments;
decomposing (S4) the segmented signals into different frequency sub-bands through computation of approximation and detail coefficients followed by down sampling over five levels in an Optimal Biorthogonal Wavelet Filter Bank (OBWFB); wherein the approximation coefficients represent low-frequency components, and the details coefficients represent high-frequency components;
applying (S5) Hjorth parameters to extract meaningful features associated with various sleeping conditions from each of the frequency sub-bands;
training (S6) a machine learning classifier to classify the extracted features in a healthy class or a sleep disorder class selected from insomnia, sleep-disordered breathing (SDB), periodic limb movements (PLM), narcolepsy, rapid eye movement behaviour disorder, nocturnal frontal lobe epilepsy (NFLE); and
displaying (S7) the classification results in real-time in the user device display.