System and method for predicting asthma exacerbations

Abstract

A breath analysis system for predicting asthma exacerbations. A breath analysis device has an array of sensor for detecting volatile organic compounds (VOCs), carbon dioxide (CO2), and other exhaled b

Description

FIELD OF THE INVENTION

[0001] The present invention relates to the field of breath analysis and respiratory health monitoring. More specifically, it relates to a method and system for detecting and analysing biomarkers in exhaled breath, utilising advanced sensor arrays, integrated CO2 monitoring, and machine learning algorithms to predict asthma exacerbations. The invention further incorporates data processing technologies, including Edge Al and deep neural networks (DNNs), to provide real-time, accurate, and reliable assessments of respiratory health, with a particular emphasis on the management and control of asthma. BACKGROUND TO THE INVENTION

[0002] Asthma, a chronic respiratory condition that affects around 300 million people worldwide, presents significant diagnostic and management challenges. Traditionally, asthma diagnosis has been heavily reliant on meticulous clinical history-taking by trained clinicians. However, this approach has limitations, frequently leading to misdiagnosis or delayed diagnosis, compromising patient care and outcomes.

[0003] Asthma affects a staggering 5.4 million individuals in the United Kingdom, imposing a substantial economic burden on the National Health Service (NHS) with annual costs exceeding £1.1 billion. This financial strain encompasses expenses related to hospital admissions, medication, and approximately 4 million visits to general practitioners. Despite these significant costs, individuals with asthma face challenges in monitoring their condition effectively a nd identifying exacerbations promptly. The difficulty lies in discerning when their lung function deteriorates, and inflammation escalates, leading to nearly 65,000 emergency asthma admissions annually, a substantial portion of which are preventable.

[0004] Severe asthma, constituting over 50% of total asthma expenditures (£1 billion / year), contributes significantly to the NHS's financial burden. Expensive medications and recurrent exacerbations are the primary drivers of these costs. Alarmingly, asthma-related deaths in the UK persist at over 1,200 annually, with premature mortality rates nearly twice the European average. Tragically, approximately 90% of these deaths could have been avoided, as highlighted by the UK National Review of Asthma Deaths (NRAD), which underscored inadequate recognition of asthma diagnosis and severity as a critical factor associated with mortality.

[0005] Asthma affects 1 in 11 children in the UK and is a common childhood chronic condition. Acute asthma attacks lead to unplanned hospital admissions, emergency hospital visits, and missed schooldays. Outcomes are worse for children and young people living in deprived areas. Detecting early signs of deterioration at home remains a challenge. Many children who have asthma attacks are identified as having “poor control” - lots of symptoms day to day. It is widely acknowledged that asthma is poorly managed due partly to the lack of a simple and effective personal monitor to help monitor it effectively and reliably.

[0006] There is a pressing need for an easy-to-use asthma test for children. Asthma is the most common chronic respiratory condition, affecting approximately 12 million children in the European Union. In many European healthcare settings, the diagnosis relies solely on clinical history and examination without additional tests. Recent reports from Europe and North America have highlighted a high rate of asthma misdiagnosis, including both over-diagnosis and under-diagnosis.

[0007] Misdiagnosis in children often occurs because respiratory symptoms are common in this age group and are frequently nonspecific, often representing episodes of viral respiratory tract infections. Some of these infections can be prolonged and present with clinical symptoms similar to asthma. Accurate diagnosis in children is crucial; over-diagnosis can lead to over-treatment with corticosteroids, which has implications for healthcare costs, risks of unnecessary side effects, and potential delays in identifying important alternative diagnoses. Conversely, underdiagnosis and under-treatment of asthma can result in unnecessary morbidity, poor quality of life, and increased mortality in low-resource settings. Additionally, existing lung function tests are often too complex for children to use effectively. Asthma is primarily a clinical diagnosis, so no objective diagnostic test is available. This underscores the need for a simple and accessible asthma test for children.

[0008] This stark reality underscores the urgent need for non-invasive, user-friendly diagnostic tools capable of accurately diagnosing and monitoring asthma and its exacerbations. Breath analysis emerges as a promising avenue, leveraging the detection of volatile organic compounds (VOCs) in human breath to create personalised signatures indicative of underlying conditions. The analysis of CO2 capnograms, focusing on specific parameters like the angle and plateau characteristics, emerges as a crucial diagnostic tool for asthma and its exacerbations. By scrutinising the angle and plateau tangent within capnograms, these metrics unveil distinct patterns indicative of asthma and exacerbations. For instance, healthy capnograms often exhibit square-shaped waveforms with smaller angles and plateau tangents during exhalation.

[0009] Conversely, asthma-related capnograms display shark-fin-shaped waveforms characterised by larger angles and plateau tangents. This discernible divergence aids in identifying and distinguishing asthma-related patterns, facilitating prompt diagnosis and prognosis. Moreover, utilising these CO2 capnogram parameters proves instrumental in monitoring exacerbations, as deviations from the norm, such as alterations in angle and plateau, signify exacerbation events. This focused analysis of CO2 capnograms, honing in on these specific geometric properties, thus stands as a pivotal method in diagnosing asthma with precision and timeliness. Non-invasive measures ofairway inflammation, with the temperature and humidity expected to be higher in asthmatics than controls and higher in more severe disease (where inflammation is related to heat and less humidity). Pressure and flow are also lower in asthmatics.

[0010] The invention’s Point of care (POC) breath test device, also referred to as Exhale-Dx, can be used to predict asthma exacerbations accurately and monitors asthma control. This easy-to-use device can simultaneously measure a panel of important biological and physiological parameters from exhaled breath and utilise machine learning algorithms. People with asthma underestimate the severity of their condition and overestimate how well it is controlled.

[0011] The presented device and machine learning algorithms can predict asthma attacks three days in advance and should, therefore, result in earlier detection of an impending asthma attack. This may enable patients to improve Self-Management (following a self-care protocol) or doctors to identify and treat airway inflammation early, which may prevent asthma exacerbations and improve asthma control. STATEMENTS OF THE INVENTION

[0012] According to a first aspect of the invention, there is provided a breath analysis system for predicting asthma exacerbations as set out in the appended claims 1 to 8.

[0013] According to a second aspect of the invention, there is provided a method of predicting asthma exacerbations as set out in the appended claims 9 to 16. BRIEF DESCRIPTION OF DRAWINGS

[0014] The invention will now be described with reference to the accompanying drawings in which:-

[0015] Figure 1 shows an Exhale-Dx device and mobile application, in accordance with the invention;

[0016] Figure 2 illustrates a Capnogram interpretation;

[0017] Figure 3 shows the algorithm detection of the end -tidal plateau by monitoring the change in the carbon dioxide concentration;

[0018] Figure 4 shows a deep neural network diagnosis model architecture;

[0019] Figure 5 shows a deep neural network self-management model architecture;

[0020] Figure 6: shows CO2 graphs displaying 5 tests done within 20 minutes time frame using the same device;

[0021] Figure 7 shows VOC graphs displaying 5 tests done within 20 minutes time frame using the same device;

[0022] Figure 8 is VOC graphs showing the difference between a non-asthmatic and an asthmatic;

[0023] Figure 9 is CO2 graphs showing the difference between non-asthmatic and asthmatic adults;

[0024] Figure 10 is VOC deconvolution graphs showing the 3 different components extracted for non-asthmatic (solid line), asthmatic (dotted line) for adults;

[0025] Figure 11 shows another example of VOC deconvolution graphs showing the 3 different components extracted for non-asthmatic (solid line), asthmatic (dotted line) for adults;

[0026] Figure 12 shows a Numerical Analysis to confirm the difference between the extracted components for asthma and non-asthmatic for adults;

[0027] Figure 13 shows Numerical Analysis to confirm the difference between the extracted components for asthma and non-asthmatic for adults;

[0028] Figure 14 shows CO2 graphs showing the difference between non asthmatic and asthmatic children;

[0029] Figure 15: VOC graphs showing the difference between a non-asthmatic and asthmatic children;

[0030] Figure 16 shows VOC deconvolution graphs showing the 3 different components extracted for nonasthma (solid line), asthmatic (dotted line) for children;

[0031] Figure 17 shows Principal Component analysis showing a clear difference between asthmatics (yellow) and non-asthmatics (purple);

[0032] Figure 18 shows Numerical Analysis to confirm the difference between the extracted components for asthma and non-asthmatic for children;

[0033] Figure 19 shows CO2 graphs showing the difference between a non-asthmatic (NA) and an asthmatic on medication (MM) and asthmatic not on medication adults (NM);

[0034] Figure 20 shows VOC graphs showing the difference between a non-asthmatic (NA) and an asthmatic on medication (MM) and asthmatic not on medication adults (NM);

[0035] Figure 21 shows Receiver operating characteristic (ROC) curve of diagnosing asthma for adults DNN;

[0036] Figure 22 shows VOC level for a year for asthmatic adults;

[0037] Figure 23 shows VOC level for a week in spring for an asthmatic adult;

[0038] Figure 24 shows VOC level for the week for an asthmatic and non-asthmatic child;

[0039] Figure 25 shows VOC level for the week for a non-asthmatic child;

[0040] Figure 29 shows the simplified flowchart of the method; [0041 ] Figure 30 shows the Flowchart of breath test;

[0042] Figure 30(a) shows the top half of the Flowchart of breath test;

[0043] Figure 30(b) shows the bottom half of the Flowchart of breath test;

[0044] Figure 31 shows the flow chart of standardization; and

[0045] Figure 32 shows the steps of the method. DETAILED DESCRIPTION OF THE INVENTION Referring to Figure 1, the invention provides a point-of-care (POC) breath analysis system comprising an apparatus and an associated mobile device application. The apparatus is configured to collect and analyse exhaled breath for the purpose of diagnosing and monitoring asthma and its exacerbations. The mobile device is communicatively coupled to the apparatus and is configured to process, display, and store the analysis results. The system enables a non-invasive and user-friendly method of assessment by detecting and interpreting volatile organic compounds (VOCs) and carbon dioxide (CO2) characteristics in exhaled breath. The system is capable of generating individualised breath profiles that may indicate underlying respiratory conditions. In particular, the system analyses CO2 capnograms to identify distinctive waveform features associated with asthmatic breathing patterns. The analysis focuses on specific capnogram parameters, including waveform angle and plateau characteristics, which have been shown to differ significantly between healthy individuals and those experiencing asthma or exacerbations. Healthy individuals typically produce square-shaped capnogram waveforms with smaller angles and relatively flat plateau regions. In contrast, individuals with asthma often produce so-called “shark-fin” shaped waveforms with larger angles and sloped plateau regions, reflective ofairway obstruction. These parameters provide a reliable diagnostic signature for identifying the presence of asthma and distinguishing between stable and exacerbated states. Furthermore, deviations in capnogram morphology, including changes in angle and plateau slope, serve as indicators of exacerbation events, enabling early detection and timely clinical response. In addition to capnographic analysis, the system measures various physical and physiological parameters from exhaled breath. This includes real-time sensing of temperature, humidity, pressure, and flow rate. Asthmatic individuals are generally observed to exhibit elevated exhaled breath temperature and humidity levels, along with reduced pressure and flow—particularly in more severe presentations. These metrics further support the comprehensive assessment of respiratory health status. The breath analysis device (Exhale-Dx™) incorporates machine learning algorithms to process the multiparametric data collected. These algorithms are capable of predicting asthma exacerbations up to three days in advance. Prediction outputs are communicated to the user via the mobile device interface, which presents results in a clear and intuitive format. This enables both patients and clinicians to engage in early intervention, through either personalised self-management protocols or timely clinical treatment. The system thus provides a significant advancement in respiratory health monitoring, offering a portable, intelligent, and predictive solution for asthma management. The invention, also referred to as Exhale-Dx™ leverages the advanced sensor array platform configured to detect multiple species and designed for seamless integration with mainstream semiconductor processes for high-volume manufacturing. These nanotechnology-based sensors offer high sensitivity and low power consumption, serving as chemical biosensors and gas sensors, which are advantageous for real-time monitoring applications. We developed an end-tidal collection method for gathering breath samples from the alveolar air to address the lack of standardised methods in VOC breath analysis. Our algorithms ensure reliable, repeatable, and reproducible measurements by controlling breath sample collection in real-time with fixed, controllable flow rates and monitoring physical parameters. We measure 13 different waveforms during exhalation, including multiple VOCs, CO2 parameters, peak exhaled flow rate, and additional physical factors like temperature and humidity. The Exhale-Dx™ device leverages Edge Al technology, allowing it to process data locally without needing an internet connection to deliver results. This on-device processing capability ensures that the process takes approximately 1 minute from when the patient logs into the application to the results display. By eliminating the need for constant internet connectivity, Exhale-Dx provides rapid, reliable, and efficient asthma diagnostics and self-management; Exhale-Dx™ ‘s results are produced in under 1 minute with a relaxed breathing rhythm to suit all patient types. The Exhale-Dx™ device was tested on adults aged 16 and above and children aged 6 to 16. These tests were conducted separately to account for physiological differences between the age groups and to ensure the device's accuracy and effectiveness across a diverse population. By validating the device on adults and children, Exhale-Dx™ demonstrates its capability to provide reliable asthma diagnostics and management tools for users of all ages. Standardisation of breath test One of the main technical challenges is that standardised methods are needed for collecting VOC breath samples. Concentrations of volatile compounds in blood are reflected by their concentrations in the exhaled air, and alveolar breath is the part of exhaled air in equilibrium with systemic blood. In contrast, end-tidal air is the last fraction of expired air, whose composition resembles alveolar air. This was achieved by integrating a carbon dioxide sensor inside the breath sampling tube and monitoring the level during exhalation. Referring now to Figure 2: • Phase 1: Beginning of exhalation (dead space ventilation). No carbon dioxide is present. • Phase 2: Mixture of carbon dioxide from the alveoli and the dead space air. • Phase 3: The carbon dioxide concentration remains constant, as mainly alveolar air is exhaled, known as the alveolar plateau. • Phase 4: Inhalation Referring to Figure 3, a software algorithm is implemented to identify the end-tidal fraction of exhaled breath and to activate chemical sensing only during this specific phase. The algorithm monitors the carbon dioxide (CO2) concentration during exhalation and detects the transition to the alveolar plateau, which represents the end-tidal portion. Activation of the chemical sensor array is synchronised with the detection of this plateau, ensuring that measurements are taken exclusively from alveolar breath. This approach enhances the reproducibility and accuracy of biomarker detection in exhaled breath samples. As the concentration of volatile organic compounds (VOCs) in exhaled air reflects their levels in the bloodstream, sampling alveolar breath—being in equilibrium with systemic blood—offers a physiologically relevant representation. The end-tidal breath, constituting the final fraction of exhalation, closely approximates alveolar air and is therefore critical for reliable diagnostic analysis. This method supports consistent breath sampling and may serve as a standardised approach for non-invasive collection of breath biomarkers. The algorithm enables the continued detection and quantification of target biomarkers under this sampling regime, without compromising diagnostic integrity. Edge Artificial Intelligence (Edge Al) An innovative component of the Exhale-Dx™ device, as shown in Figure 1, is the integration of Edge Artificial Intelligence (Edge Al) technology. This enables rapid and efficient analysis of breath samples through local on-device data processing, eliminating reliance on continuous internet connectivity. The Edge Al implementation ensures diagnostic results can be produced within approximately one minute of sample collection, significantly enhancing the responsiveness and usability of the system in both clinical and remote settings. Edge Al also maintains high diagnostic accuracy and consistency, both of which are essential for effective asthma management. The Al processing routines are embedded directly into the microcontroller housed within the Exhale-Dx™ apparatus. This embedded system architecture reduces latency and enhances system resilience, allowing the device to operate independently of cloud resources during active analysis. Within the Edge Al framework, the system employs two pre-trained deep neural networks (DNNs). One DNN is configured for asthma self-management classification, while the other is tailored for diagnostic prediction. These DNNs are not trained on the device but are deployed in a fixed, validated state, where they perform inference tasks based on sensor input data. This dual-network configuration enables the device to output both immediate diagnostic assessments and longer-term management recommendations with a high degree of reliability. In parallel with local analysis, the platform is supported by a cloud-based backend infrastructure powered by Amazon Web Services (AWS). This includes components such as API Gateway, Lambda, EC2, DynamoDB, and Cognito. The cloud infrastructure facilitates secure storage, remote access to patient data, and remote updates to the device software and algorithms. The AWS service region is configurable to meet local data governance and privacy requirements, including compliance with GDPR and equivalent standards. All transmitted patient data is secured through comprehensive encryption and identity protection protocols, ensuring that the system meets international standards for data security, scalability, and patient privacy. This hybrid architecture—combining on-device Edge Al and cloud-based data services— positions the Exhale-Dx™ system as a robust, intelligent solution for real-time respiratory health monitoring and personalised care delivery. Diagnosis Referring to Figure 4, principal component analysis (PCA) is used as an exploratory analysis to assist with determining hyperparameters in the selected deep learning (DL) algorithm. PCA captures the most significant variations in the data by transforming the original set of parameters into a smaller number of principal components. This dimensionality reduction enables improved identification of patterns and correlations among the parameters, which is essential for distinguishing between different asthma control levels and phenotypes. In addition, PCA highlights the most influential parameters, allowing the deep learning models to assign appropriate weights and improve predictive accuracy. As a result, PCA enhances the diagnostic process by clarifying how various biomarkers and lung function parameters contribute to asthma management, leading to more accurate and reliable outcomes. This approach allows identification of the most critical components, which are subsequently assigned higher weighting. A deep neural network (DNN) model is developed for the diagnosis of asthma. Each layer in the model is designed to focus on specific biomarkers or lung function parameters, with weightings determined based on the results from the PCA. For binary classification—distinguishing between asthmatic (1) and non-asthmatic (0) individuals—the model architecture incorporates several advanced neural network blocks. The model begins with an input layer, which receives raw data from a variety of biomarkers and lung function parameters. This is followed by convolutional layers that automatically extract and learn features from the input data, identifying patterns that are useful for asthma diagnosis. To enhance performance, the architecture integrates Dense Blocks, which connect each layer to every other layer in a feed-forward manner. This structure maximises feature reuse and improves gradient flow during training. In addition to Dense Blocks, the model includes Inception Modules, which enable multi-level feature extraction by applying convolution filters of different sizes within the same module. This allows the model to capture diverse features at multiple scales, thereby improving its ability to differentiate between asthmatic and non-asthmatic profiles. Batch Normalisation layers are incorporated to normalise outputs from previous layers, thereby stabilising and accelerating the training process by reducing internal covariate shifts. The architecture also includes fully connected layers, which aggregate the features extracted by the dense and specialised blocks to produce the final classification. Dropout layers are inserted at strategic points in the network to prevent overfitting, by randomly setting a fraction of the input units to zero during training. The output layer uses a sigmoid activation function, generating a probability score ranging from 0 to 1 to indicate whether a patient is asthmatic. The DNN model is trained using a comprehensive dataset that includes patient records, historical data, and clinician inputs. An Adam optimiser is used to dynamically adjust the learning rate for improved convergence. By integrating the advanced neural network blocks and optimisation techniques, the model is able to learn complex patterns in the input data and provide accurate and reliable asthma diagnoses. To minimise the error between predicted outcomes and actual data, a robust loss function is employed during training. Specifically, binary cross-entropy loss is used, which is well-suited for classification problems involving two classes. The loss function measures the difference between predicted probability values and actual class labels. It penalises incorrect classifications more heavily, thereby helping the model to more effectively adjust weights and biases throughout training. This supports the production of outputs that closely match the correct classification targets. N iy,                i ^Loss / Yyactualj ' log ( Vpredictedi) J i=l In which N is the amount of data points, yactuail is the actual data and ypredictedjs the predicted data. A cohort of 40 patients was utilised fortraining and validating the deep neural network (DNN) model. The group consisted of 25 male and 15 female participants, ranging in age from 17 to 85 years. Of these, 20 individuals had been previously diagnosed with asthma by general practitioners and were undergoing pharmacological treatment for asthma management. The remaining 20 participants exhibited no known pulmonary conditions and served as the non-asthmatic control group. The dataset was randomly partitioned into three segments for model development: • 60% allocated for training, • 30% for validation, and • 10% reserved for testing. During training, the DNN model was repeatedly exposed to the training data across multiple epochs. Each epoch involved: • A forward pass, during which the input data was processed to generate predictions; and • A backward pass, during which the model computed errors and updated internal weights accordingly. Through this iterative learning process, the model identified complex patterns and relationships among the input features. The training data encompassed both the full patient cohort and a broad range of predefined parameter thresholds sourced from previous research studies and clinician-defined boundaries. Specific parameters considered to be diagnostically significant were assigned greater weight within the algorithm to enhance model accuracy and clinical relevance. To improve generalisability and reduce the risk of overfitting, the model was also subjected to cross-validation techniques during the initial development phase. Although the dataset was relatively small, the application of cross-validation served to validate the model’s performance beyond its training subset. The dataset was segmented into multiple folds, each of which was used to train and validate the model iteratively, allowing evaluation on unseen data in each cycle. This five-fold cross-validation approach enabled optimisation of hyperparameters, assessment of model robustness, and early identification of any overfitting tendencies. Furthermore, it provided insights into the model’s limitations under constrained data conditions and informed the strategy for future data acquisition. The process was designed to ensure that the model maintained predictive accuracy and reliability across a broader range of real-world scenarios. Self-Management Principal Component Analysis (PCA) is employed as an exploratory tool to optimise hyperparameters in the selected deep learning (DL) algorithm. The objective of PCA is to identify the most significant components in the dataset, which are subsequently assigned higher weighting during model training. PCA identifies the dominant patterns and trends across various parameters, including volatile organic compound (VOC) levels, carbon dioxide (CO2) waveforms, and other relevant biomarkers, by transforming the original data into principal components. This dimensionality reduction enables more efficient and accurate real-time tracking of asthma control and the risk of exacerbations. By focusing computational resources on the most relevant features, PCA reduces the data processing load and improves system response time. This streamlined data pipeline ensures that the most critical information is prioritised for remote monitoring. In doing so, PCA facilitates more timely and precise patient feedback, empowering individuals to manage their asthma more proactively and enabling healthcare providers to deliver data-driven, remote clinical guidance. The emphasis on key indicators supports more effective asthma control across various settings. Referring to Figure 5, the system includes a 13-layer Deep Neural Network (DNN) architecture specifically configured for multiclass classification to predict three distinct asthma control states: well-controlled, poorly controlled, and uncontrolled asthma. The architecture integrates multiple advanced neural network components to enhance predictive performance and system reliability. The model incorporates Residual Blocks, which are designed to address the vanishing gradient problem commonly observed in deep learning systems. These blocks introduce shortcut connections that preserve gradient flow throughout the network, thereby enabling the construction of deeper architectures without performance degradation. This structural enhancement allows the network to learn more complex dependencies within the input data, resulting in more accurate and refined output predictions. In addition, Attention Mechanisms are embedded within the DNN architecture. These mechanisms dynamically assign weights to various input features, allowing the model to focus on the most salient aspects of the data for differentiating between levels of asthma control. By improving the interpretability and relevance of input features, the attention layers contribute to more precise and clinically meaningful predictions. The DNN model is trained on a large dataset consisting of patient records, previously defined parameter thresholds from existing research, and inputs provided by clinicians. This comprehensive training set ensures that the model can generalise effectively across different patient populations and clinical scenarios. The architecture also supports adaptive learning, enabling the system to continually refine its internal representations and maintain accuracy as new data becomes available. The combination of Residual Blocks, Attention Mechanisms, and a robust training dataset allows the 13-layer DNN to perform accurate and reliable asthma self-management classification. This integration enhances the network's depth and robustness, while enabling personalised, real-time feedback critical for effective asthma control and patient engagement. To further optimise performance in the self-management context, an alternative loss function is applied. Specifically, the Mean Squared Error (MSE) loss function is utilised to measure and minimise the deviation between predicted and actual asthma control outcomes. MSE assigns higher penalties to larger errors, which is especially important when small deviations can have significant clinical implications in asthma management. The use of MSE incentivises the model to generate more accurate predictions, thus supporting improved self-care decisions by the user and more effective long-term disease control. n i N actual ypredicted-^) --- + ~ ’ (yactualp Vpredicted i = l i=l In which, T is the amount of days to take into account that is set to 5 right now. N is the amount of data points, yactua^ is the actual data for the first day and ypredicted^5 the predicted data for the first day. A cohort of 20 patients was used for the training and validation of the self-management deep neural network (DNN) model. The group comprised 11 male and 9 female participants, aged between 17 and 85 years. Of these, 10 participants were asthmatic individuals who had been diagnosed by their general practitioners. These patients represented a range of asthma severities and were all undergoing treatment with medication at the time of the study. The remaining 10 individuals had no known lung conditions and served as a non-asthmatic control group. The training and validation process for the self-management DNN followed a structured and comprehensive methodology. The dataset was randomly split, with 60% allocated for training, 30% for validation, and 10% reserved for testing. The model was trained over multiple epochs. Each epoch included a forward pass, during which input data was processed to generate predictions, and a backward pass, during which the prediction errors were propagated back through the network to adjust internal weights. This iterative training procedure allowed the model to identify and learn complex relationships and patterns within the data, improving predictive performance overtime. To further assess the robustness and generalisability of the model, a 10-fold cross-validation technique was employed. The dataset was divided into 10 equal subsets. In each fold, nine subsets were used for training and one subset for validation, with each subset serving as the validation set once. This method enabled evaluation of the model across various data splits, allowing for thorough testing of its ability to generalise to unseen data. In addition to performance evaluation, cross-validation facilitated the optimisation of key hyperparameters such as learning rate, batch size, and the number of layers and neurons. The self-management DNN was trained using a large and diverse dataset that included thousands of data points. These data points encompassed patient-specific information, breath biomarkers, and lung function measurements collected over a two-year monitoring period. This extensive and heterogeneous dataset enabled the model to capture a wide range of behavioural and physiological variations associated with different asthma control levels. The rigorous training and validation approach ensured the model's reliability and accuracy across varying patient profiles. As a result, the self-management DNN provided consistent and personalised predictions, supporting improved asthma control and management. This enhanced the utility of the system in delivering actionable recommendations, thereby improving patient outcomes and overall quality of life. Results Multiple pre-clinical trials were conducted involving both adult and paediatric participants to evaluate the efficacy and reliability of the Exhale-Dx™ platform. These trials included individuals diagnosed with asthma as well as non-asthmatic individuals with no known respiratory conditions, enabling a comprehensive assessment of the diagnostic capabilities of the system across a broad user base. In addition, blind testing was performed with adult participants to further validate the performance of the device under controlled conditions. The trials also demonstrated the platform’s ability to efficiently support remote monitoring of patients, which is essential for decentralised and real-time asthma management. Preliminary testing was also initiated with paediatric participants. Early results from these assessments indicate promising applicability of the system for use in younger populations. The testing undertaken across diverse demographic groups underscores the robustness and versatility of the Exhale-Dx™ platform, highlighting its potential to significantly advance the diagnosis and management of asthma across all age ranges. Reproducibility The standardised breath test was comprehensively evaluated through a series of repeated measurements administered to a group of patients. The study consisted of five separate tests performed within a 20-minute timeframe. The results were analysed to assess the consistency and accuracy of the test methodology. Figures 6 and 7 present the corresponding carbon dioxide (CO2) and volatile organic compound (VOC) graphs, which display the averaged data obtained from these patients. These graphical representations provide a detailed overview of the observed variations and trends in CO2 and VOC levels across the repeated tests. Figure 6: shows CO2 graphs displaying 5 tests done within 20 minutes time frame using the same device. Figure 7 shows VOC graphs displaying 5 tests done within 20 minutes time frame using the same device. The results demonstrate consistent and stable performance of the testing procedure. CO2 levels remained within a ±5% range, while VOC levels exhibited variation within a range of ±10% to ±15%. These findings confirm the reliability and reproducibility of the breath test methodology under repeated conditions, supporting its suitability for routine clinical and remote monitoring applications. Diagnosis CO2 and VOC Graphs Adults The analysis of carbon dioxide (CO2) and volatile organic compound (VOC) data enables clear differentiation between asthmatic and non-asthmatic adult subjects, based on a set of defined physiological metrics. In particular, the examination of CO2 alpha values, which represent the rate of change in CO2 concentration during exhalation, in conjunction with the CO2 plateau width, reveals characteristic waveform patterns indicative of asthma. These variations are attributed to differences in airway resistance and respiratory mechanics commonly observed in individuals with asthma. In addition to CO2-based measurements, the VOC graphs yield further discriminatory features. Specifically, the VOC plateau width and VOC maximum value serve as critical parameters for identifying differences in volatile compound emissions between asthmatic and non-asthmatic groups. Figure 8 is VOC graphs showing the difference between a non-asthmatic and an asthmatic. Figure 9 is CO2 graphs showing the difference between non-asthmatic and asthmatic adults Figures 8 and 9 illustrate these findings. Figure 9 shows the distinction in CO2 alpha values and plateau width between the two groups, while Figure 8 highlights the variation in VOC plateau width and maximum value. These results support the effectiveness of the breath test in distinguishing asthmatic from non-asthmatic profiles, confirming the diagnostic value of the underlying measurement methodology. Deconvolution Adults Further analysis was undertaken to distinguish between asthmatic and non-asthmatic patients using a deconvolution technique applied to the volatile organic compound (VOC) data. This analysis utilised a Gaussian Mixture Model (GMM) to cluster the VOC data into three distinct Gaussian components, with the objective of identifying underlying patterns that may represent physiologically distinct states within the patient groups. Initially, the gmmClustering function was applied to the VOC data collected from both asthmatic and nonasthmatic subjects, resulting in three discrete clusters. These clusters correspond to different breath emission patterns and enable a component-level breakdown of the composite VOC signal. Following clustering, the optimizeGmmParameter function was used to refine the Gaussian parameters of each cluster. This function adjusts the fit to more accurately reflect the observed VOC signal using Gaussian curve modelling. The optimisation procedure was performed separately for asthmatic and non-asthmatic datasets to capture the distinct emission characteristics associated with each group. The resulting VOC profiles were then plotted to visually compare the distributions between the two populations. As shown in Figures 10 and 11, the digitised VOC data are overlaid with fitted Gaussian curves. Solid lines correspond to non-asthmatic individuals, while dashed lines represent asthmatic profiles. Each of the three Gaussian components is visually distinguished using different colours, thereby highlighting the differences in VOC emission patterns across the two cohorts. To quantify these observed differences, several key parameters were extracted from the Gaussian fits, including amplitude (peak height), mean (central value), standard deviation (distribution spread), and Full Width at Half Maximum (FWHM). These metrics were calculated for each component and compared across the asthmatic and non-asthmatic groups to provide a detailed assessment of VOC profile variation. In particular, the FWHM parameter offered insight into the width and dispersion of the VOC peaks, reflecting the variability and complexity of breath emission characteristics. The comparative analysis of these Gaussian parameters, combined with visual inspection of the plotted curves, enabled effective differentiation between asthmatic and non-asthmatic subjects. The outcomes of this analysis are illustrated in Figures 12 and 13, supporting the utility of VOC deconvolution via GMM as a valuable method for respiratory condition classification. CO2 and VOC Graphs Children The analysis of carbon dioxide (CO2) and volatile organic compound (VOC) data enables clear differentiation between asthmatic and non-asthmatic paediatric subjects using a defined set of physiological metrics. Specifically, the CO2 alpha values, which represent the rate of change in CO2 levels during exhalation, along with the CO2 plateau width, are used to identify waveform patterns characteristic of asthma. These patterns are indicative of differences in airway resistance and respiratory effort and are consistently observable in both adult and paediatric populations. In addition to CO2-based metrics, the VOC profiles provide further diagnostic value. Key parameters such as VOC plateau width and VOC maximum value offer critical insights into the emission of volatile compounds in exhaled breath. These parameters reflect physiological differences between asthmatic and non-asthmatic individuals, independent of age. Figures 14 and 15 illustrate these differences. Figure 14 presents the variation in CO2 alpha and plateau width between asthmatic and non-asthmatic children, while Figure 15 displays the corresponding differences in VOC plateau width and maximum value. The data presented in these figures support the effectiveness of the breath test methodology in accurately distinguishing between asthmatic and nonasthmatic paediatric profiles. Deconvolution Children To distinguish between asthmatic and non-asthmatic children, an advanced deconvolution analysis was applied to the volatile organic compound (VOC) data. A Gaussian Mixture Model (GMM) was employed to cluster the VOC signals into three distinct Gaussian components. This method was designed to identify underlying patterns in the VOC emission profiles that may reflect distinct physiological states within the paediatric population. The gmmClustering function was applied to VOC data from both asthmatic and non-asthmatic groups, producing three clusters corresponding to different breath emission patterns. Following this, the optimizeGmmParameter function was used to refine the Gaussian parameters for each component. This function adjusts the Gaussian fits to accurately represent the measured VOC data. The optimisation was performed independently for asthmatic and non-asthmatic datasets to capture the unique characteristics of each group. The resulting VOC profiles were plotted to facilitate visual comparison between the two groups. As shown in Figure 16, the digitised VOC data are overlaid with fitted Gaussian curves. Solid lines represent data from non-asthmatic children, while dashed lines represent data from asthmatic children. Each Gaussian component is assigned a distinct colour, illustrating the differences in VOC emission patterns between the groups. To quantify these differences, key Gaussian parameters were extracted from the fitted curves. These include amplitude (peak height), mean (central position), standard deviation (spread), and Full Width at Half Maximum (FWHM). These metrics were calculated for each component and compared between the two groups to provide a comprehensive characterisation of VOC profile variability. The FWHM, in particular, offers insight into the breadth of VOC peaks, which reflects underlying variability in breath composition. It is noted that when applying this analytical method to a paediatric population, the absolute values and threshold parameters may require adjustment. This is due to physiological and metabolic differences between children and adults. Such calibration ensures that the method remains accurate and reliable in differentiating between asthmatic and non-asthmatic conditions in paediatric subjects. The quantified results of this analysis are illustrated in Figure 18. PCAAdults Principal Component Analysis (PCA) was conducted to evaluate the discriminative power of various input parameters in differentiating between asthmatic and non-asthmatic adults. The results, shown in Figure 17, indicate that certain parameters do not produce a clear separation between the two groups. In the PCA plot, asthmatic individuals are represented in yellow, and non-asthmatic individuals in purple. While some overlap is observed, the PCA results suggest the potential for identifying distinct asthma phenotypes based on latent feature clusters. In the PCA graph, the x-axis corresponds to distinct carbon dioxide (CO2) waveform traits, while the y-axis represents different volatile organic compound (VOC) characteristics. The distribution of data points reveals the formation of discernible clusters within the asthmatic group, supporting the hypothesis that VOC-related features contribute significantly to the stratification and classification of asthma endotypes. This outcome reinforces the rationale for incorporating the full range of available parameters—both CO2 and VOC—into the deep learning (DL) model. The use of a comprehensive parameter set increases the model’s dimensional coverage and predictive capability when distinguishing between asthmatic and nonasthmatic profiles. ROC Curve Adults The Area Under the Curve (AUC) is a metric derived from the Receiver Operating Characteristic (ROC) curve, which plots the true positive rate against the false positive rate across varying classification thresholds. The ROC curve for asthma diagnosis using the described system is shown in Figure 21. An AUC value of 0.90 was obtained, indicating a high level of discriminative performance of the diagnostic model. The corresponding classification accuracy achieved was 93%. These results demonstrate the model’s effectiveness in accurately identifying asthmatic versus non-asthmatic individuals. Future optimisation efforts include expanding the patient cohort to further improve predictive performance. Specifically, the objective is to increase diagnostic accuracy to 95% or higher, while maintaining or improving the current AUC value, thereby preserving the model’s robustness and generalisability as the dataset scales. Self-Management CO2 and VOC graphs Adults The analysis of carbon dioxide (CO2) and volatile organic compound (VOC) data not only enables differentiation between asthmatic and non-asthmatic individuals but also reveals significant distinctions between asthmatic patients undergoing pharmacological treatment and those who are not. Examination of the CO2 waveform data shows variations in CO2 alpha values and plateau widths, which reflect the physiological impact of medication on respiratory function. Asthmatic patients receiving treatment typically exhibit more stable CO2 profiles and reduced plateau widths in comparison to untreated individuals. These patterns are indicative of improved airflow regulation and decreased airway resistance. Such differences are clearly illustrated in the CO2 graphs presented in Figures 19 and 20, which demonstrate how medication affects the rate of CO2 concentration change and the duration of elevated CO2 levels during exhalation. Similarly, the VOC data provide insight into how pharmacological intervention influences the composition of exhaled breath. Notable differences are observed in VOC plateau width and VOC maximum value between medicated and non-medicated asthmatic patients. Patients on medication generally exhibit narrower VOC plateau widths and lower maximum values, which reflect reduced airway inflammation and altered emission of specific volatile compounds. These trends are illustrated in the VOC graphs shown in Figures 19 and 20, which visually capture the modulation of VOC profiles due to treatment. The clear contrast between the two groups supports the utility of the breath test system in assessing treatment efficacy and guiding asthma management strategies. Ongoing monitoring of VOC profiles further allows forevaluation of medication effectiveness and patient adherence to prescribed therapies. Changes in VOC emission patterns may signal how well a treatment is controlling airway inflammation, with certain compounds increasing or decreasing in response to therapeutic efficacy. Additionally, consistent tracking of VOC trends may help identify non-adherence by revealing deviations from expected biomarker profiles. This information provides clinicians with valuable data to adjust treatment plans, enhance patient engagement, and improve clinical outcomes. Monitoring graphs Adults A cohort of 20 patients was monitored over a period ranging from three months to two years to evaluate changes in respiratory health over time. The cohort included ten asthmatic individuals presenting with varying levels of disease severity, as well as ten non-asthmatic individuals with no known pulmonary conditions, who served as control subjects. The monitoring protocol involved regular assessments combining volatile organic compound (VOC) analysis with other pulmonary function parameters. By tracking these biomarkers consistently throughout the observation period, the system was used to gain insight into each participant’s respiratory status and to detect signs of disease progression or stability. The longitudinal nature of the monitoring enabled the identification of subtle and progressive deterioration in respiratory health, particularly in asthmatic patients. The VOC profiles captured changes in airway inflammation and other metabolic indicators. Distinct patterns in VOC emissions were observed in association with asthma exacerbations and worsening symptom control. In addition to VOC measurements, supplementary lung function tests were incorporated, including the analysis of carbon dioxide (CO2) parameters and peak expiratory flow rate (PEFR). These combined measures provided a comprehensive evaluation of respiratory function and supported the early detection of physiological changes indicative of clinical deterioration. The data collected through this extended monitoring protocol demonstrated the system’s effectiveness in tracking the dynamic and evolving nature of respiratory health. The corresponding VOC and CO2 profiles obtained during this study are shown in Figures 22 and 23, highlighting the diagnostic and prognostic value of continuous breath-based monitoring. Monitoring graphs Children A two-week monitoring study was conducted on a cohort of ten paediatric patients to evaluate the progression of respiratory health over a short but clinically relevant period. The cohort included five children diagnosed with asthma and five children without any known respiratory conditions. The study focused on analysing volatile organic compounds (VOCs) in exhaled breath, along with other key pulmonary parameters, to identify early indicators of respiratory deterioration. Regular assessments were carried out throughout the monitoring period to capture dynamic changes in each child’s respiratory profile. The combined data provided a comprehensive overview of respiratory health trends in the paediatric subjects under observation. The analysis yielded significant findings related to the onset of respiratory decline in some participants. Notably, alterations in the VOC profiles served as early indicators of airway inflammation and other underlying respiratory abnormalities. These variations often preceded the emergence of clinical symptoms, supporting the utility of VOC analysis as a non-invasive early warning system. In addition to VOC data, several other physiological parameters were monitored, including respiratory rate, oxygen saturation, and lung function metrics such as carbon dioxide (CO2) parameters and peak expiratory flow rate (PEFR). These complementary measurements validated the observed VOC changes and contributed to a more holistic assessment of each patient’s respiratory condition. The ability to detect subtle yet clinically meaningful changes within a two-week timeframe underscores the potential of this breath analysis platform for early intervention and precise management of respiratory diseases in paediatric populations. The results of this study are illustrated in Figures 24 and 25, demonstrating the responsiveness and effectiveness of the monitoring system in children. PCAAdults The analysis demonstrated that certain parameters did not distinctly differentiate between individuals experiencing deterioration in respiratory health, as shown in Figure 27. However, the data suggest the potential for developing a classification system based on a traffic light indication scheme, wherein green corresponds to well-controlled asthma, amber to poorly controlled asthma, and red to uncontrolled asthma. In the Principal Component Analysis (PCA) graphs, the x-axis represents distinct carbon dioxide (CO2) waveform traits, while the y-axis delineates various volatile organic compound (VOC) characteristics. The graphical legend applies colour-coding aligned with the traffic light scheme, allowing visual identification of asthma control levels within the asthmatic population. The presence of discernible clusters within the asthmatic group indicates that VOC parameters significantly contribute to the monitoring and classification of asthma attacks. As illustrated in Figure 27, these clusters within the labelled class subsets highlight the utility of VOC-based features in enhancing the stratification of asthma subtypes or endotypes. This refined classification capability enables a more nuanced understanding of asthma's heterogeneity and supports improvements in the predictive accuracy of the deep learning (DL) model. Such an approach enhances the model’s capacity to forecast asthma exacerbations and contributes directly to personalised self-management and the timely implementation of clinical interventions. ROC Curve Adults The Area Under the Curve (AUC) is a performance metric derived from the Receiver Operating Characteristic (ROC) curve, which plots the true positive rate against the false positive rate across varying classification thresholds. The ROC curve for the self-management model of asthma in adults is shown in Figure 26. An AUC value of 0.90 was achieved, indicating a high level of model discrimination between different asthma control states. The corresponding classification accuracy was measured at 91%, confirming the reliability of the deep learning model in predicting asthma control levels for adult patients. Planned future steps involve increasing the number of patients in the training dataset to validate the model across a broader population. The objective is to maintain the same level of accuracy while expanding the prediction window to five days or more, without reducing the AUC. This will enhance the model's clinical utility for early intervention and self-management support. ROC Curve Children The Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) is used to evaluate the overall performance of a diagnostic or predictive model, with the curve itself representing the relationship between the true positive rate and the false positive rate at various threshold settings. The ROC curve for the paediatric asthma self-management model is shown in Figure 28. Analysis of the model’s performance yielded an AUC value of 0.64, with a corresponding diagnostic accuracy of 71%. These results indicate a moderate level of discriminative capability in predicting asthma control levels in children. To improve the robustness and generalisability of these findings, future development efforts will focus on increasing the sample size of paediatric participants. Additionally, the goal is to extend the prediction window to five days in advance, while maintaining or improving both the AUC and overall diagnostic accuracy. This would enhance the system's effectiveness for early intervention and proactive paediatric asthma management. In conclusion, the invention presents a method to predict asthma exacerbations. Platform The system, referred to as Exhale-Dx™, incorporates an advanced sensor array platform engineered for the detection of multiple chemical and biological species. The platform is compatible with standard semiconductor manufacturing processes, enabling scalable, high-volume production. The integrated sensors are based on nanotechnology, providing high sensitivity and low power consumption. These characteristics render the sensors suitable for use as chemical, biosensors, and gas sensors, and are particularly advantageous for continuous, real-time monitoring applications. Standardised Breath Test To address the lack of standardisation in volatile organic compound (VOC) breath analysis, Exhale-Dx™ employs an end-tidal breath collection method. This technique targets alveolar air, the portion of exhaled breath in equilibrium with systemic blood, ensuring physiologically relevant sample acquisition. A software algorithm is used to detect the end-tidal breath fraction in real-time and to activate chemical sensing only when this fraction is present. This approach ensures reproducibility in the detection of exhaled breath biomarkers and provides a framework for establishing a standardised method for breath sample collection. Pre-Processing (Standardisation and PCA) During exhalation, the system records thirteen waveform parameters, including VOC levels, carbon dioxide (CO2) metrics, peak exhaled flow rate, and physical properties such as temperature and humidity. These data are processed using proprietary algorithms designed to maintain a constant, controllable flow rate and to monitor physical conditions, thereby ensuring accurate and repeatable measurements. Preprocessing includes the application of Principal Component Analysis (PCA), which is used to optimise hyperparameters for the deep learning algorithm. PCA identifies the most significant features within the data, allowing higher weighting to be assigned to the most informative components. This enhances realtime tracking of asthma control levels and detection of exacerbation risks. Edge Al and Forecasting The system integrates Edge Al capabilities, enabling local data processing directly on the device without the need for an internet connection. This allows the platform to deliver diagnostic results typically within one minute from patient login to result display. The self-management functionality is driven by a 13-layer deep neural network (DNN) designed to perform multiclass classification of asthma control outcomes. The DNN architecture incorporates residual blocks, which mitigate the vanishing gradient problem and improve the model's capacity to train deeper structures. Attention mechanisms are also embedded in the network to enhance the model’s ability to prioritise relevant input features, thereby improving prediction accuracy. The model is trained on a large, diverse dataset consisting of patient records and breath biomarkers, which supports its generalisability across various clinical scenarios and populations. Figure 29 illustrates a simplified flowchart of the overall method. Figure 30 presents the specific procedural flow of the breath test operation. Device Operational Workflow The following describes the operation of a breath analysis device incorporating sensor calibration, BLE communication, and an algorithmic breath test process. 1. Power-Up and Initialisation • Upon powering up, the device activates a light indicator and sets it to orange, indicating the system is booting. • The pump and heater are switched on to bring the system to operational temperature. • A sensor calibration routine is initiated. • The system checks whether the sensors are ready: o If sensors fail to initialise: ■ The pump and heater are turned off. ■ The indicator light changes to red, indicating a fault. o If sensors are ready: ■ The indicator light is changed to yellow. ■ The solenoid valve is toggled on and off to verify function. 2. BLE Activation • The system begins Bluetooth Low Energy (BLE) advertising, enabling communication with external devices such as smartphones or tablets. 3. Algorithm Case Selection • A decision is made based on the operational mode: o If in Training Mode: ■ The system enters a continuous measurement loop to collect training data. o If in System Mode, the breath test algorithm is activated. Breath Test Algorithm (System Mode) The breath test algorithm consists of multiple internal states and transitions based on flow readings and CO2 plateau detection. State: IDLE • The system continuously reads flow measurements. • If flow values do not exceed a defined threshold, the system remains in IDLE. • If flow readings exceed the threshold, the system transitions to START. State: START • The system continues to read sensor data. • Once a CO2 plateau is detected in the waveform, the system transitions to the “Start Plateau Reading” state. State: START PLATEAU READING • The solenoid valve is turned on. • The system begins sample collection, acquiring 20 breath samples. State: END PLATEAU READING • The solenoid valve is turned off. • The system collects a further 40 samples during the end-plateau phase. State: SUBMIT RESULTS • The collected measurements are stored or transmitted. • All system parameters are reset to prepare for the next test cycle. STEPS: 1. System Configuration: A breath analysis device incorporating an advanced sensor array platform to detect multiple species, including VOCs, CO2, and other biomarkers. A dedicated mobile application is used during breath testing to control the breath sample collection and show the results. 2. The sensor array is configured for seamless integration with mainstream semiconductor manufacturing processes, allowing for high-volume production. 3. The sensor array exhibits high sensitivity and low power consumption, enabling its use in real-time monitoring applications such as chemical sensors, biosensors, and gas sensors. 4. The user starts the mobile application. 5. The patient breathes into the breath analysis system (Exhale-Dx™ Device) via a viral bacterial filter. 6. Control of Flow Rates: The system ensures reliable, repeatable, and reproducible measurements by controlling the breath sample collection in real time, maintaining fixed, controllable flow rates, and monitoring physical parameters. 7. End-Tidal Collection Method: A standardised method for collecting VOC breath samples using an end-tidal collection technique, capturing breath from the alveolar air portion of exhaled breath. 8. CO2 Sensor Integration: A CO2 sensor embedded within the breath sampling tube to monitor CO2 levels during exhalation, ensuring accurate identification of the end-tidal breath fraction. 9. Algorithm for End-Tidal Detection: A software algorithm that identifies the end-tidal breath fraction by monitoring changes in CO2 concentration and activates chemical sensing only when the end-tidal fraction is detected. 10. Multivariate Measurement: The system measures 13 distinct waveforms during exhalation, including multiple VOCs, CO2 parameters, peak exhaled flow rate, and additional physical factors such as temperature and humidity. 11. OPTIONALY: PCA for Data Simplification: Principal Component Analysis (PCA) is employed to identify the most significant components in the data, reduce dimensionality, and assign higher weighting to the most critical factors. 12. Edge Al Integration: The device incorporates Edge Al technology, enabling local data processing without requiring an internet connection. 13. Multiclass Classification DNN: A 13-layer DNN designed to predict asthma control outcomes, with classes including well-controlled asthma, poorly controlled asthma, and uncontrolled asthma. 14. Residual Blocks and Attention Mechanisms: The DNN architecture incorporates Residual Blocks to enhance gradient flow and Attention Mechanisms to prioritise critical input features, improving the model's prediction accuracy. 15. MSE Loss Function: The DNN model for self-management uses a Mean Squared Error (MSE) loss function to minimise prediction errors, particularly emphasising minor deviations critical for asthma management. 16. The results are shown on the mobile application and saved in a secure online data server for further review. This is a traffic light indication system output for the three days ahead. Green is well-controlled, amber is poorly controlled, and red is uncontrolled. 17. OPTIONALY: Validation Process: The DNN undergoes a rigorous validation process, including 10-fold cross-validation, to ensure robust performance and generalisation. Therefore, the invention provides an embodiment of a breath analysis system comprising a sensor-integrated diagnostic platform, operable according to the following functional sequence: System Configuration: Providing a breath analysis system comprising an advanced sensor array platform configured to detect multiple analyte species, including volatile organic compounds (VOCs), carbon dioxide (CO2), and additional biomarkers, the system being operable in conjunction with a dedicated mobile application. Sensor Array Integration: Configuring the sensor array for compatibility with standard semiconductor manufacturing processes to facilitate high-volume fabrication. Sensor Array Operation: Operating the sensor array with high sensitivity and low power consumption, enabling its use in real-time monitoring applications as chemical, biosensors, or gas sensors. Application Initiation: Initiating the mobile application by the user to control the breath sampling process and enable data display. Breath Sample Introduction: Instructing the patient to exhale into the Exhale-Dx™ breath analysis device via a viral bacterial filter. Flow Rate Control: Controlling the breath sample collection in real-time to ensure accurate, repeatable, and reproducible measurements, wherein the system maintains constant and controllable flow rates and monitors physical sampling conditions. End-Tidal Sample Collection: Collecting VOC samples using a standardised end-tidal breath collection method, capturing air from the alveolar portion of the exhaled breath. CO2 Monitoring: Measuring exhaled CO2 levels via a CO2 sensor integrated within the breath sampling tube. End-Tidal Detection Algorithm: Executing a software algorithm configured to detect the end-tidal breath fraction by monitoring the change in CO2 concentration and triggering chemical sensing only upon detection of the end-tidal phase. Multivariate Signal Measurement: Measuring a plurality of waveform signals during exhalation, including at least 13 parameters such as multiple VOC levels, CO2 waveform characteristics, peak exhaled flow rate, and physical parameters including temperature and humidity. [Optional] PCA for Dimensionality Reduction: Applying Principal Component Analysis (PCA) to identify and weight the most significant components among the multivariate dataset, reducing dimensionality and enhancing algorithm performance. Edge Al Processing: Performing data analysis using Edge Al technology, wherein processing is carried out locally on-device, independent of internet connectivity. Asthma Control Prediction via DNN: Executing a deep neural network (DNN) model having 13 layers to classify asthma control states, the classification comprising at least three classes: well-controlled asthma, poorly controlled asthma, and uncontrolled asthma. DNN Architecture Optimisation: Implementing residual blocks within the DNN to preserve gradient flow and attention mechanisms to enhance feature weighting and improve predictive accuracy. Loss Function Application: Applying a Mean Squared Error (MSE) loss function to optimise prediction accuracy, particularly accounting for minor deviations in respiratory parameters relevant to asthma management. Results Display and Storage: Displaying the predictive outcome on the mobile application interface and saving the data to a secure cloud-based server. The output is presented via a traffic light indication system representing predicted asthma control status forthe following three days: green for well-controlled, amber for poorly controlled, and red for uncontrolled. [Optional] Model Validation: Subjecting the DNN model to a validation procedure comprising 10-fold cross-validation, thereby ensuring robust generalisability and model stability across diverse datasets. The invention provides a breath analysis system configured forthe prediction of asthma exacerbations. The system comprises a breath analysis device incorporating a sensor array platform capable of detecting a range of exhaled biomarkers, including volatile organic compounds (VOCs), carbon dioxide (CO2), and other species present in breath. The sensor array is designed for compatibility with standard semiconductor manufacturing techniques, enabling scalable, high-volume production, and is characterised by high sensitivity and low power consumption, which make it suitable for real-time physiological monitoring. A dedicated mobile application interfaces with the breath analysis device and is used to control breath sample collection as well as to display results to the user. The system includes a breath input interface fitted with a viral or bacterial filter through which the user exhales. A flow control module is provided to regulate the breath flow rate in real time, ensuring consistent and reproducible sampling conditions. A CO2 sensor embedded within the breath sampling path monitors exhaled CO2 levels and enables accurate identification of the end-tidal breath fraction. A software algorithm executed by a processing unit selectively activates chemical sensing only upon confirmation of the end-tidal phase, thereby improving biomarker reliability. The system is further configured to acquire a set of at least thirteen waveform parameters during exhalation, including but not limited to VOC concentration profiles, CO2 waveform characteristics, peak exhaled flow rate, temperature, and humidity. These measurements are transmitted to a secure server or cloud infrastructure for storage and further analysis. In some embodiments, the processing unit applies Principal Component Analysis (PCA) to reduce the dimensionality of the acquired data and assign greater weighting to the most influential features. In certain configurations, an embedded Edge Al module enables local data processing directly on the device without requiring an internet connection. The system optionally includes a trained deep neural network (DNN) that classifies asthma control status into at least three categories: well-controlled, poorly controlled, and uncontrolled. The DNN architecture may include residual blocks to improve gradient propagation in deep models, and attention mechanisms to dynamically prioritise relevant input features. The DNN may be trained using a mean squared error (MSE) loss function to optimise sensitivity to subtle changes in input parameters relevant to asthma control. The resulting output may be displayed using a traffic light indication system (green, amber, red), providing intuitive feedback to users regarding their condition. In some implementations, the DNN is validated using 10-fold cross-validation to ensure its generalisability and robustness across diverse patient data. The invention provides a breath analysis system configured for the prediction of asthma exacerbations. The system comprises a breath analysis device incorporating a sensor array platform capable of detecting a range of exhaled biomarkers, including volatile organic compounds (VOCs), carbon dioxide (CO2), and other species present in breath. The sensor array is designed for compatibility with standard semiconductor manufacturing techniques, enabling scalable, high-volume production, and is characterised by high sensitivity and low power consumption, which make it suitable for real-time physiological monitoring. A dedicated mobile application interfaces with the breath analysis device and is used to control breath sample collection as well as to display results to the user. The system includes a breath input interface fitted with a viral or bacterial filter through which the user exhales. A flow control module is provided to regulate the breath flow rate in real time, ensuring consistent and reproducible sampling conditions. A CO2 sensor embedded within the breath sampling path monitors exhaled CO2 levels and enables accurate identification of the end-tidal breath fraction. A software algorithm executed by a processing unit selectively activates chemical sensing only upon confirmation of the end-tidal phase, thereby improving biomarker reliability. The system is further configured to acquire a set of at least thirteen waveform parameters during exhalation, including but not limited to VOC concentration profiles, CO2 waveform characteristics, peak exhaled flow rate, temperature, and humidity. These measurements are transmitted to a secure server or cloud infrastructure for storage and further analysis. In some embodiments, the processing unit applies Principal Component Analysis (PCA) to reduce the dimensionality of the acquired data and assign greater weighting to the most influential features. In certain configurations, an embedded Edge Al module enables local data processing directly on the device without requiring an internet connection. The system optionally includes a trained deep neural network (DNN) that classifies asthma control status into at least three categories: well-controlled, poorly controlled, and uncontrolled. The DNN architecture may include residual blocks to improve gradient propagation in deep models, and attention mechanisms to dynamically prioritise relevant input features. The DNN may be trained using a mean squared error (MSE) loss function to optimise sensitivity to subtle changes in input parameters relevant to asthma control. The resulting output may be displayed using a traffic light indication system (green, amber, red), providing intuitive feedback to users regarding their condition. In some implementations, the DNN is validated using 10-fold cross-validation to ensure its generalisability and robustness across diverse patient data. The invention further provides a method for predicting asthma exacerbations using a breath analysis system configured to detect and interpret exhaled biomarkers. The method comprises configuring a breath analysis device equipped with a sensor array platform to detect multiple species, including VOCs, CO2, and additional biomarkers present in exhaled breath. The sensor array is designed for integration into standard semiconductor manufacturing processes to facilitate mass production and is characterised by high sensitivity and low power consumption, enabling real-time physiological monitoring. The method includes initiating a mobile application which controls the breath sampling procedure and interfaces with the user. A breath sample is collected via a viral or bacterial filter attached to the device, and the system controls the flow rate of the exhaled breath in real time to ensure measurement repeatability and accuracy. The method involves capturing breath specifically from the alveolar air portion using an end-tidal collection approach. A CO2 sensor monitors the CO2 waveform during exhalation, and an algorithm detects the end-tidal breath fraction based on changes in CO2 concentration. Chemical sensing is activated only when this end-tidal phase is detected, ensuring accurate and consistent biomarker acquisition. The method further comprises measuring thirteen distinct waveform parameters during exhalation, including VOC levels, CO2 parameters, peak exhaled flow rate, temperature, and humidity. The resulting data are displayed via the mobile application and securely stored on a remote server for further review. In some embodiments, Principal Component Analysis (PCA) is applied to reduce data dimensionality and assign greater weight to the most informative parameters. Edge Al processing may also be implemented to enable all data analysis to occur locally on the device without requiring internet connectivity. The method optionally includes the step of classifying asthma control states using a trained deep neural network (DNN) capable of distinguishing between well-controlled, poorly controlled, and uncontrolled asthma. The DNN may include residual blocks for improved gradient flow in deep layers and attention mechanisms for dynamically focusing on relevant features. The network may be trained using a mean squared error (MSE) loss function to minimise prediction error with emphasis on clinically relevant deviations. The classification output may be displayed to the user in a traffic light format, with green, amber, and red corresponding to respective control levels. In some configurations, the method further comprises validating the DNN using 10-fold cross-validation to ensure consistent and robust predictive performance across patient populations.

Claims

1. A breath analysis system for predicting asthma exacerbations, the system comprising:(a) a breath analysis device comprising a sensor array platform configured to detect multiple species including volatile organic compounds (VOCs), carbon dioxide (CO2), and other exhaled biomarkers;(b) the sensor array integrated with standard semiconductor manufacturing processes to enable high-volume production and configured to operate with high sensitivity and low power consumption to permit real-time physiological monitoring;(c) a mobile application configured to control breath sample collection and display results;(d) a breath input interface including a viral or bacterial filter through which a patient provides an exhaled breath sample;(e) a flow control module configured to regulate breath sample flow rate in real time to maintain stable, repeatable, and reproducible measurement conditions;(f) an embedded CO2 sensor configured to monitor CO2 concentration levels during exhalation and identify an end-tidal breath fraction;(g) a processing unit configured to execute an algorithm that activates chemical sensing only upon detection of the end-tidal breath fraction;(h) the sensor array further configured to acquire at least thirteen waveform parameters during exhalation, including VOC profiles, CO2 waveform characteristics, peak exhaled flow rate, temperature, and humidity;(i) a secure server or cloud infrastructure configured to receive and store measured data for further review.

2. The system of claim 1, wherein the processing unit is further configured to apply Principal Component Analysis (PCA) to the acquired data to reduce dimensionality and assign higherweighting to significant features.

3. The system of any preceding claim, wherein the processing unit comprises an Edge Al module configured to perform local data processing within the breath analysis device without requiring an external network connection.

4. The system of any preceding claim, further comprising a trained deep neural network (DNN) configured to classify an asthma control state into at least three categories: well-controlled, poorly controlled, and uncontrolled.

5. The system of claim 4, wherein the deep neural network comprises:(a) residual blocks to preserve gradient flow and support deep model architecture; and(b) attention mechanisms configured to assign greater importance to relevant input features during classification.

6. The system of any of claims 4 or 5, wherein the deep neural network is trained using a mean squared error (MSE) loss function, optimised to reduce small deviations important for asthma management accuracy.

7. The system of claim 4, wherein the output of the deep neural network is presented to a user as a traffic light indication system, comprising: green to indicate well-controlled asthma, amber to indicate poorly controlled asthma, and red to indicate uncontrolled asthma.

8. The system of any of claims 4 to 7, wherein the deep neural network is validated using 10-fold cross-validation to ensure generalisation and robust prediction accuracy.

9. A method for predicting asthma exacerbations, comprising:(i) configuring a breath analysis device with a sensor array platform capable of detecting multiple species including volatile organic compounds (VOCs), carbon dioxide (CO2), and other exhaled biomarkers;(ii) integrating the sensor array with standard semiconductor manufacturing processes to enable high-volume production;(iii) enabling real-time monitoring through the sensor array with high sensitivity and low power consumption; (iv) initiating a mobile application to control breath sample collection and display results;(v) collecting an exhaled breath sample from a patient through a breath interface comprising a viral or bacterial filter;(vi) controlling the breath sample flow rate in real time to ensure stable, repeatable, and reproducible measurement conditions;(vii) capturing alveolar air using an end-tidal collection method;(viii) monitoring CO2 concentration levels during exhalation using an embedded CO2 sensor;(ix) identifying an end-tidal breath fraction via a software algorithm by detecting changes in the CO2 waveform;(x) activating VOC chemical sensing only upon detection of the end-tidal breath fraction;(xi) acquiring a set of at least thirteen exhalation waveforms including VOCs, CO2 profile parameters, peak exhaled flow rate, temperature, and humidity;(xii) displaying results through the mobile application and storing the measured data on a secure remote server for subsequent review.

10. The method of claim 9, further comprising applying Principal Component Analysis (PCA) to the acquired data to reduce dimensionality and assign increased weighting to the most influential features.

11. The method of any preceding claim, wherein the breath analysis device includes an embedded processor implementing Edge Al to perform all processing locally without requiring an external network connection.

12. The method of any preceding claim, wherein a trained Deep Neural Network (DNN) classifies the asthma control state into at least three categories: well-controlled, poorly controlled, and uncontrolled.

13. The method of claim 12, wherein the deep neural network comprises:(a) residual blocks to preserve gradient flow and permit deeper network architectures; and(b) attention mechanisms configured to prioritise relevant input features in the prediction task.

14. The method of any of claims 12 or 13, wherein the deep neural network is trained using a Mean Squared Error (MSE) loss function to minimise prediction errors, with emphasis on small deviations critical to asthma management.

15. The method of claim 12, wherein the output of the deep neural network is presented as a traffic light indication system, comprising: green for well-controlled asthma, amber for poorly controlled asthma, and red for uncontrolled asthma.

16. The method of any of claims 12 to 15, further comprising validating the deep neural network using a 10-fold cross-validation approach to ensure generalisation and robust performance.A

Images