Telemonitoring of respiratory health

Abstract

The present disclosure relates to a computer-implemented method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising receiving, by a processing device, voice recordings from a patient via a mobile application, extracting, by the processing device, acoustic features from the received voice recordings, analyzing, by the processing device, the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics, establishing, by the processing device, an adaptive baseline for the patient based on the analyzed acoustic features, comparing, by the processing device, subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status, and generating, by the processing device, an output indicating changes in the patient's respiratory health status.

Description

[0001] Title: telemonitoring of respiratory health

[0002] Description:

[0003] The present invention generally relates to the field of digital health monitoring, and more particularly to the field of Al-driven telemonitoring of respiratory health in patients with chronic respiratory diseases using voice analysis.

[0004] Chronic respiratory diseases, including asthma and Chronic Obstructive Pulmonary Disease (COPD), represent a significant global health burden, affecting millions of people worldwide. Effective management of these conditions is crucial to improving patients' quality of life and reducing the risk of exacerbations, hospitalizations, and other complications. Traditionally, the management of chronic respiratory diseases has relied on periodic clinical assessments and patient-reported symptoms. These approaches, however, are often limited in their ability to capture the day-to-day variations in a patient’s condition, leading to delays in the detection of worsening symptoms or exacerbations.

[0005] In recent years, digital health technologies have increasingly been adopted to supplement traditional methods of monitoring respiratory health. These technologies include mobile applications and wearable devices that can track various health metrics, such as respiratory rate, oxygen saturation, and patient-reported symptoms. While these advancements offer the potential for more continuous monitoring, they are often limited by the reliance on intermittent data collection and the subjective nature of self-reported symptoms. As a result, there is a recognized need for more objective and reliable methods to monitor respiratory health in real-time.

[0006] Despite the advancements in telemonitoring technologies, several challenges remain. Current systems often lack the ability to accurately predict exacerbations or changes in a patient’s condition before they become clinically apparent. This shortcoming can lead to missed opportunities for early intervention, which is critical in preventing the progression of respiratory diseases. Moreover, the existing systems tend to require additional equipment or devices that may not be accessible or convenient for all patients, particularly in low-resource settings.

[0007] It is therefore a goal of the present invention to provide an improved method for monitoring respiratory health status in patients with chronic respiratory diseases, thereby overcoming the above-mentioned disadvantages of the prior art at least in part.

[0008] In an aspect of the present invention, there is provided, a computer- implemented method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving, by a processing device, voice recordings from a patient via a mobile application; extracting, by the processing device, a plurality of acoustic features from the received voice recordings; analyzing, by the processing device, a fusion of the extracted acoustic features using a machine learning-based algorithm to predict a change in the patient's respiratory health status; establishing, by the processing device, an adaptive baseline for the patient, wherein the adaptive baseline is periodically recalibrated over time to account for variability in the patient's disease trajectory based on clinical outcomes; comparing, by the processing device, a prediction from the machine learningbased algorithm based on subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating, by the processing device, an output indicating changes in the patient's respiratory health status.

[0009] Aspects of the present invention relate to a computer-implemented method for monitoring respiratory health status in patients with chronic respiratory diseases. A computer-implemented method may be understood as a process or procedure carried out by a computer system that is programmed to execute specific tasks. Monitoring respiratory health status refers to the continuous or periodic assessment of a patient's respiratory condition, particularly in identifying changes that may indicate deterioration or improvement. Chronic respiratory diseases, such as asthma and Chronic Obstructive Pulmonary Disease (COPD), are long-term conditions that affect the airways and other structures of the lung, often requiring ongoing management.

[0010] The method comprises receiving voice recordings from a patient via a mobile application. A voice recording may be understood as a digital audio file that captures the sound of a patient's voice, typically collected through a device such as a smartphone. The mobile application, in this context, refers to software installed on a mobile device that facilitates the collection and transmission of these voice recordings to a processing device. This provides for non-invasive, regular data collection from the patient without the need for additional specialized equipment, making the monitoring process more accessible and convenient for the patient.

[0011] The method further includes extracting a plurality of acoustic features from the received voice recordings. Acoustic features are specific measurable properties of sound waves that can be derived from the voice recordings. These features include but are not limited to pitch, tone and amplitude and provide a multidimensional representation of a patient's vocal characteristics. The features include, but are not limited to, several distinct categories. Frequency-related features quantify pitch and vocal stability and may include the fundamental frequency (F0 or pitch) and its cycle- to-cycle variation (jitter). Amplitude-related features measure loudness and its stability, such as overall intensity and its cycle-to-cycle variation (shimmer). Spectral features describe the timbre and resonance of the voice, prominently including Mel- frequency cepstral coefficients (MFCCs) and vocal tract formants. Voice quality features, such as the harmonics-to-noise ratio (HNR), assess the level of hoarseness or breathiness. Finally, temporal features capture the rhythm and timing of speech, including speech rate, articulation rate, and the frequency and duration of pauses.

[0012] The extraction process is typically performed by the processing device using algorithms designed to identify relevant characteristics of the voice that may correlate with the patient's respiratory health. This provides for enabling the method to quantify and analyze aspects of the patient's voice that may change due to their respiratory condition, facilitating the detection of subtle variations that might not be apparent through subjective assessment.

[0013] The method also involves analyzing the extracted acoustic features. Specifically, the method comprises analyzing, by the processing device, a fusion of the extracted acoustic features using a machine learning-based algorithm to predict a change in the patient's respiratory health status.

[0014] The term fusion of features as used herein is to be understood in the field of machine learning as the technical operation of combining a plurality of extracted acoustic features into a common representation for subsequent processing by a machine-learning model. In particular, the invention contemplates that features obtained from different acoustic domains, such as spectral features (e.g. Mel- frequency cepstral coefficients), temporal features (e.g. jitter, speech rate), and energy-related features (e.g. amplitude, loudness), are combined into a single feature vector or unified input set. This process of feature fusion ensures that the machinelearning algorithm receives as input not merely individual features evaluated in isolation, but a multidimensional dataset in which the correlation and interaction between features can be exploited. Feature fusion may be implemented, for example, by concatenating numerical values of the extracted features into a vector, by normalising and scaling features into a common domain prior to input, or by applying dimensionality-reduction techniques such as principal component analysis to construct a fused representation. By providing such fused features as input, the machinelearning model is enabled to capture complex relationships across different acoustic domains, thereby improving the sensitivity and robustness of predictions of respiratory health status.

[0015] A machine learning-based algorithm refers to a computational model that is trained on a combination of multiple features to recognize complex patterns and make predictions. In this case, the algorithm is not designed to merely identify deviations in individual vocal characteristics, but rather to predict a clinical outcome by learning the correlations between a fusion of features and the patient's respiratory health status. This provides for a more robust and accurate prediction of health status changes, such as an exacerbation.

[0016] The method further establishes an adaptive baseline for the patient. Critically, this adaptive baseline is periodically recalibrated over time to account for variability in the patient's disease trajectory based on clinical outcomes. This means the baseline is not a static or simple personalized average of vocal characteristics, but is a dynamic reference that is updated in view of the patient's actual clinical progression, such as following a recovery from an exacerbation. This provides for improving the personalization and accuracy of the monitoring process by distinguishing meaningful deteriorations in health from normal day-to-day fluctuations in the patient's condition.

[0017] The method includes comparing a prediction from the machine learning-based algorithm, which is based on subsequent analyzed acoustic features, with the adaptive baseline to detect changes in the patient's respiratory health status. This comparison is therefore not a simple feature-by-feature threshold check, but a higher-level assessment where the model's prediction of a health status change is evaluated against a clinically relevant, time-sensitive baseline. This allows for timely detection of changes in health status, enabling earlier intervention and potentially reducing the risk of severe exacerbations or complications.

[0018] Finally, the method generates an output indicating changes in the patient's respiratory health status. The output is the result provided by the processing device after the analysis and comparison, which could be in the form of a notification, report, or other types of alerts communicated to the patient or healthcare provider. This output is crucial for informing the patient or healthcare provider about potential changes in health status that may require attention or intervention. This enhances the responsiveness of the healthcare management process, providing actionable insights that can be used to adjust treatment plans or take preventive measures in a timely manner.

[0019] In existing systems for monitoring respiratory health, machine learning is typically generally applied at the stage of processing raw speech signals in order to extract elementary features such as vowels, pauses or simple acoustic markers. The outcome of such approaches typically consists of a limited number of single features, for example jitter, shimmer or pause duration, which may then be assessed against a personalised threshold or general averaged baseline. While these methods can indicate gross deviations in vocal behaviour, they remain highly sensitive to day-to- day and week-to-week variability of patients, and they lack sufficient sensitivity to reliably detect the onset of clinical events such as exacerbations. Moreover, since these prior systems treat features largely in isolation, they fail to capture the correlations and interactions between different acoustic domains that carry clinically relevant information. As a consequence, such solutions have insufficient accuracy for distinguishing genuine health deterioration from normal variation, and do not provide the level of predictive power required for effective patient management.

[0020] The present invention provides a different and surprising technical solution by inverting the approach: a plurality of acoustic features is first extracted using domainspecific preprocessing optimised for respiratory conditions, feature selection is performed, and only then is machine learning applied to this fusion of features as the input of the model. By training the model on multiple heterogeneous features and clinically validated outcome data, the system is able to learn correlations across domains and to predict exacerbations with significantly greater robustness than threshold-based assessments of single features. In addition, the invention introduces an adaptive baseline which is recalibrated over time in accordance with actual clinical outcomes, thereby distinguishing meaningful deterioration from natural day-to-day fluctuations. This combination of multi-feature fusion, correlation-based predictive modelling, and clinically anchored adaptive baselines provides a technical effect of improved accuracy and reliability in respiratory health monitoring, enabling earlier and more precise intervention in patient care.

[0021] In an example, the periodic recalibration of the adaptive baseline may be triggered following a clinical event. A clinical event may be understood as a significant change in the patient's health status, such as a diagnosed exacerbation following a medical decision, a change in treatment, or a hospitalization. An effect of this feature is that it enhances the sensitivity and context-awareness of the monitoring system. By anchoring the recalibration to a tangible change in the patient's condition, the system can more effectively distinguish between normal day-to-day symptom variability and a genuine shift in the patient's underlying disease state, thereby reducing false positives and improving the precision of health status predictions.

[0022] In an example, extracting the acoustic features may comprise domain-specific preprocessing of the voice recording, including applying respiratory-disease-specific cut-offs for feature extraction parameters, such as frame length and hop length, rather than using standard or default settings common in general audio processing. Frame length refers to the size of the short-term window over which the audio signal is analyzed, and hop length refers to the step size or interval at which these analysis windows are processed. An effect of this feature is that it enhances the performance of the machine learning algorithm by incorporating domain-specific knowledge. By tailoring the preprocessing steps to the specific nuances of voice signals affected by respiratory conditions, the extracted features are of a higher quality and contain more relevant information, which leads to more accurate and reliable predictions of the patient's health status.

[0023] In an example, the method may further comprise utilizing a model interpretation tool to identify and prioritize the acoustic features contributing to the prediction of the change in the patient's respiratory health status. A model interpretation tool, such as the well-known SHapley Additive exPlanations (SHAP) method, is an algorithm used to explain the output of a machine learning model by quantifying the contribution of each input feature to the final prediction. As such it increases the explainability of the model. An effect of this feature is that it provides transparency into the decision-making process of the machine learning algorithm. This builds trust and enhances the clinical utility of the system by allowing healthcare providers to understand which specific voice characteristics are most predictive of a change in health status, facilitating more informed clinical decision-making and potentially highlighting areas for further refinement of the predictive model. In another aspect of the present invention, there is provided, a computer- implemented method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving, by a processing device, voice recordings from a patient via a mobile application; extracting, by the processing device, acoustic features from the received voice recordings; analyzing, by the processing device, the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establishing, by the processing device, an adaptive baseline for the patient based on the analyzed acoustic features; comparing, by the processing device, subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating, by the processing device, an output indicating changes in the patient's respiratory health status.

[0024] In another aspect of the present invention there is provided a data processing apparatus for monitoring respiratory health status in patients with chronic respiratory diseases, the apparatus comprising: a mobile application configured to receive voice recordings from a patient; a processing device configured to: extract acoustic features from the received voice recordings; analyze the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establish an adaptive baseline for the patient based on the analyzed acoustic features; compare subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generate an output indicating changes in the patient's respiratory health status.

[0025] In yet another aspect of the present invention there is provided a computer program comprising instructions which, when executed by a processing device, cause the processing device to carry out a method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving voice recordings from a patient via a mobile application; extracting acoustic features from the received voice recordings; analyzing the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establishing an adaptive baseline for the patient based on the analyzed acoustic features; comparing subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating an output indicating changes in the patient's respiratory health status.

[0026] In yet another aspect of the present invention there is provided a computer- readable storage medium storing a program which, when executed by a processing device, causes the processing device to carry out a method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving voice recordings from a patient via a mobile application; extracting acoustic features from the received voice recordings; analyzing the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establishing an adaptive baseline for the patient based on the analyzed acoustic features; comparing subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating an output indicating changes in the patient's respiratory health status.

[0027] Aspects of the present invention relate to a computer-implemented method for monitoring respiratory health status in patients with chronic respiratory diseases. A computer-implemented method may be understood as a process or procedure carried out by a computer system that is programmed to execute specific tasks. Monitoring respiratory health status refers to the continuous or periodic assessment of a patient's respiratory condition, particularly in identifying changes that may indicate deterioration or improvement. Chronic respiratory diseases, such as asthma and Chronic Obstructive Pulmonary Disease (COPD), are long-term conditions that affect the airways and other structures of the lung, often requiring ongoing management.

[0028] The method comprises receiving voice recordings from a patient via a mobile application. A voice recording may be understood as a digital audio file that captures the sound of a patient's voice, typically collected through a device such as a smartphone. The mobile application, in this context, refers to software installed on a mobile device that facilitates the collection and transmission of these voice recordings to a processing device. This provides for non-invasive, regular data collection from the patient without the need for additional specialized equipment, making the monitoring process more accessible and convenient for the patient.

[0029] The method further includes extracting acoustic features from the received voice recordings. Acoustic features are specific measurable properties of sound waves, such as pitch, tone, and amplitude, that can be derived from the voice recordings. This extraction process is typically performed by the processing device using algorithms designed to identify relevant characteristics of the voice that may correlate with the patient's respiratory health. This provides for enabling the method to quantify and analyze aspects of the patient's voice that may change due to their respiratory condition, facilitating the detection of subtle variations that might not be apparent through subjective assessment.

[0030] The method also involves analyzing the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics. A machine learning-based algorithm refers to a computational model that is trained on data to recognize patterns and make predictions based on new inputs. In this case, the algorithm is designed to identify deviations in vocal characteristics, which may include changes in pitch, breathiness, or other acoustic features, that are indicative of changes in the patient's respiratory health. This provides for enhancing the accuracy and reliability of detecting these changes, as the algorithm can learn from a large dataset and continuously improve its predictions, reducing the likelihood of false positives or missed detections.

[0031] The method further establishes an adaptive baseline for the patient based on the analyzed acoustic features. An adaptive baseline refers to a reference point that is dynamically adjusted over time to reflect the normal range of variations in the patient's vocal characteristics. This baseline is beneficial because it accounts for the individual differences in voice and the natural progression of the chronic respiratory disease. This provides for improving the personalization of the monitoring process, ensuring that alerts or notifications are based on meaningful deviations from what is normal for that particular patient, rather than on a fixed or generic standard.

[0032] The method includes comparing subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status. This comparison is the process by which the method identifies significant deviations from the established baseline, which may suggest a deterioration or improvement in the patient's condition. This allows for timely detection of changes in health status, enabling earlier intervention and potentially reducing the risk of severe exacerbations or complications. By continuously updating the comparison with the patient's own adaptive baseline, the method provides a more accurate and context-sensitive assessment of the patient's respiratory health.

[0033] Finally, the method generates an output indicating changes in the patient's respiratory health status. The output is the result provided by the processing device after the analysis and comparison, which could be in the form of a notification, report, or other types of alerts communicated to the patient or healthcare provider. This output is crucial for informing the patient or healthcare provider about potential changes in health status that may require attention or intervention. This enhances the responsiveness of the healthcare management process, providing actionable insights that can be used to adjust treatment plans or take preventive measures in a timely manner. By implementing the described method, the invention provides an improved approach for monitoring respiratory health status in patients with chronic respiratory diseases, overcoming the disadvantages of prior art at least in part. The method leverages advanced algorithms to offer a reliable and personalized assessment based on voice recordings, facilitating more accurate and timely detection of changes in the patient's condition. This approach enhances the effectiveness of telemonitoring by providing continuous, non-invasive monitoring, thereby supporting better management of chronic respiratory diseases.

[0034] In an example, the method may include the feature wherein the adaptive baseline is updated over time to account for natural changes in the patient's disease trajectory. It may be provided that an adaptive baseline is understood as a dynamically adjusting reference that reflects the ongoing changes in a patient’s condition over time. An effect of this feature is that it enhances the accuracy of monitoring by continuously refining the baseline against which new data is compared, ensuring that the method remains sensitive to relevant changes in the patient's respiratory health while reducing false positives that could arise from natural fluctuations in the disease.

[0035] In an example, the method may further comprise the feature of predicting, by the processing device, the onset of exacerbations based on detected deviations from the adaptive baseline. It may be provided that the onset of exacerbations refers to the early stages of a significant worsening in the patient’s respiratory condition. An effect of this feature is that it allows for the timely identification of exacerbations, which is crucial for prompt medical intervention. By detecting deviations from a dynamically updated baseline, this feature helps in providing accurate predictions, thereby enabling proactive management and potentially reducing the severity of exacerbations.

[0036] In an example, the method may further comprise the feature wherein the voice recordings are received periodically, preferably, daily from the patient. It may be provided that periodic reception of voice recordings involves collecting data at regular intervals, such as daily, which ensures a consistent flow of information for monitoring purposes. An effect of this feature is that it enables continuous and up-to-date assessment of the patient's respiratory health, providing a more reliable and comprehensive understanding of their condition over time. Regular data collection is particularly advantageous in tracking gradual changes and responding to health variations promptly.

[0037] In an example, the method may further comprise utilizing a validated symptom score as a ground truth for training the machine learning-based algorithm. It may be provided that a validated symptom score is a clinically recognized metric that quantifies the severity of symptoms in a standardized manner. Such validated domainspecific scores, add domain-specific knowledge to the model and increases accuracy for use in the population with respiratory conditions. An effect of this feature is that it provides a robust foundation for training the machine learning algorithm, ensuring that its predictions are grounded in clinically relevant data. This improves the reliability of the algorithm and enhances its applicability in real-world clinical settings.

[0038] In an example, the method may include the feature wherein the validated symptom score is one of a plurality of clinically recognized metrics. Such metrics may include but are not limited to, for example, the COPD Assessment Test (CAT), the Clinical COPD Questionnaire (CCQ), or the Exacerbations of Chronic Pulmonary Disease Tool (EXACT) score for patients with chronic obstructive pulmonary disease, or the Asthma Control Questionnaire (ACQ) for patients with asthma. The EXACT score for example, is configured for tracking changes in daily symptoms of chronic obstructive pulmonary disease. It may be provided that the EXACT or any of the other scores is a tool specifically designed to monitor daily fluctuations in symptoms related to COPD, which is already FDA approved. An effect of this feature is that it aligns the machine learning algorithm with a metric that is specifically tailored to the disease in question, thereby enhancing the precision of the monitoring process for COPD patients. This targeted approach ensures that the method is particularly effective in managing COPD, offering more nuanced insights into the patient’s condition.

[0039] In an example, the method may comprise the validation of the input and / or output of the model by health-care professionals. The interpretation and valuing of the input features or model output by individuals with domain specific knowledge increases the accuracy and the trustworthiness of the model. An effect of this feature is that the machine learning algorithm input and output are features that are clinically relevant and meaningful for use in the population with respiratory conditions.

[0040] In an example, the method may further comprise the feature of continuously updating the adaptive baseline and the machine learning-based algorithm based on longitudinal tracking of the patient's voice recordings. It may be provided that longitudinal tracking refers to the continuous or repeated monitoring of a patient's condition over an extended period. An effect of this feature is that it allows the system to adapt to the patient’s evolving health status, making the monitoring process more responsive and personalized. Continuous updates to the baseline and algorithm ensure that the system remains accurate and relevant, reflecting the current state of the patient’s respiratory health.

[0041] In an example, the method may further comprise the feature of providing realtime alerts and recommendations to the patient via the mobile application based on the detected changes in respiratory health status. It may be provided that real-time alerts refer to immediate notifications delivered to the patient as soon as significant changes in health status are detected. An effect of this feature is that it enhances the timeliness of interventions, enabling patients to take prompt action based on the system's assessments. Real-time recommendations further empower patients to manage their condition more effectively, potentially preventing exacerbations and improving overall health outcomes.

[0042] In an example, the method may be configured for monitoring by use of only a smartphone with a microphone. It may be provided that this configuration allows the method to be implemented using widely available consumer technology without the need for additional specialized equipment. An effect of this feature is that it significantly lowers the barrier to adoption, making the monitoring process more accessible to a broad patient population. The use of a smartphone also facilitates the seamless integration of the monitoring system into the patient’s daily routine, promoting consistent use and data collection. In an example, the method may include the feature wherein the chronic respiratory diseases include at least one of asthma and Chronic Obstructive Pulmonary Disease (COPD). It may be provided that this inclusion allows the method to be applicable to a wide range of respiratory conditions that share common symptoms and monitoring needs. An effect of this feature is that it broadens the applicability of the method, making it useful for managing multiple types of chronic respiratory diseases. This versatility ensures that the system can be adapted to different patient populations with varying needs.

[0043] In an example, the method may include the feature wherein the machine learning-based algorithm comprises a regression model trained on spectral and phonation features extracted from the voice recordings. It may be provided that a regression model is a type of machine learning model that predicts continuous outcomes, and spectral and phonation features refer to the specific characteristics of the voice that can be analyzed to assess respiratory health. An effect of this feature is that it allows for precise modeling of the relationship between voice characteristics and health status, leading to accurate and nuanced assessments of the patient’s condition.

[0044] In an example, the method may include the feature wherein the acoustic features include Mel-frequency cepstral coefficients (MFCC). It may be provided that MFCCs are a representation of the short-term power spectrum of a sound signal, commonly used in speech and audio processing. An effect of this feature is that it enhances the ability of the machine learning algorithm to accurately capture the relevant characteristics of the patient's voice, providing a more reliable basis for assessing respiratory health. The use of MFCCs is particularly effective in distinguishing subtle variations in voice that may be indicative of changes in health status.

[0045] In an example, the method may include the feature wherein extracting the MFCC features comprises applying a pre-emphasis filter to the voice recording, performing short-time Fourier transform (STFT) on the filtered signal, mapping the powers of the spectrum obtained from the STFT onto the mel scale, taking the logarithm of the powers at each of the mel frequencies, and applying the discrete cosine transform (DCT) to the list of mel log powers. It may be provided that this sequence of operations is designed to process the voice signal in a way that highlights the most relevant features for analysis. An effect of this feature is that it ensures the extraction of high-quality acoustic features, which are crucial for the accurate functioning of the machine learning algorithm.

[0046] In an example, the method may include the feature wherein the number of MFCC features extracted is at least 2, preferably between 1 and 100, preferably between 2 and 50, 4 and 40, and most preferably between 10 and 20. It may be provided that varying the number of MFCC features allows for optimization of the feature set used by the machine learning algorithm, balancing between computational efficiency and accuracy. An effect of this feature is that it enables the method to be tailored to different processing capacities and accuracy requirements, making it adaptable to various implementation contexts.

[0047] In an example, the method may include the feature wherein the machine learning-based algorithm is a K-Nearest Neighbors (KNN) algorithm trained using 10- fold cross-validation to predict respiratory health events. It may be provided that the KNN algorithm is a simple, yet effective machine learning method that classifies data points based on the majority vote of its neighbors, and 10-fold cross-validation is a technique used to evaluate the performance of the algorithm by partitioning the data into ten parts and training on nine parts while testing on the remaining one. An effect of this feature is that it provides a robust and validated model for predicting respiratory health events, ensuring that the algorithm performs well across different datasets.

[0048] In an example, the method may include the feature wherein the KNN algorithm is trained with a training set to an accuracy of at least 50%, more preferably at least 60%, at least 70%, at least 80%, and most preferably 85% in classifying respiratory health events. It may be provided that the accuracy of the KNN algorithm refers to its ability to correctly classify health events based on the training data. An effect of this feature is that it ensures a high level of confidence in the predictions made by the algorithm, making it a reliable tool for monitoring respiratory health. In an example, the method may include the feature wherein the machine learning-based algorithm is a Support Vector Regression (SVR) model trained on spectral features to predict daily symptom score deviations. It may be provided that SVR is a type of regression algorithm that aims to predict continuous values by finding the best-fit line within a threshold in a high-dimensional space, and spectral features refer to the characteristics of the sound signal that can be analyzed to assess respiratory health. An effect of this feature is that it allows for precise modeling of the relationship between voice characteristics and symptom severity, providing accurate and actionable predictions of health status.

[0049] In a further example, the method may include the feature wherein the SVR model is trained to an accuracy of at least 50%, more preferably at least 60%, at least 70%, at least 80%, and most preferably 85% in classifying respiratory health states as improvement, stable, or deterioration. It may be provided that the accuracy of the SVR model refers to its ability to correctly predict the classification of health states. An effect of this feature is that it ensures the model provides reliable assessments of the patient's condition, which can inform timely and appropriate.

[0050] In an example, the method may include the feature wherein the extracted acoustic features comprise a fusion of features from multiple acoustic domains, including spectral, temporal, and energy-related features. It may be provided that a fusion of features from multiple acoustic domains refers to the combination of distinct characteristics from different aspects of the sound signal, such as spectral (e.g., frequency distribution), temporal (e.g., time-related patterns), and energy-related features (e.g., amplitude and loudness). An effect of this feature is that it allows for a more comprehensive analysis of the voice data, capturing a wider range of information that could be indicative of changes in the patient's respiratory health. This approach enhances the model's ability to detect subtle variations in the voice that may be missed when only a single domain is analyzed, thereby improving the accuracy and robustness of the monitoring system.

[0051] In an example, the method may further comprise utilizing a model interpretation tool, such as SHapley Additive exPlanations (SHAP), to identify and prioritize the most salient acoustic features contributing to the prediction of changes in the patient's respiratory health status. It may be provided that SHAP is a tool used in machine learning to explain the output of models by attributing the importance of each feature to the final prediction. An effect of this feature is that it provides transparency in the decision-making process of the machine learning algorithm, allowing healthcare providers to understand which specific voice features are most predictive of health status changes. This transparency can increase trust in the system, facilitate clinical decision-making, and potentially highlight areas for further refinement in the algorithm.

[0052] In an example, the method may include the feature wherein the adaptive baseline is periodically recalibrated based on the continuous collection of voice recordings and patient-reported outcomes, with the baseline being reset at predefined intervals, such as every four weeks. It may be provided that periodic recalibration refers to the regular adjustment of the baseline to reflect the most current state of the patient’s condition. An effect of this feature is that it ensures the monitoring system remains sensitive and accurate over time, accounting for natural variations in the patient’s health and avoiding potential drifts in the baseline that could lead to inaccurate assessments. This dynamic adjustment supports the long-term reliability and precision of the telemonitoring method.

[0053] In an example, the method may include the feature wherein the extracted acoustic features include specific speech characteristics, such as pitch and jitter. It may be provided that pitch refers to the perceived frequency of sound waves produced by the vocal cords, while jitter measures the frequency variation from cycle to cycle in vocal fold vibration. An effect of this feature is that it allows the system to detect changes in voice that are closely related to respiratory function, as these characteristics are often impacted by respiratory distress or exacerbations. Incorporating these features enhances the system's ability to identify early signs of deteriorating respiratory health.

[0054] In an example, the method may include the feature wherein the voice recordings are received at predefined intervals, including morning, midday, and evening sessions. It may be provided that collecting voice recordings at multiple points throughout the day allows for capturing a comprehensive representation of the patient's vocal characteristics, accounting for potential diurnal variations in voice. An effect of this feature is that it improves the overall accuracy and reliability of the monitoring system by providing a more consistent data stream for analysis, which is essential for detecting subtle changes in respiratory health.

[0055] In an example, the method may include the feature wherein the machine learning-based algorithm is trained and validated using a dataset specific to a particular patient population, such as COPD patients with predefined characteristics. It may be provided that population-specific models refer to algorithms that are trained on data from patients who share similar demographic or clinical characteristics, ensuring that the predictions are highly relevant to the target group. An effect of this feature is that it increases the predictive accuracy and clinical relevance of the monitoring system, making the method more effective in managing the respiratory health of specific patient populations, thereby supporting personalized medicine approaches.

[0056] In a further embodiment of the present disclosure, the method for monitoring respiratory health status in patients with chronic respiratory diseases may be implemented in a distributed system comprising at least two devices: a mobile device for recording and transmitting voice data and a remote server or cloud-based processing device for analyzing the data. In this embodiment, the mobile device, which may be a smartphone, a dedicated voice recorder, or a wearable device with microphone capabilities, is configured to capture voice recordings from the patient. The mobile device transmits the recorded voice data to the remote server over a network connection, such as Wi-Fi, 4G, or 5G, for subsequent processing and analysis.

[0057] The method comprises receiving, at the remote processing device, voice recordings transmitted from the mobile device. The remote processing device is further configured to extract acoustic features from the received voice recordings, analyze the extracted features using a machine learning-based algorithm, and establish an adaptive baseline for the patient’s respiratory health status. The use of the remote server or cloud device allows for the offloading of computationally intensive tasks from the mobile device, leveraging the greater processing power and storage capacity of the server. This ensures the method remains efficient and scalable, especially for patients using low-resource devices.

[0058] In some implementations, the analysis performed by the remote server includes the steps of detecting deviations in vocal characteristics that may indicate changes in the patient's respiratory health. These changes are then communicated back to the mobile device, allowing the patient to receive real-time feedback via notifications or alerts generated by the mobile application. Additionally, the remote server can integrate data from other health monitoring systems or medical records, further enhancing the accuracy of the analysis and enabling the system to deliver personalized recommendations or alerts to the patient or healthcare provider.

[0059] This distributed architecture provides the benefit of performing voice data collection locally while offloading intensive data analysis to a more capable remote device. The mobile device can thus conserve battery and processing resources, while the remote server, with access to advanced machine learning models, can process larger datasets and continuously update the patient's adaptive baseline as new data becomes available. Furthermore, this setup enables integration with cloud-based health systems, providing additional flexibility and scalability for broader telehealth applications.

[0060] The above-mentioned and other features and advantages of the invention are illustrated in the following description with reference to the enclosed drawings which are provided by way of illustration only and which are not limitative to the present invention.

[0061] Fig. 1. shows a schematic view of an embodiment of a data processing apparatus in accordance with an aspect of the present disclosure;

[0062] Fig. 2. shows a schematic view of an embodiment of an computer-implemented method in accordance with another aspect of the present disclosure; and Fig. 3. shows a schematic view of an embodiment of substeps of the computer- implemented method of Fig. 2.

[0063] Figure 1 shows a data processing apparatus 1 for monitoring respiratory health status in patients with chronic respiratory diseases, such as asthma and / or Chronic Obstructive Pulmonary Disease (COPD).

[0064] The data processing apparatus 1 , in this embodiment a smartphone of the patient, comprises an internal computer-readable storage medium 3, a microphone 5 for receiving voice recordings 2 from a patient, a processing device 7 and a display 9 for generating output to the patient.

[0065] The computer-readable storage medium 3 stores a computer program comprising instructions, which, when executed by the processing device 7, causes the processing device 7 to carry out the steps of a computer-implemented method 100 in accordance with another aspect of the present disclosure.

[0066] The computer-implemented method 100 is implemented as part of a mobile application that provides continuous monitoring and feedback to patients with chronic respiratory diseases, enhancing self-management and early intervention capabilities.

[0067] The computer-implemented method 100 for monitoring respiratory health status in patients with chronic respiratory diseases, as schematically shown in figure 2, comprises the steps of: receiving 101 , by the processing device 7, voice recordings 2 from the patient via the microphone 5 and via the mobile application, wherein the voice recordings 2 are received periodically from the patient; extracting 103, by the processing device 7, acoustic features from the received voice recordings 2; analyzing 105, by the processing device 7, the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; - establishing 107, by the processing device 7, an adaptive baseline for the patient based on the analyzed acoustic features, wherein the adaptive baseline is updated over time to account for natural changes in the patient's disease trajectory; comparing 109, by the processing device 7, subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status, wherein the detected changes in respiratory health status are used to provide personalized treatment advice or lifestyle recommendations to the patient; predicting 111 , by the processing device 7, the onset of exacerbations based on detected deviations from the adaptive baseline; and generating 113, by the processing device 7, an output, via the display 9, indicating changes in the patient's respiratory health status, wherein real-time alerts and recommendations are provided to the patient via the mobile application based on the detected changes in respiratory health status.

[0068] The computer-implemented method 100 utilizes a validated symptom score as a ground truth for training the machine learning-based algorithm, wherein the validated symptom score is the EXACT (Exacerbations of Chronic Pulmonary Disease Tool) score, configured for tracking changes in daily symptoms of chronic obstructive pulmonary disease.

[0069] The adaptive baseline and the machine learning-based algorithm are continuously updating based on longitudinal tracking of the patient's voice recordings 2.

[0070] The acoustic features include in particular M el-frequency cepstral coefficients (MFCC), wherein the number of MFCC features extracted is preferably between 10 and 20. The step of extracting 103 the MFCC features from the received voice recordings 2 comprises the following substeps as shown in more detail in figure 3: applying 103A a pre-emphasis filter to the voice recording 2; performing 103B short-time Fourier transform (STFT) on the filtered signal; mapping 103C the powers of the spectrum obtained from the STFT onto the mel scale; taking 103D the logarithm of the powers at each of the mel frequencies; and applying 103E the discrete cosine transform (DCT) to the list of mel log powers.

[0071] The machine learning-based algorithm comprises a regression model trained on spectral and phonation features extracted from the voice recordings 2. The machine learning-based algorithm is a K-Nearest Neighbors (KNN) algorithm trained using 10- fold cross-validation to predict respiratory health events, wherein the KNN algorithm is trained with a training set to an accuracy of preferably 85% in classifying respiratory health events. Furthermore, the machine learning-based algorithm is a Support Vector Regression (SVR) model trained on spectral features to predict daily symptom score deviations, wherein the SVR model is trained to an accuracy of preferably 85% in classifying respiratory health states as improvement, stable, or deterioration.

[0072] CLAUSES

[0073] 1. A computer-implemented method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving, by a processing device, voice recordings from a patient via a mobile application; extracting, by the processing device, acoustic features from the received voice recordings; analyzing, by the processing device, the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establishing, by the processing device, an adaptive baseline for the patient based on the analyzed acoustic features; comparing, by the processing device, subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating, by the processing device, an output indicating changes in the patient's respiratory health status.

[0074] 2. The method of claim 1 , wherein the adaptive baseline is updated over time to account for natural changes in the patient's disease trajectory, and / or wherein the voice recordings are received periodically, preferably, daily from the patient.

[0075] 3. The method of claim 1 , further comprising predicting, by the processing device, the onset of exacerbations based on detected deviations from the adaptive baseline and / or further comprising utilizing a validated symptom score as a ground truth for training the machine learning-based algorithm, and preferably wherein the validated symptom score is the EXACT (Exacerbations of Chronic Pulmonary Disease Tool) score, configured for tracking changes in daily symptoms of chronic obstructive pulmonary disease.

[0076] 4. The method of claim 1 , further comprising continuously updating the adaptive baseline and the machine learning-based algorithm based on longitudinal tracking of the patient's voice recordings. 5. The method of claim 1 , further comprising providing real-time alerts and recommendations to the patient via the mobile application based on the detected changes in respiratory health status.

[0077] 6. The method of claim 1 , wherein the method is configured for monitoring by use of only a smartphone with a microphone.

[0078] 7. The method of claim 1 , wherein the chronic respiratory diseases include at least one of asthma and Chronic Obstructive Pulmonary Disease (COPD).

[0079] 8. The method of claim 1 , wherein the machine learning-based algorithm comprises a regression model trained on spectral and phonation features extracted from the voice recordings.

[0080] 9. The method of claim 1 , wherein the acoustic features include Mel-frequency cepstral coefficients (MFCC), and preferably wherein extracting the MFCC features comprises: applying a pre-emphasis filter to the voice recording; performing short-time Fourier transform (STFT) on the filtered signal; mapping the powers of the spectrum obtained from the STFT onto the mel scale; taking the logarithm of the powers at each of the mel frequencies; applying the discrete cosine transform (DCT) to the list of mel log powers, and more preferably wherein the number of MFCC features extracted is at least 2, preferably between 1 and 100, between 2 and 50, 4 and 40 and most preferably between 10 and 20.

[0081] 10. The method of claim 1 , wherein the machine learning-based algorithm is a K- Nearest Neighbors (KNN) algorithm trained using 10-fold cross-validation to predict respiratory health events, and preferably wherein the KNN algorithm is trained with a training set to an accuracy of at least 50%, more preferably at least 60%, at least 70%, at least 80% and most preferably 85% in classifying respiratory health events. 11. The method of claim 1 , wherein the machine learning-based algorithm is a Support Vector Regression (SVR) model trained on spectral features to predict daily symptom score deviations, and preferably wherein the SVR model is trained to an accuracy of at least 50%, more preferably at least 60%, at least 70%, at least 80% and most preferably 85% in classifying respiratory health states as improvement, stable, or deterioration.

[0082] 12. The method of claim 1 , further comprising using the detected changes in respiratory health status to provide personalized treatment advice or lifestyle recommendations to the patient.

[0083] 13. The method of claim 1 , wherein the method is implemented as part of a mobile application that provides continuous monitoring and feedback to patients with chronic respiratory diseases, enhancing self-management and early intervention capabilities.

[0084] 14. A data processing apparatus for monitoring respiratory health status in patients with chronic respiratory diseases, the apparatus comprising: a mobile application configured to receive voice recordings from a patient; a processing device configured to: extract acoustic features from the received voice recordings; analyze the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establish an adaptive baseline for the patient based on the analyzed acoustic features; compare subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generate an output indicating changes in the patient's respiratory health status.

[0085] 15. A computer program comprising instructions which, when executed by a processing device, cause the processing device to carry out a method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving voice recordings from a patient via a mobile application; extracting acoustic features from the received voice recordings; analyzing the extracted acoustic features using a machine learning-based algorithm to detect changes in vocal characteristics; establishing an adaptive baseline for the patient based on the analyzed acoustic features; comparing subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating an output indicating changes in the patient's respiratory health status.

Claims

28CLAIMS1. A computer-implemented method for monitoring respiratory health status in patients with chronic respiratory diseases, the method comprising: receiving, by a processing device, voice recordings from a patient via a mobile application; extracting, by the processing device, a plurality of acoustic features from the received voice recordings; analyzing, by the processing device, a fusion of the extracted acoustic features using a machine learning-based algorithm to predict a change in the patient's respiratory health status; establishing, by the processing device, an adaptive baseline for the patient, wherein the adaptive baseline is periodically recalibrated over time to account for variability in the patient's disease trajectory based on clinical outcomes; comparing, by the processing device, a prediction from the machine learningbased algorithm based on subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generating, by the processing device, an output indicating changes in the patient's respiratory health status.

2. The method of claim 1 , wherein the periodic recalibration of the adaptive baseline is triggered following a clinical event, such as an exacerbation recovery period.

3. The method of claim 1 , wherein extracting the acoustic features comprises domain-specific preprocessing of the voice recording, including applying respiratory- specific cut-offs for feature extraction parameters such as frame length and hop length.

4. The method of claim 1 , further comprising utilizing a model interpretation tool, such as SHapley Additive exPlanations (SHAP), to identify and prioritize the acoustic features contributing to the prediction of the change in the patient's respiratory health status.

5. The method of claim 1 , further comprising utilizing a validated symptom score as a ground truth for training the machine learning-based algorithm, and preferably wherein the validated symptom score is the EXACT (Exacerbations of Chronic Pulmonary Disease Tool) score, configured for tracking changes in daily symptoms of chronic obstructive pulmonary disease.

6. The method of claim 1 , further comprising continuously updating the machine learning-based algorithm based on longitudinal tracking of the patient's voice recordings.

7. The method of claim 1 , further comprising using the output to provide realtime alerts and personalized treatment advice or lifestyle recommendations to the patient via the mobile application.

8. The method of claim 1 , wherein the machine learning-based algorithm comprises a regression model trained on spectral and phonation features extracted from the voice recordings.

9. The method of claim 1 , wherein the machine learning-based algorithm is a K- Nearest Neighbors (KNN) algorithm trained using 10-fold cross-validation to predict respiratory health events.

10. The method of claim 1 , wherein the machine learning-based algorithm is a Support Vector Regression (SVR) model trained on spectral features to predict daily symptom score deviations.

11. The method of claim 1 , wherein the acoustic features include Mel-frequency cepstral coefficients (MFCC), and preferably wherein extracting the MFCC features comprises: applying a pre-emphasis filter to the voice recording; performing short-time Fourier transform (STFT) on the filtered signal; mapping the powers of the spectrum obtained from the STFT onto the mel scale;taking the logarithm of the powers at each of the mel frequencies; applying the discrete cosine transform (DCT) to the list of mel log powers.

12. The method of claim 1 , wherein the chronic respiratory diseases include at least one of asthma and Chronic Obstructive Pulmonary Disease (COPD).

13. The method of claim 1 , wherein the method is configured for monitoring by use of only a smartphone with a microphone.

14. A data processing apparatus for monitoring respiratory health status in patients with chronic respiratory diseases, the apparatus comprising: a mobile application configured to receive voice recordings from a patient; a processing device configured to: extract a plurality of acoustic features from the received voice recordings; analyze a fusion of the extracted acoustic features using a machine learning-based algorithm to predict a change in the patient's respiratory health status; establish an adaptive baseline for the patient, wherein the adaptive baseline is periodically recalibrated over time to account for variability in the patient's disease trajectory based on clinical outcomes; compare a prediction from the machine learning-based algorithm based on subsequent analyzed acoustic features with the adaptive baseline to detect changes in the patient's respiratory health status; generate an output indicating changes in the patient's respiratory health status.

15. A computer program comprising instructions which, when executed by a processing device, cause the processing device to carry out the method of any one of claims 1 to 13.

Images