Method and system for music based diagnosis of physiological properties

Abstract

Embodiments of the present disclosure undertake physiological measurements of a subject whilst the subject is listening to music. The physiological measurements, together with predetermined characteristics of the music, are then used as inputs into a machine learning network, such as a graph convolution neural network (GCNN), which undertakes a classification based on the physiological measurements and musical properties as to whether the subject has a particular disease or pathological condition, such as for example hypertension. The use of the music whilst obtaining the physical measurements resulted in an increase in accuracy of the measurement of the disease or condition prevalence of greater than 10 to 15%. This improvement is due to the modulatory effects of music on the autonomic nervous system augmenting differences between people with distinct autonomic profiles.

Description

[0001] P608486PC00

[0002] METHOD AND SYSTEM FOR MUSIC BASED DIAGNOSIS OF PHYSIOLOGICAL PROPERTIES

[0003] Field

[0004] The present disclosure relates to a method and system for music-based diagnosis of physiological properties of a human subject. In one embodiment, the physiological property is whether the human subject has hypertension.

[0005] Background

[0006] Music listening leads to changes in cardiovascular functional parameters in individuals. Dousty et al. demonstrated the effect of different types of music on electrocardiogram (ECG) signals [1]. Using ECG sensors, Wu et al. evaluated changes in heart rate variability in the presence of music [2]. Time delay stability in music signals along with the cardiac signals of musicians and listeners was investigated by Solinski et al. [3], showing TDS during specific music events. Mixed linear models have been used to predict players' heart rates given their interpretation decisions, accounting for up to 60% of the variability in beat-to-beat heart intervals [4]. Loomba & Arora [5] reviewed literature on music's impact on vitals - (systolic and diastolic) BP and heart rate. Other research claim that music influences not only emotions but also cardiac rhythms as expressed via physiological signals (ECG, photoplethysmography (PPG), respiration), blood pressure (BP), and skin conductance [6]. Furthermore, music-based interventions affect hypertensive individuals and those with abnormal endothelial function differently from those with normal physiological function [7- 10].

[0007] BP is one of the crucial factors which is considered for cardiac well-being

[0011] . Abnormality in the cardiovascular system is apprehended by BP values beyond the normal range of 120 / 80 mmHg. Clinically, BP beyond 139 / 89 mmHg indicates hypertension. Hypertension stresses the blood vessels. If untreated, hypertension can lead to heart failure or renal problems. Early diagnosis can allow interventions to be prescribed to avoid life threatening outcomes. Hypertension risk has been analysed by Cano et al. using ECG signal features and KNN classifier with 83% accuracy

[0012] . Liang et al. was able to classify hypertensives with an Fl-score of 84.34% using ECG and PPG signals. None have attempted to classify hypertensives using physiological signals collected from participants during music listening

[0013] . P606567PC00

[0008] Summary of the Disclosure

[0009] Cardiovascular risk impacts the cardiac autonomic response. Individuals with different states of autonomic function respond to music differently. Furthermore, our research shows that the presence of music augments the difference between individuals with distinct autonomic traits. This amplification of autonomic differences can be exploited in cardiovascular diagnostics.

[0010] To our knowledge, there is no prior work on music based cardiovascular diagnostics. Cardiovascular diagnostics in the form of hypertension diagnosis has been studied without music. For example, Cano et al. (2022) used physiological signals like ECG with machine learning to classify hypertensives with 83% accuracy. With music as an integral part of the investigation, we set out to provide a method and system to distinguish hypertensives from normal individuals. Analysis of concurrent ECG, respiration, and pulse signals whilst listening to music of altered loudness and tempi was performed. Graph convolution neural network (GCNN) was implemented for node classification where each node in the graph represented a single individual. The physiological signal features formed the node features and the edges were established based on the attributes of the music they heard. Music was central to the building of the graph connectivity.

[0011] Embodiments of the present disclosure undertake physiological measurements of a subject whilst the subject is listening to music. The physiological measurements, together with predetermined characteristics of the music, are then used as inputs into a machine learning network, such as a graph convolution neural network (GCNN), which undertakes a classification based on the physiological measurements and musical properties as to whether the subject has a particular disease or pathological condition, such as for example hypertension. The use of the music whilst obtaining the physical measurements resulted in an increase in accuracy of the measurement of the disease or condition prevalence of greater than 10 to 15%. This improvement is due to the modulatory effects of music on the autonomic nervous system augmenting differences between people with distinct autonomic profiles.

[0012] Embodiments of the present disclosure therefore provide for the use of music when collecting the physiological data, and then processing the physiological data together with predetermined properties of the music using a machine learning network, such as a GCNN, to classify the physiological data as being indicative of a pathological condition or not, such as hypertension. The effect of this feature is to increase the accuracy of classification of a subject as having the pathological condition, such as hypertension, the accuracy for which increased by up to 10-15%. P606567PC00

[0013] In view of the above, a first aspect of the present disclosure provides a system for identification of at least one target pathological disease or condition in human subjects, comprising: a processor, arranged in use to: receive a plurality of predefined physiological measurements of a human subject and to derive a set of physiological features therefrom, the physiological measurements having been obtained in the presence of a reproduction of one or more known pieces of music; input the physiological features into a machine learning network, the machine learning network having been trained taking into account musical features of the one or more known pieces of music and the identity of the predefined physiological measurements; and classify, using the trained machine learning network, the human subject to which the physiological features relate as having the at least one target pathological disease or condition based on the input physiological features.

[0014] The advantages of such a system for identifying at least one target pathological disease or condition in human subjects are numerous. For example, the use of music during the collection of physiological measurements enhances the accuracy of the classification of subjects as having a pathological condition, such as hypertension. The accuracy of the classification increased by up to 10-15%. In addition, music has modulatory effects on the autonomic nervous system, which augments the differences between individuals with distinct autonomic profiles. This amplification can be exploited in cardiovascular diagnostics.

[0015] The system also offers a non-invasive and enjoyable method for hypertension diagnosis that can be integrated into daily activities. This approach can potentially avoid "white coat syndrome," where a subject's blood pressure temporarily increases during clinical measurements, giving a false impression of hypertension. Moreover, the system can be used for early hypertension detection and adaptive digital music therapeutics, especially as physiological sensors become more prevalent in consumer devices such as headphones or earbuds.

[0016] These advantages make the system a valuable tool for diagnosing pathological conditions in a non-invasive, accurate, and user-friendly manner

[0017] In one example embodiment, the machine learning network is a Graph Convolution Neural Network, with nodes of the network representing subjects having known classifications of the at least one target pathological disease or condition and the edges of the network representing the musical features of the one or more known pieces of music. In this respect, the edges not only carry the music features, they also are used to connect people who listened to the same piece of music P606567PC00

[0018] In one example embodiment, the musical features include one or more selected from the set comprising: loudness, tempo, timbre, and spectral centroid. Other embodiments may include other features, either in combination with or as alternatives to these features.

[0019] In one example embodiment, the at least one target pathological disease or condition is one or more of: hypertension; hypotension; or blood pressure within a predefined range.

[0020] Further features and advantages will be apparent from the appended claims.

[0021] Further features and advantages of the present disclosure will become apparent from the following description, presented by way of example only, and with reference to the accompanying drawings, wherein like reference numerals refer to like parts, and wherein:

[0022] Figure 1 is a diagram illustrating how physiological data is collected from a human subject;

[0023] Figure 2 is a series of graphs illustrating how feature measures are extracted from the physiological data collected from the human subject;

[0024] Figure 3 is a schematic diagram of the classification process;

[0025] Figure 4 is a block diagram of a machine learning computer system arranged to implement embodiments of the present disclosure;

[0026] Figure 5 is a flow diagram illustrating the method of training the machine learning computer system; and

[0027] Figure 6 is a flow diagram illustrating the method of operating the trained machine learning computer system to detect physiological properties of the human subject.

[0028] Overview of Embodiments of the Disclosure

[0029] One of the core insights of the present disclosure is that music modulates the autonomic nervous system, amplifying differences in cardiovascular responses between individuals with distinct autonomic profiles. This amplification can be exploited diagnostically by collecting physiological measurements during music listening and analyzing them alongside musical features using machine learning. P606567PC00

[0030] In view of the above the present disclosure relates to a music-based diagnostic tool for hypertension using a Graph Convolution Neural Network (GCNN) that analyzes ECG, respiration, and pulse signals. Embodiments of the disclosure leverage the impact of music on cardiovascular function, noting that music amplifies autonomic differences, which can be exploited for cardiovascular diagnostics. We understand this to be the first known attempt to use music-based diagnostics for hypertension, differentiating it from existing ECG- (or other physiological signal-)based methods which do not involve music.

[0031] Embodiments of the disclosure use a GCNN for node classification, where each node represents an individual subject and edges are based on the music they heard. Node features are obtained by analyzing the electrocardiogram (E) data, respiration (R) data, and pulse signals (P) during music listening. The edge features are derived from the music: loudness, tempo, timbre, and spectral centroid. In this respect, to train the network, data was collected from 70 individuals listening to different versions of Chopin's Nocturne (this could be a different piece, or a different set of pieces), with ECG, respiration, and pulse signals processed and analyzed to derive features used in the GCNN.

[0032] Training the Graph Convolutional Neural Network (GCNN) uses physiological measurements and known hypertensive states of training subjects. The steps are as follows:

[0033] 1. Acquire E, R, and P measurements in silence (for system evaluation purposes): This step involves collecting physiological measurements ECG, respiration, and pulse (respectively E, R, and P) data from the training subjects while they are in a silent environment.

[0034] 2. Acquire E, R, and P measurements while listening to known music: This step involves collecting the same physiological measurements while the training subjects are listening to known music.

[0035] 3. Signal process E, R, and P measurements: The collected measurements are then processed (as described further later) to extract the relevant E, R and P features.

[0036] 4. Generate additional features E*R*P, E*P, E*R, R*P: Additional features are then generated by combining the E, R, and P measurements in various ways.

[0037] 5. Train GCNN using {E, R, P, E*P, E*R, R*P, E*R*P} training features derived for each training subject, and the known hypertensive state of each training subject'. The GCNN is then trained using the extracted physiological features and music features (such as Loudness, Tempo, Timbre, and Spectral Centroid, but in other embodiments other features may be used) and the known hypertensive states of the training subjects. P606567PC00

[0038] 6. Store trained GCNN network: The trained GCNN network is then stored for future use in classifying new subjects.

[0039] Once trained, the GCNN can then be used to classify new subjects whose hypertensive state is unknown. This is performed as follows:

[0040] 1. Acquire E, R, and P measurements in silence-. This step involves collecting physiological measurements (E, R, and P) from the subject while they are in a silent environment. Please note that this is an optional feature required for evaluation purposes, and is not essential and may not be performed in real-world embodiment.

[0041] 2. Acquire E, R, and P measurements while listening to music. This step involves collecting the same physiological measurements while the subject is listening to music.

[0042] 3. Signal process E, R, and P measurements-. The collected measurements are then processed to extract relevant features, as will be described in more detail later.

[0043] 4. Generate additional features E*R*P, E*P, E*R, R*P: Additional features are then generated by combining the E, R, and P measurements in various ways.

[0044] 5. Execute the GCNN using {E, R, P, E*P, E*R, R*P, E*R *P} features derived for new subject-. The GCNN is then executed using the new extracted physiological features and music features (Loudness, Tempo, Timbre, and Spectral Centroid).

[0045] 6. Add new node to GCNN with musical edge features-. A new node is added to the Graph Convolutional Neural Network (GCNN) with the extracted features to represent the subject to be classified.

[0046] 7. Classify new node as hypertensive or not: The GCNN classifies the new node based on the features, determining whether the subject is hypertensive or not.

[0047] When deployed on test subjects whose hypertensive state is to be determined, the GCNN showed improved classification accuracy with music, achieving up to 0.90 accuracy when music features were included, indicating the effectiveness of music in enhancing diagnostic accuracy. With such accuracy embodiments of the disclosure can be used for early hypertension detection and adaptive digital music therapeutics, especially as physiological sensors become more prevalent in consumer devices such as headphones, or earbuds such as Apple® AirPods®. In addition, embodiments of the disclosure offer a non-invasive, pleasurable method for hypertension diagnosis that can be integrated into daily activities, potentially avoiding so-called "white coat syndrome" - where a subject's blood pressure temporarily increases during the act of measuring of the blood pressure in the clinic or doctor's office, thus giving a false impression of the subject being hypertensive. P606567PC00

[0048] Detailed Description

[0049] An embodiment of the present disclosure will now be described. Figure 1 is a schematic that shows how data is collected in embodiments of the disclosure. Here a human subject 10 is shown sat on a chair. The human subject 10 has connected to them an ECG monitor 462, a respiration monitor 464, and a pulse signal monitor 466. The ECG monitor 462, the respiration monitor 464 and the pulse signal monitor 466 are each connected to a computer system 40, which monitors and records the signals produced by the monitors. In addition, the computer system 40 acts to control a music reproduction system 44, to reproduce music at the same time as it is monitoring the physiological signals from the human subject 10.

[0050] To gain sufficient training data, multiple human subjects 10 are subjected to monitoring, involving physiological monitoring of their ECG, respiration, and pulse signal, whilst listening to music. The cohort for this study comprised of 70 (41 women) normal and hypertensive individuals (aged 18-80 years, mean = 47.76±12.90, average BMI= 24.42±3.6). Two versions of Chopin's Nocturne in F sharp minor, Op. 48 No. 2, performed by Josef Hofmann — the original (normal) (duration 7:51) and artificially sped up (3: 10) — was played to participants via a disklavier piano. Only tempo differed between the two versions. Three sets of physiological signals - ECG (462) via a H10 (Polar Electro Oy, Kempele, Finland) sensor, respiration (464) via a BIOPAC (BIOPAC Systems, Goleta, USA) band, and pulse signal via a CNAP sensor (466) (CN Systems, Graz, Austria) - were simultaneously acquired from this cohort whilst they listened to the music. The total duration of data collection was 12:51 if they heard the normal version and 8: 10 if they heard the sped up version, including the initial 5 minutes of silence as baseline. Gold standard blood pressure BP values were recorded during the initial baseline silence.

[0051] Figure 5 shows the system training method in more detail. More particularly, at step 5.2 a processing loop is started to obtain training data from each training subject. For each training subject, firstly, at step 5.4 ECG, respiration, and pulse signals (E, R, and P) are acquired from the training subject, whilst no music is being played. This is to provide baseline measurements against which obtained via music can be compared. Once the baseline measurements have been obtained, next, at step 5.6, the ECG, respiration, and pulse signals (E, R, and P) are then measured again whilst the subject is listening to known music.

[0052] Once the ECG, respiration, and pulse signals have been obtained, at step 5.8 signals are processed to filter and extract the necessary information. As shown in Figure 2, the ECG was filtered with the Butterworth band pass filter (BPF) with cut-off frequency 0.2-500 Hz P606567PC00

[0053] (sampling frequency (sF)= 3000Hz) and order 2. The respiration signal was filtered with rolling and moving averages, and window size 200 frames (sF= 13 Hz). BPF with cut-off frequency 0.4-8Hz (sF= 300Hz) and order 2 filtered the pulse signal. Filtered ECG signals were mean standardised. The product of QT and RR intervals was extracted from the ECG signal as the E feature. In this respect, in an ECG (electrocardiogram), the QT interval measures the time from the start of the QRS complex (beginning of ventricular depolarization) to the end of the T wave (end of ventricular repolarization), and represents the total time for the ventricles to contract and then recover, whereas the RR interval is the time between two successive R-wave peaks on the ECG, and is used to calculate the heart rate and assess the regularity of the heart rhythm

[0054] For the R feature, the ratio of the area of the inspiration to the expiration portions of the respiratory cycle was set as the R feature. The augmentation index of the pulse signal, the ratio of the amplitude of the first wave to that of the second wave in a cardiac cycle, gave the P feature. The combined individual features, (E, R, and P), are then used at s.5.10 to get the derived features: ER= E*R, EP=E*P, RP=R*P, and ERP=E*R*P. The cohort — each member having 7 features — was segregated into training / validation / test sets in the proportion 80 / 10 / 10. These method steps are repeated for each training subject, so that a set of seven features is obtained for each training subject.

[0055] Next, at s.5.14, the identifying of raised BP individuals was conducted using node classification from GCNN as shown in Fig. 2. Here, each node represents one individual characterised by the 7 features extracted from the signals. The edge connectivity was established by the version of music heard by the individual. The edge features were evaluated by processing the music signal to obtain 4 features: loudness, tempo, timbre, and spectral centroid. Graph properties are utilised to procure the attributes which identify the unknown node (individual) as potentially hypertensive or normal.

[0056] With regards to the neural network (GCNN), the input to the network comprised of 70 x 7 nodes and 2 x 489 edges mapped to 1 x 7 actual data. Edge features were included as 489x 4 matrix in the input. This followed three layers of graph convolution with 6 hidden channels using the rectified linear unit (RELU) as the non-linearity function. The hyperparameters involved in the network execution are the Adams optimizer as the training algorithm, learning rate of 0.01 and weight decay rate of 6e-5. The training of the network was executed with 56 nodes for 200 epochs. The validation and testing were done with the resulting predicted 1 x 7 data array as the end product of network execution. The confusion matrix was used to derive the result matrices. Once the neural network has been trained it is stored at s.5.16 for later use to classify unknown subjects. P606567PC00

[0057] The block diagram in Figure 4 describes a computer system 40 used for classifying a subject as hypertensive using a machine learning network. The system comprises the following elements: a) Physiological Measurements (46): This component acquires physiological measurements (E, R, and P) from the subject, as described previously. b) Music Reproducer (44): This component plays known music to the subject while the physiological measurements are being taken. c) CPU (402): The central processing unit that executes the instructions of the ML diagnostic program 420 and the ML training program 418, under the control of the control program 412 and operating system 410. d) Memory (404): The memory stores the ML training program, ML diagnostic program, trained ML network, and other necessary data to allow the CPU to run the programs. e) I / O (406): The input / output interface for communication with external devices. f) O / S (410): The operating system that manages the computer system's resources. g) Control Program (412): The control program that coordinates the overall operation of the system. h) ERP Training Data (416): The training data used to train the machine learning network. This is derived from the 70 volunteer subjects as described above, and comprises the 7 feature vector (E, R, P, ERP, EP, ER, and PR) for each volunteer subject. i) ML Training Program (424): The program used to train the machine learning network, using the ERP training data 416. j) Music data (414): The music data played to the subjects via the music reproducer 44 during physiological data collection, and their features (for example, loudness, tempo, timbre, and spectral centroid). k) ERP Input Data for New Subject (422): The physiological measurements taken in silence and while listening to music are used as input data for the new subject. They are processed by the ML Diagnostic program 408 in the manner described above to get the 7-feature vector (E, R, P, ERP, EP, ER, and PR) characterising the P606567PC00 new test subject, and which are used with music features to classify the new test subject as hypertensive or not in the GCNN. l) ML Diagnostic Program (420): This program receives the E, R, P measurements for the new subject, and generates the relevant features i.e. the 7-feature vector (E, R, P, ERP, EP, ER, and PR) characterising the new test subject. This is then used as the ERP input data 422 for the new subject as an input into the trained ML network 410. m) Trained ML Network (424): The extracted and generated features for the new subject i.e. the 7 feature vector (E, R, P, ERP, EP, ER, and PR) characterising the new test subject, the music they listened to and their features are fed into a pretrained machine learning network, being in preferred embodiments a GCNN. n) New Subject Classification Data (426): The machine learning network classifies the new subject as hypertensive or not based on the input data and extracted features.

[0058] With the above system elements, the system operates as follows. The control program (412) coordinates the overall operation, and the ML training program (418) uses the ERP training data (416) and music features (414) to train the machine learning network. Once the network has been trained, the physiological measurements (46) and music reproducer (44) provide input to the ERP input data for a new subject (422). The ERP input data (422) is processed and controlled by the ML diagnostic program (420), which runs on the CPU (402) and uses memory (404) for storage. The ML diagnostic program (420) uses the trained ML network (424) to classify the new subject, and the results are stored as new subject classification data (426). The I / O (406) interface allows communication with external devices, and the O / S (410) manages the system's resources.

[0059] Figure 6 shows the process of classifying a new subject whose hypertensive state is not known, using the trained machine learning network. The steps are performed for each new subject to be classified (s.6.2). Firstly, at s.6.4, acquire E, R, and P measurements in silence. This step involves collecting the physiological measurements (E, R, and P) as described earlier from the subject while they are in a silent environment. Next, at s.6.6, acquire the same physiological measurements E, R, and P while listening to known music. In this respect the music is known to the extent that the features of the music match the music features on which the classifier network has been trained. That is, the music itself does not need to be known, but its features must have been calculated. This also leads to the possibility that different music having the same features as that on which the network was trained could be used. P606567PC00

[0060] Once the physiological measurements have been obtained, at s.6.8, signal process the E, R, and P measurements to extract relevant features as described above with respect to Figure 2. Then, at s.6.10, generate additional features ERP, EP, ER, RP by combining the E, R, and P measurements in various ways i.e. to produce the 7-feature vector set (E, R, P, ERP, EP, ER, and PR).

[0061] Once the feature vector has been obtained for the new test subject, a new node representing the new test subject is added to the Graph Convolutional Neural Network (GCNN) with the extracted features (s.6.12), and then the GCNN classifies the new node based on the features, determining whether the subject is hypertensive or not (s.6.14). An output signal indicating the hypertensive classification is then made, for example by displaying the classification on the display screen 42.

[0062] Results

[0063] Table 1 shows the classification results. The performance metrics of the classifier were obtained by implementing 5-fold cross-validation as well as with the leave-one-out method on the dataset. The importance of music in this approach was studied by comparing the results with that derived in the absence of music information in the following study.

[0064] Without edges, the GCNN becomes a multilayer perceptron (MLP). During baseline period (absence of music) the MLP gave classification results as 0.73 accuracy, 0.72 recall, 0.73 specificity, and 0.74 Fl-score (first line of Table I). The outcome of MLP classification during music but without incorporating music information was 0.75 accuracy, 0.76 recall, 0.74 specificity, and 0.77 Fl-score (Table I, line 2). When adding music information by introducing edges connecting individuals who listened to the same music version gave 0.81 accuracy, 0.80 recall, 0.81 specificity, and 0.79 Fl-score (3rd line of Table I). Inclusion of expressive music features in the form of edge features produced the results: 0.90 accuracy, 0.88 recall, 0.89 specificity, and 0.88 Fl-score (Table I, line 4).

[0065] Table I: Summary of results P606567PC00

[0066] Table II: Comparison of results with different inputs

[0067] The decision of considering all seven features as GCNN input was made after inspecting the results of the different input options. Table II shows the different results for each input set. Observe that the performance of individual signal features was weaker (0.69 accuracy, 0.68 recall, 0.67 specificity, and 0.68 Flscore), combined features performed better but were not outstanding (0.71 accuracy, 0.73 recall, 0.70 specificity, and 0.72 Fl score), whereas all the features combined dominated the classification assessment.

[0068] In conclusion, the GCNN's ability to classify hypertensives increased in the presence of music, i.e. with the edge connectivity features. This showed that connecting the cohorts based on the music they listened to (segregated by expressive features such as loudness and tempo) was markedly more efficient in distinguishing hypertensives from normal BP individuals. The improvement of the music-based model suggests that the modulatory effects of music on the autonomic nervous system augments differences between people with distinct autonomic profiles.

[0069] Various modifications may be made to the above-described embodiments to provide further embodiments, any and all of which are intended to be encompassed by the appended claims. For example, whilst we describe above the machine learning architecture that is used in the example embodiment to be a Graph Convolutional Neural Network (GCN or GCNN), in other example embodiments a different machine learning architecture may be used, for example any of the following :

[0070] Graph Attention Networks (GATs): These networks use attention mechanisms to assign different weights to different nodes in a neighborhood, allowing the model to focus on the most relevant nodes during the aggregation process;

[0071] GraphSAGE (Graph Sample and Aggregation): This method generates embeddings by sampling and aggregating features from a node's local neighborhood. It is designed to handle large graphs more efficiently; P606567PC00

[0072] Graph Isomorphism Networks (GINs): GINs are designed to be as powerful as the Weisfeiler-Lehman graph isomorphism test, making them capable of distinguishing a wide variety of graph structures;

[0073] Message Passing Neural Networks (MPNNs): These networks generalize GCNs by using a message-passing framework where nodes exchange information with their neighbors through messages, which are then aggregated;

[0074] Graph Recurrent Neural Networks (GRNNs): These networks incorporate recurrent neural network architectures to capture sequential dependencies in graph data;

[0075] Graph Autoencoders (GAEs): These are unsupervised models that learn to encode graph data into a latent space and then decode it back to the original graph structure. They are often used for tasks like link prediction and node clustering; or

[0076] Temporal Graph Networks (TGNs): These networks are designed to handle dynamic graphs where the structure and features of the graph change over time.

[0077] Moreover, in other variations, instead of requiring separate sensors to measure ECG, R, and P, in further embodiments modern health monitoring smart devices such as smart watches and earphones can be provided with suitable sensors to measure the ECG, R, and P signals that are needed as input into the trained network of the present disclosure. This then opens up the possibility of routine monitoring of hypertension and other conditions using the techniques of the present disclosure during everyday activities. For example, with a suitably equipped earphone (for example such as shown in US20230225659A1), or smart watch implementing the system of the present disclosure, routine and everyday monitoring of users as hypertensive can be undertaken, with appropriate information messages provided depending on the results of the monitoring (for example to inform a user that they are or are not hypertensive that day).

[0078] Moreover, as music listening is pleasurable, office BP measurements could be made to be more like home BP measurements. Many people spend numerous hours listening to music through their headphones. As in-ear headphones begin to acquire the ability to record measurements, physiological data during music listening may become ubiquitous, and the invention could be used for early detection of hypertension or adaptive digital music therapeutics with feedback. In this respect, hypertension is the most important risk for cardiovascular disease. Monitoring blood pressure values is essential to cardiovascular prevention and treatment. Music listening is pervasive in the general population and is accessed throughout the day. As physiological sensors become more prevalent in headphones and ear buds, early diagnosis of hypertension during music listening using P606567PC00 deep learning techniques will become an important technology for cardiovascular healthcare. In other embodiments different physiological features other than or in addition to the E, R, P, EP, ER, PR, EPR features described above may be used. Moreover, alternative and / or additional music features not limited to loudness, tempo, and spectral centroid may also be employed.

[0079] Within the above described embodiments a binary classification of a subject exhibiting hypertension or not is made. However, in other embodiments a more granular classification may be provided, by appropriate training of the neural network with appropriately classified training subjects. For example, the machine learning system may be trained to recognize ternary or quaternary classifications, with ternary classifications being for example non-elevated BP, elevated BP, or HT following clinical guidelines, or a quaternary classification further including whether a subject is hypotensive in addition to the ternary classifications. Taking this concept further, with a large enough training cohort the machine learning system can further be trained to classify a subject into measurement bands of blood pressure measurements in mmHg, for example to classify a subject as having blood pressure within a measurement band 10 mmHg wide, or less. This would effectively make the system capable not just of classifying an individual as hyper- or hypotensive but also taking a measurement of their blood pressure. At a limit, the classification bands might only be 1 mmHg wide, which then effectively makes the classification system a blood pressure sensor. Such a more granular system might be particularly applicable when the sensor is being used in an earbud or smart watch, or other worn sensor system.

[0080] In order to allow for the above, a significantly larger training cohort is required for the machine learning system, with multiple training examples having measured blood pressures within the granular bands to be measured.

[0081] References

[0082] 1] Dousty, M., et al. The effects of sedative music, arousal music, and silence on electrocardiography signals. Journal of electrocardiology, 3, 44, 396-el.2011

[0083] [2] Wu, M.H. and Chang, T.C. Evaluation of Effect of Music on Human Nervous System by Heart Rate Variability Analysis Using ECG Sensor. Sensors & Materials, 33. 2021.

[0084] [3] Solinski, M., Reed, C.N. and Chew, E., Time Delay Stability Analysis of Pairwise Interactions Amongst Ensemble-Listener RR Intervals and Expressive Music Features. Computing in Cardiology (CinC), IEEE, 50, 1-4. 2023 P606567PC00

[0085] [4] Solihski, Mateusz, Courtney N. Reed, and Elaine Chew. "A framework for modeling performers' beat-to-beat heart intervals using music features and Interpretation Maps." Frontiers in Psychology 15 (2024): 1403599.

[0086] [5] Loomba, R.S., et al. Effects of music on systolic blood pressure, diastolic blood pressure, and heart rate: a meta-analysis. Indian heart journal, 3, 64, 309-313. 2012

[0087] [6] Bretherton, B., Deuchars, J. and Windsor, W.L., The effects of controlled tempo manipulations on cardiovascular autonomic function. Music & Science, 2, p.2059204319858281. 2019.

[0088] [7] Somayaji K, Frenkel M, Tabaza L, Visotcky A, Ruck TK, Ofori EK, Widlansky ME, Kulinski J. Acute effects of singing on cardiovascular biomarkers. Front Cardiovasc Med. 2022 Jul 18;9:869104. doi: 10.3389 / fcvm.2022.869104. PMID: 35924212; PMCID: PMC9339901.

[0089] [8] Pope, Vanessa, Mateusz Solinski, Pier D. Lambiase, and Elaine Chew. "Raised Blood Pressure Alters Reactivity to Musical Features." In Congress of the European Society of Cardiology 2024.

[0090] [9] Somayaji K, Frenkel M, Tabaza L, Visotcky A, Ruck TK, Ofori EK, Widlansky ME, Kulinski J. Acute effects of singing on cardiovascular biomarkers. Front Cardiovasc Med. 2022 Jul 18;9:869104. doi: 10.3389 / fcvm.2022.869104. PMID: 35924212; PMCID: PMC9339901.

[0091]

[0010] Pope, Vanessa, Mateusz Solinski, Pier D. Lambiase, and Elaine Chew. "Raised Blood Pressure Alters Reactivity to Musical Features." In Congress of the European Society of Cardiology 2024.

[0092]

[0011] Ghosh, A., Torres, et al. Detection of essential hypertension with physiological signals from wearable devices. IEEE Engineering in Medicine and Biology Society, 8095-8098. 2015

[0093]

[0012] Cano, J.,et al. Cuffless Hypertension Risk Assessment and the Significance of Calibration. Computing in Cardiology, IEEE, 498, 1-4.2022

[0094]

[0013] Liang, Y.,et al. Hypertension assessment via ECG and PPG signals: An evaluation using MIMIC database. Diagnostics, 3, 8, 65.2018.

Claims

P608486PC00Claims1. A system for identification of at least one target pathological disease or condition in human subjects, comprising: a. a processor, arranged in use to: i. receive a plurality of predefined physiological measurements of a human subject and to derive a set of physiological features therefrom, the physiological measurements having been obtained in the presence of a reproduction of one or more known pieces of music; ii. input the physiological features into a machine learning network, the machine learning network having been trained taking into account musical features of the one or more known pieces of music and the identity of the predefined physiological measurements; and iii. classify, using the trained machine learning network, the human subject to which the physiological features relate as having the at least one target pathological disease or condition based on the input physiological features.

2. A system according to claim 1, wherein the machine learning network is a Graph Convolution Neural Network, with nodes of the network representing subjects having known classifications of the at least one target pathological disease or condition and the edges of the network representing the musical features of the one or more known pieces of music.

3. A system according to claims 1 or 2, wherein the musical features include one or more selected from the set comprising: loudness, tempo, timbre, and spectral centroid.

4. A system according to any of the preceding claims, wherein the at least one target pathological disease or condition is one or more of: hypertension; hypotension; or blood pressure within a predefined range.

5. A system according to any of the preceding claims, wherein the physiological measurements include measurements derived from one or more of:A) an ECG signal;B) a respiration signal (R); and / orC) a pulse pressure signal (P).P608486PC006. A system according to claim 5, wherein the ECG signal is processed to generate a feature E as one of the physiological features, the feature E corresponding to a product of QT and RR intervals extracted from the ECG signal.

7. A system according to claims 5 or 6, wherein the respiration signal is processed to generate feature R as one of the physiological features, the feature R corresponding to the ratio of the area of the inspiration to the expiration portions of the respiratory cycle.

8. A system according to any of claims 5 to 7, wherein the pulse signal is processed to generate feature P as one of the physiological features, the feature P corresponding to an augmentation index of the pulse signal, being a ratio of an amplitude of a first wave to that of a second wave in a cardiac cycle.

9. A system according to any of claims 5 to 8, wherein the physiological features comprise a plurality of features selected from the group comprising:E, R, P, ER= E*R, EP=E*P, RP=R*P, and ERP=E*R*P10. A system according to claim 9, wherein the physiological features comprise, for each human subject, all of the features E, R, P, ER= E*R, EP=E*P, RP=R*P, and ERP=E*R*P.

11. A method for identification of at least one target pathological disease or condition in human subjects, comprising: a. receiving a plurality of predefined physiological measurements of a human subject and deriving a set of physiological features therefrom, the physiological measurements having been obtained in the presence of a reproduction of one or more known pieces of music; b. inputting the physiological features into a machine learning network, the machine learning network having been trained taking into account music listening information and musical features of the one or more known pieces of music and the identity of the predefined physiological measurements; and c. classifying, using the trained machine learning network, the human subject to which the physiological features relate as having the at least one target pathological disease or condition based on the input physiological features.

12. A method according to claim 11, wherein the machine learning network is a Graph Convolution Neural Network, with nodes of the network representing subjectsP608486PC00 having known classifications of the at least one target pathological disease and the edges of the network representing music listening information) and the musical features of the one or more known pieces of music.

13. A method according to claims 11 or 12, wherein the musical features include one or more selected from the set comprising: loudness, tempo, timbre, and spectral centroid.

14. A method according to any of claims 11 to 13, wherein the at least one target pathological disease is one or more of : hypertension; hypotension; or blood pressure within a predefined range..

15. A method of training a machine learning network to identify at least one target pathological disease in human subjects, comprising: a. receiving a plurality of sets of predefined physiological measurements respectively relating to a set of human subjects having known pathology for at least one target pathological disease, the sets of physiological measurements having been obtained in the presence of a reproduction of one or more known pieces of music; b. for the plurality of sets of physiological measurements, deriving respective sets of physiological features therefrom; c. for the one of more known pieces of music, deriving one or more musical features thereof, and music listening information; and, d. training a machine learning network with the plurality of sets of physiological features, respective known disease pathologies, music listening information, and the one or more derived musical features to obtain a trained machine learning network.

Images