WO2013106700A1 - Systems and methods for determining mental and physical health using multi-scale metrics - Google Patents

Abstract

A system and method is provided for subject monitoring and acquisition of data for determining the mental and/or physical health of a subject. In an embodiment, the system and method provide automated subject monitoring and signal collection (acquisition) of data, for example activity data, associated with one or more of a plurality of metrics for mental and physical health, processing of the data to extract features associated with mental and/or physical health and then classifying the features extracted from the data using one or more classifiers into behavioral subject categories. The system and method can simultaneously combine multiple metrics to determine mental and/or physical health of a subject, which is robust to artifacts and missing data.

Description

SYSTEMS AND METHODS FOR DETERMINING MENTAL AND

PHYSICAL HEALTH USING MULTI-SCALE METRICS

Cross-Reference to Related Application

This application claims priority to U.S. provisional application entitled "SYSTEMS AND METHODS FOR DETERM INING MENTAL AND PHYSICAL HEALTH USING MULTI- SCALE METRICS" having Serial No. : 61/631 ,822, filed on January 12, 2012, which is expressly incorporated herein by reference as if fully set forth herein in its entirety.

Cross-Reference To Related Documents

This application incorporates by reference the following papers as if they were fully set forth herein expressly in their entireties:

(A) Osipov et al., "Detection of abnormal activity for early warning of patient deterioration using mobile networks," CDT In Healthcare Innovation Summer Project, 201 1 , pp. 1 -12 (Attached hereto as Appendix A)

(B) Gederi et al., "Fusion of Image and Signal Processing for Detection of Abnormal Sleep Structure," CDT In Healthcare Innovation Summer Project, 201 1 , pp. 1 -12 (Attached hereto as Appendix B)

(C) Joachim Behar - St Hilda's College, "Analysis of accelerometer data for apnea screening," Master of Science in Biomedical Engineering at the University of Oxford, submitted August 201 1 , pp. 1 -60 (Attached as Appendix C to the above-referenced U.S. provisional application )

(D) Roebuck, "Detection of Abnormal Respiratory Sounds and Sleep Apnoea Using Audio Recording," University of Oxford summer report, submitted August 23, 201 1 , 60 pages (Attached as Appendix D to the above-referenced U.S. provisional application)

(E) J. Behar et al. An automated OSA screening application for smartphones.

Submitted to IEEE, Transactions on Information Technology in Biomedicine (Attached hereto as Appendix E)

Field of the Invention

The present disclosure is directed to novel systems and methods for determining the mental and physical health of a subject. The present disclosure is directed, in particular, to systems and methods for collecting data representative of one or more behavioral disorders, such as sleep-related disorders, and classifying the data using a multi-scale metrics method and machine learning to determine mental and/or physical health. Background

It is thought that the effects of various sleep-related disorders are extensive, impacting sufferers physically, psychologically and financially [J.L. Hossain and CM.

Shapiro. See, e.g., The prevalence, cost implications, and management of sleep disorders: an overview. Sleep Breath., 6(2):85-102, 2002]. Changes in physical activity and sleep structure have been shown to correlate with a range of short and long term medical conditions from cardiac problems to mental health issues [R.G. Foster and K. Wulff, "The rhythm of rest and excess," Nat Rev Neurosci, vol. 6, no. 5, pp. 407-414, May 2005. [Online]. Available: http://dx.doi.org/10.1038/nrn16701fA. Wirz-Justice, "Chronobiology and psychiatry," Sleep Medicine Reviews, vol. 11 , no. 6, pp. 423-427, Dec. 2007.[Online].

Available: http://www.smiv-iournal.com/article/S1087-0792(07)00118-9/abstract.lfK. Wulff, S. Gatti, J.G. Wettstein, and R.G. Foster, "Sleep and circadian rhythm disruption in psychiatric and neurodegenerative disease," Nat Rev Neurosci, vol. 11 , no. 8, pp. 589-599m 2010. [Online]. Available: http://dx.doi.org/10.1038/nrn2868].

For example, the health effects of sleeping disorders span a wide range: from the apparently simple daytime sleepiness to the more severe effects of increased risk of cardiovascular disease and stroke [T. Young, P.E. Peppard, and D.J. Gottlieb.

Epidemiology of obstructive sleep apnea: a population health perspective. Am. J. Resp. Crit. Care, 1965(9): 1217, 2002]. In fact, excessive daytime sleepiness (EDS) is the cause of hundreds of road traffic accidents, and has even been linked to disasters such as Chernobyl [J.L. Hossain and CM. Shapiro. The prevalence, cost implications, and management of sleep disorders: an overview. Sleep Breath., 6(2):85-102, 2002]. It is estimated that 93% of females and 82% of males with obstructive sleep apnea (OSA) go undiagnosed and untreated [T. Young, L. Evans, L. Finn, and M. Palta. Estimation of the clinically diagnosed proportion of sleep apnea syndrome in middle-aged men and women. Sleep, 20(9): 705-706, 1997].

Historically, activity and sleep monitoring have taken the form of infrequent and relatively brief recordings (such as overnight polysomnography studies) taken on days (or nights) that may not be representative of the actual behavior or physiology of the patient [A. Sadeh, "The role and validity of actigraphy in sleep medicine: An update," Sleep Medicine Reviews, vol. 15, no. 4, pp. 259-267, Aug. 2011. [Online]. Available:

http://www.sciencedirect.com/science/article/pii/S1087079210001292]. For example, the recordings may be anomalous due to either temporal changes (associated with circadian fluctuations for example), recording location or screening bias [Hug, Clifford & Reisner, Critical Care 201 1 ] or causal factors (such as 'white coat syndrome' [T.G. Pickering, "Clinical applications of ambulatory blood pressure monitoring: the white coat syndrome," Clinical and Investigative Medicine. Mdecine Clinique EtExperimentale, vol. 14, no. 3, pp. 212-217, June 1991 , PMID: 1893653. [Online]. Available:

http://www.ncbi.nlm.niri.qov/pubmed/1893653l.fApendix A1 Moreover, potentially important fluctuations in physiology between such tests are therefore not captured.

Although continuous (day-to-day) monitoring is preferable, it may be expensive and require the analysis of large and often noisy tracts of data. Traditional inexpensive continuous or long term monitoring methods for physical activity are based upon actigraphy, generally using a wrist-worn accelerometer [A. Sadeh, "the role and validity of actigraphy in sleep medicine: An update," Sleep Medicine Reviews, vol. 15, no. 4, pp. 259-267, Aug. 2011. [Online]. Available:

http://www.sciencedirect.com/science/article/pii/S1087079210001292]. However, accelerometry is of limited use and provides a three-dimensional view of physical activity and optionally levels of ambient light. At the same time, new methods for actigraphy data analysis are being used, such as Fourier analysis, Wavelet analysis and entropy based methods [P. Indie, P. Salvatore, C. Maggini, S. Ghidini, G. Ferraro, R.J. Baldessarini, and G. Murray, "Scaling behavior of human locomotor activity amplitude: Association with bipolar disorder," PLoS ONE, vol. 6, no. 5, p. e20650, May 2011. [Online]. Available:

http://dx.doi.org/10.1371/journal.pone.0020650], [E. R. Hauge, J. Berle, J.J. Oedegaard, F. Holsten, and O.B. Fasmer, "Nonlinear analysis of motor activity shows differences between schizophrenia and depression: A study using Fourier analysis and sample entropy," PLoS ONE, vol. 6, no. 1 , p. e16291 , Jan. 2011. [Online]. Available:

http://dx.doi.org/10.1371/journal.pone.0016291], [R. Meadows, S. Venn, J. Hislop, N.

Stanley, and S. Arber, "Investigating couples' sleep: an evaluation of actigraphic analysis techniques." J Sleep Res, vol. 14, no. 4, p. 377386, Dec. 2005 [Online]. Available:

http://dx.doi.Org/10.111 1/j.1365-2869.2005.00485.x]. However, none of these techniques capture the changes in nonlinear complexity of the actigraphy over different time scales. Actigraphy is commonly used in sleep monitoring, although it is of limited use outside of normal populations. The field of sleep is perhaps the most well documented, understood and explored field of physiological monitoring (perhaps because other fields of medical monitoring suffer from the confounder that the patient is either conscious or suffering from multifactorial problems, for which there are often a series of unpredictable exogenous interventions, such as in the intensive care unit).

Currently, the 'gold standard' in terms of sleep disorder diagnosis is a sleep study, or an overnight polysomnogram (PSG). A PSG is carried out overnight in a hospital, using multiple sensors to detect apneas, hypoapneas and arousals. However, PSGs are expensive and are limited by the number of beds available in the hospital and the number of sleep specialists who can interpret the data. Further, some patients experience different sleep patterns due to artificial condition of the sleep laboratory. Moreover, a PSG often requires multiple monitoring modalities and intrusive electrodes attached to the patient which may disrupt sleep. Many home sleep recording systems on the market aim to reduce the financial cost and reach a larger population by reducing the number of parameters recorded. Essentially though, such systems are mini-PSG devices. However, the patient, who has no medical or technical training, has to place the sensors in the correct positions. If done incorrectly, the results may be inconclusive even if an expert is available to read the data. Accuracy is improved by introducing more parameters, which makes the device more cumbersome as well as more expensive and can interfere with the patients' sleep leading to an unrepresentative result. In the past few years, sleep specialists have begun to use systems which monitor alternative parameters such as audio and video. Recording audio data does not require accurate placement of sensors, thereby removing the likelihood of inconclusive data. A key issue in analyzing such data is the identification of events within the data and disentangling artifacts from real data. Trained specialists are not always readily available to analyze the data and may not be able to do so just from audio and

accelerometry data.

Summary

The present systems and methods address and overcome the aforementioned disadvantages. They provide automated subject monitoring and acquisition of data for determining the mental and/or physical health of the subject. In an embodiment, they provide automated subject monitoring and signal collection (acquisition) of data, for example activity data, associated with one or more of a plurality of metrics for mental and physical health, processing of the data to extract features associated with mental and/or physical health and then classifying the features extracted from the data using one or more classifiers into behavioral subject categories. The present systems and methods can simultaneously combine multiple metrics to determine mental and physical health of a subject, which is robust to artifacts and missing data. A subject can be, for example, a mammalian subject, in particular a human subject.

In an embodiment, we apply metrics to evaluate changes in activity data over multiple scales and apply this technique to subjects with one or more behavioral disorders, or psychiatric illnesses, to provide an automated method for classifying subjects or detecting changes in disease severity. The present systems and methods can, for example, acquire activity data of a subject and identify irregularity in the activity data. The irregularity in the data may involve changes in complexity of the data, and the metrics may involve multi-scale metrics. By multi-scale metrics, we mean any metric which quantifies an individual's activity over multiple time scales. It typically, though not necessarily, involves multiple metrics evaluated over different window sizes and then fused together into one number (such as probability of a given diagnosis) using for example, a machine learning approach. For example, we can measure entropy over several time scales, select the most useful scales, and then learn how to weight each of them together (using for example a Support Vector Machine (SVM), a linear classifier or a multivariate classifier) to classify entropies. The present systems and methods can be effective for extracting irregularity in the activity data and classifying the data representative of various disorders of the subject and the severity of the disorder(s).

In an embodiment, for example, a method is provided including: a) acquiring activity data of a subject from a device for monitoring the subject; b) pre-processing the data; c) identifying irregular pre-processed data and extracting features from the irregular data relevant to the health of the subject; and d) classifying the extracted features by mapping the extracted features onto categories, for example behavioural categories, representative of the disorder in the health of the subject.

In an embodiment, a system is provided comprising: system at least one computing device; and at least one application executable in the computing device, the at least one application comprising: a) logic that receives activity data of a subject from a device for monitoring the subject; b) pre-processes the data; c) analyses the pre-processed data and extracts features from the irregular data relevant to the health of the subject; and d) classifies the extracted features by mapping the extracted features onto categories, for example behavioural categories, representative of the disorder in the health of the subject.

In one or more embodiments, the data may be related to and/or include physical and/or physiological data. The data may be pre-processed to remove outlier data and/or normalise data. The data may be identified for relevant behavioural features, including for example irregularity in activity data. The features extracted may be behavioural features, for example behavioural complexity related features. Classifying the extracted features may include classifying the severity of a disorder.

In one or more embodiments a device may be provided for monitoring physical or physiological activity of the subject from which the activity data is acquired. The device for acquiring activity data may be selected from one or more of an acceleration or motion measuring device, an audio microphone, a video recorder, a finger photoplethysmogram, a pulse oximeter, a nasal cannula, a nasal cannula in conjunction with a microphone, an electrocardiogram (ECG), a smart mobile device, or a Hotler-type recorder. The device for acquiring the data may be, or also include, a mobile smart device including one or more sensors selected from the group of an accelerometer, a gyroscope, a microphone and a camera.

In one or more of the embodiments, a novel data analysis technique, Multi-scale Entropy (MSE), is applied to activity data, for example accelerometer data, for the purpose of extracting behavioral complexity related features. In one or more embodiments a subject is monitored for activity using video equipment, video data is recorded, filtered and a classifier is used to distinguish subjects with OSA syndrome (OSAS) from those without OSAS. The classifier can be an automated classifier using multi-scale metrics derived from the video data.

In one or more embodiments, a new approach to classifying subjects is provided which uses one or more of a Support Vector Machine (SVM), a linear classifier, or a multivariate classifier (such as a Naive Bayes Neural Networks or Random Forests) to classify entropies extracted from activity data, for example motion detected in videos over varying time scales.

In any one or more embodiments the present systems and methods use Multi-scale Entropy (MSE) to analyze audio, accelerometer and/or video data acquired during sleep and to distinguish between apneic subjects and benign snorers. The entropy values can be classified using, for example, a Support Vector Machine (SVM) a linear classifier, or a multivariate classifier which can provide similar or superior results to Linear Predictive Coding (LPC) and cepstral analysis without the need to detect specific events as MSE is applied to the entire recording.

Exemplary activity data that can be monitored can include one or more of acceleration, motion, heart rate, respiration rate, blood oxygen saturation, sweat, temperature, skin resistance, posture, location, audio, noise (electrical or acoustic), video and use of a monitoring device (such as text messaging, making a phone call, web surfing, or turning the device off or on). Exemplary irregularities in the activity that can be identified can include one or more of an irregularity in circadian rhythm, heart rate, skin conductance, pulse oximetry, pulse rate, respiration rate, sound, posture, motion, acceleration, location (via GPS, WiFi, etc), or other behavioural measures (such as text messaging, making a phone call, web surfing, or turning a device off or on). Exemplary disorders can include one or more of sleep disorders (such as apnea), bipolar disorders, depression, schizophrenia, affective disorders, cardiovascular disease, chronic obstructive pulmonary disease, diabetes, other neurological disorders such as dementia or Parkinson's disease, or deterioration during post-operative recovery or post discharge from hospital.

The systems, devices, methods, features, and advantages will be or become apparent to one with skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features and advantages be included within this description, be within the scope of the present disclosure, and be protected by the accompanying claims. Brief Description of the Drawings

Many aspects of the present disclosure can be better understood with reference to the following drawings. The components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present invention. Moreover, in the drawings, like reference numerals designate corresponding parts throughout the several views.

FIG. 1 depicts a flow chart in which various exemplary aspects of the present systems and methods are provided.

FIG. 2 depicts lengths of activity recordings for subject categories with median lines of the distributions.

FIG. 3 illustrates a block diagram describing an embodiment of a mobile phone application system of the present disclosure.

FIG. 4 is a boxplot distribution of Multi-scale Entropy with m = 2 and r = 0:2 on scales from 1 to 10 with record length of 10 days, taken from logarithm of activity.

FIG. 5 depicts Multi-scale Entropy of a single subject with m = 2 and r = 0:2 on scales from 1 to 400 with record length of 10, 20, 30, 40 days, taken from logarithm of activity.

FIG. 6 depicts smoothed Multi-scale Entropy of a single subject with m = 2 and r = 0:2 on scales from 1 to 400, taken from logarithm of activity, smoothing performed over 5, 9 and 15 points (10, 18 and 30 minutes respectively).

FIG. 7 depicts Multi-scale Entropy of a single subject with m = 2 and r = 0:2 on scales from 1 to 400, with record length of 10 days taken from logarithm of activity, starting at 1 st, 10th, 20th and 30th day.

FIG. 8 shows the non-parametric distributions of the standard activity metrics, using mean (for 10 days) of 15 MSE scales calculated on logarithm of most active 10 hours (L5) data.

FIG. 9 illustrates distribution of normalized activity metrics with record length of 10 Days, the distributions presented in the following order - schizophrenia (red), unemployed controls (green), affective disorder high risk (cyan), affective disorder low risk (magenta) and affective disorder unknown risk (blue), and the boxes representing 25th and 75th percentiles of the distribution and outliers are displayed as circles.

FIG. 10 illustrates separation between the two OSA and non-OSA classes using three of the best sample entropy {SampEn) features (Table II), the green triangles representing OSA patients and red circles representing non-OSA patients, the patients corresponding to the OSA data points which are close to the non-OSA group diagnosed with minor supine OSA (marked with blue arrows in the figure). FIG. 11 illustrates boxplots of SampEn values of m = 2, r = 0:15, and scale factors

(T)

of 1 to 30, with the blue boxes belonging to non-OSA patients and the red boxes to the OSA patients.

FIG. 12 illustrates selecting the best λ (m and r pairs), and T parameters for Classification, the image showing the Δ₀ values calculated based on the boxplot method explained in Section III-C4, Appendix B, with higher A_Q values indicating less overlap between the distributions and their corresponding features are more suitable to be used for classification.

FIG. 13 illustrates selecting the best λ (m and r pairs), and T parameters for Classification, the image showing the A_B values calculated based on the Bhattacharyya distance method explained in Section III-C4, Appendix B, with higher values indicating

less overlap between the distributions and their corresponding features are more suitable to be used for classification.

FIG. 14 depicts the results of upper body segmentation method on a frame of a video recorded at the IB E sleep lab.

FIG. 15 depicts boxplots of sample entropy distributions of OSA (red boxes) and non- OSA (black boxes) patients.

FIG. 16 (a)-(b) depict (a) an OSA severity score based on probability estimates of an SV classifier vs. patients grouped based on the OSA severity category assigned by clinitians.

FIG. 17 depicts the mapping of different sleeping positions using pitch and roll angles.

FIG. 18 is a block diagram of an embodiment of a system for determining the mental and physical health of a subject using multi-scale metrics.

FIG. 19 is a block diagram of an embodiment of a processing system shown in FIG.

18.

Detailed Description

Described below are various embodiments of the present systems and methods for determining mental and physical health, in particular systems and methods for collecting data representative of mental and physical health and classifying the data using a multi- scale metrics method to determine mental and physical health of a subject are detailed. Although particular embodiments are described, those embodiments are mere exemplary implementations of the system and method. One skilled in the art will recognize other embodiments are possible. All such embodiments are intended to fall within the scope of this disclosure. Moreover, all references cited herein are intended to be and are hereby incorporated by reference into this disclosure as if fully set forth herein.

We have discovered that one or a plurality of systems and methods can be employed for collecting and analyzing various data representative of mental and/or physical health. For example, physical and/or physiological activity of a subject in the context of one or more disorders can be used for diagnosis and evaluation of symptoms when certain features are extracted. The present systems and methods can be particularly effective in collecting and classifying data representative of behavioral disorders, such as sleep-related disorders (for example obstructive sleep apnea), bipolar disorders, depression, schizophrenia, affective disorders, cardiovascular disease, chronic obstructive pulmonary disease, diabetes, other neurological disorders such as dementia or Parkinson's disease, or deterioration during post-operative recovery or post discharge from hospital.

With reference to FIG. 1 , a flow chart is depicted in which various exemplary aspects of the present systems and methods are provided. In this embodiment, a system or method 100 is provided involving a device for monitoring a subject 102. The subject may be monitored for physical and/or physiological activity, for example. Through use of this device, data such as activity data of a subject, can be acquired 104. The data may be pre- processed 106. The pre-processing may be used to reduce outliers and/or normalise data. An example of pre-processing of the data includes filtering the data to remove un-needed or un-wanted data, including for example noise. The pre-processed data can be analyzed to identify irregular data, such as irregular activity data and features of the data related to the health of the subject extracted 108. These features can include features related to the health of the subject. As an example, the features can include behavioral features representative of a disorder in the health of the subject. The extracted features can then be classified 110 onto categories representative of the health of the subject, for example of a disorder in the health of the subject and the severity of the disorder. Various aspects of the present disclosure are discussed in more detail below.

In an embodiment an activity, such as physical and/or physiological activity, is measured or monitored, and collected or acquired. For example, an actigraphic assessment of a subject can be carried out by using an acceleration or motion measuring device. In another assessment, subject/patient behavior can be recorded using an audio microphone, and in another video recordings can be acquired. Other monitoring devices that can be used include a finger photoplethysmogram or pulse oximeter (to monitor for example blood oxygen saturation), a nasal cannula (to monitor for example nasal airflow), a nasal cannula in conjunction with a microphone (to monitor for example nasal sound), an electrocardiogram (ECG) (to measure heart activity), a smart mobile device, or a Hotler-type recorder (to allow assessment of data offline afterwards). The acquired data can, thus, include activity data, such as one or more of motion, heart rate, sweat, temperature, skin resistance, posture, location, audio, noise (electrical or acoustic), video and use of a monitoring device (such as text messaging, making a phone call, web surfing, or turning the device off or on).

As a non-limiting example, subjects with obstructive sleep apnoea (OSA) syndrome experience repeated periods of apnoea/hypopnea and arousals. Thus, in an embodiment a technique based on motion estimation in videos is used to characterize the syndrome in a subject. Alternatively, or in addition, audio data during sleep can be collected and used to distinguish between an apnoeic subject and a benign snorer. We have also developed similar approaches applied to mental health data, for example in schizophrenia, where changes in circadian rhythms, as well as very short term behavioral changes can be detected using similar sensors to those described above.

Signal processing methods can then be used for pre-processing of the data and extracting features. In one or more aspects, the data can be pre-processed to remove outlier data and/or normalize data. As an example, we have developed systems and methods used in conjunction with monitoring of subject activity and collection of activity data to filter the data. The data can be filtered using one or more signal processing techniques. For example, in the case of actigraphic assessment of patients using an acceleration measuring device the data can be processed in order to improve accuracy of results. (See, Section lll-A, Appendix A). In an aspect, one or more of the following rules can be applied:

1 ) All records can be re-sampled for 2 minutes epoch according to the Actiwatch-L User Manual ["The actiwatch user manual v 7.2," 2008. [Online]. Available:

http://www.camntech.eom/files The_Actiwatch_User_Manual_V7.2.pdf].

2) First and last days of every recording may be excluded as transition periods.

3) Days with activity level below 50% of the mean for this particular recording may also excluded. These days include significant period of watch-off and could not be used in analysis since they are unrepresentative.

4) Recordings with less than 10 days may be excluded. As can be seen from FIG. 2, certain categories of subjects, for example AFD high and low patients, had a median of recording length equal to 10 days and increase of the recording length would lead to under sampling of these categories.

5) Only the first 10 days of every recording may be included into analysis (except the first and the last day).

6) Periods of missing data (watch off) may be ignored (taken as no activity). These periods usually represent slower time and this activity pattern is unknown, so missing data is substituted with a known value to make sure that it will contribute to MSE patterns matching on higher scales. As another example, in the case of video sleep clinic data, since the smallest body movements during sleep are normally associated with respiratory movements and the respiratory rate for adults at rest averages 12-18 breath per minute, we may down sample to a rate of, for example, 1 Hz which is enough for taking respiration movements into accounts. Moreover, down sampling extracted video motion signals to 1 Hz allows us to match the frequency of high frequency actigraphic recordings making the two signals easier to monitor and compare. Activity can then be detected through complex motion detection, or simply by calculating the average pixel difference between each own-sampled frame. Additionally, one may elect to track or detect only the upper body and remove background clutter, providing a more specific locus of points related to the activity, and calculate the changes in the segments.

We can then identify and extract data relevant to the behavioral disorder(s) of interest. For example, we can identify irregularity in the acquired data. The irregularity may be one or more of an irregularity in circadian rhythm, heart rate variability, skin conductance responses, pulse oximetry, pulse rate, respiration, sound, posture, motion, acceleration, location (via GPS, WiFi, etc.) or other behavioural measures (such as text messaging, making a phone call, web surfing, or turning a device off or on). We can then analyze the identified irregular data with the purpose of extracting behavioral complexity related features indicative of one or more behavioral disorders. In an embodiment multi-scale metrics can be applied to the data for this purpose.

Traditional entropy measures analyze the predictability of a time series using a single scale. If complexity is defined to be associated with "meaningful structural richness", showing the correlations over different spatio-temporal scales, the regularity of a time series has no clear link to the complexity of that time series. For example, the entropy measurement of a randomized time series is higher than the time series it was originated from as the complexity and correlation of the signal is destroyed by a tempral shuffling process [M. Costa, A. Goldberger, and C. Peng, "Multiscale entropy analysis of complex physiologic time series," Physical Review Letters, vol. 89, no. 6, p. 68102, 2002].

Multi-scale entropy (MSE) was introduced as a method of measuring the complexity of physiologic or physical finite length time series. The aim of MSE is to take into account the multiple time scales inherent in such time series. The algorithm can include two steps. The first step is called the "coarse-graining" process in which the time series ... , x_N] is divided into non-overlapping windows of length τ. The values within each window are averaged to construct a new coarse-grained time series y ^w}:

If r=1 , the coarse-grained time series would be the original time series.

In the second step, entropy is measured for each coarse- grained time series corresponding to the scale factor r. MSE can be calculated using different measures of entropy ["Multiscale entropy analysis of biological signals," Physical Review E, vol. 71 , no. 2, p. 021906, 2005]. In this research sample entropy proposed by Richman and Moorman ["Physiological time-series analysis using approximate entropy and sample entropy," American Journal of Physiology-Heart and Circulatory Physiology, vol. 278, no. 6, p.

H2039, 2000] was used as a metric of entropy. Sample entropy takes two parameters, the pattern length m, and the similarity criterion r. The pattern length is effectively the length of the 'word' which is compared to other 'words' or segments in the time series and r, a positive real value and is usually chosen between 10% to 20% of the standard deviation of the time series, effectively defines the quantization or dictionary size through which the comparison is made. In this embodiment the data is normalized to zero mean with a standard deviation of 1. It should be noted that sample entropy is a modification to the approximate entropy measure by Pincus ["Approximate entropy as a measure of system complexity" Proceedings of the National Academy of Sciences, vol. 88, no. 6, p. 2297, 1991] but is less dependent on the time series length and demonstrates relative consistency over a broader range of r, m, and N values.

Two patterns of length m match if every point in the first pattern is within distance r from the corresponding point in the second pattern. Here, the distance between two vectors is defined as the maximum absolute difference between the components of those two vectors. Let's define vector to be a pattern of

length m (i can only go up to N - m to ensure that vector x_m+1 of length m+1 is also defined). is defined as the number of vectors x_m(j) that have a distance smaller than r with xm (i), i ≠ ;^'to exclude the self matches. is the probability

that the distance between vector xm (i) and any other vector xm j) is smaller than r. U^m{r) is the probability of any two vectors of length m being within distance r of each other. The probability U^m (r) would therefore be

Richman and Moorman define sample entropy

which is estimated by the statistic

The MSE analysis can then be performed over multiple scale factors, r {e. g. between 1 and 60). (For each successive scale factor, pairs of points are averaged together).

Entropy can be measured, for example, for all scale factors and for pattern lengths m, m e M= {2, 3, 4, 5, 6} and similarity criteria r, r e R = {0.1 , 0.15, 0.2, 0.25}. For simplicity we define parameter λ for all combinations of pairs of m and r as λ_; = {m^ = M(t/5), η = mod 4)}, 1 < i < 20. A total of 1200 entropy values can therefore be calculated for each time series.

Features indicative of various disorders can then be extracted from the calculated entropy values. For example, in an embodiment involving actigraphic assessment of patients using an acceleration device, the calculated entropies for each pair of pattern length and similarity criterion can be visually analyzed using boxplots. Boxplots are graphical representations of groups of numerical data summarizing the information about each group in five numbers, (i) group's minimum (lower whisker), (ii) group's lower quartile (Q1 ), (iii) group's median, (iv) group's upper quartile (Q3), and (v) group's maximum (upper whisker). A quartile is one of three points that divides a data set into four equal segments. If any of the data points are considered outliers they will be presented as separate points outside the box in the plot. The set of parameters that give the best separation between the two classes of OSA and non-OSA may be selected by subtracting the distance between the 25^th percentile of OSA boxes and 75^,h percentile of the non-OSA boxes added to the difference between the lower whisker of OSA boxes with the upper whisker of that of non-OSA. The sum of the difference between these values can be used as an estimate of the separation between the two groups.

The Bhattacharyya Distance is a measure of similarity of two continuous or discrete probability distributions and in classification it is used as a measure of separability between two classes. This distance is calculated from the Bhattacharyya coefficient p which is a measure of the amount of overlap between two probability distributions. For discrete distributions p and q over the domain X the Bhattacharyya coefficient is defined as

with 0 < p < 1. The Bhattacharyya distance A_B is then defined as

so that [T.Kailath, "The divergence and Bhattacharyya distance measures

in signal selection," Communication Technology, IEEE Transactions on, vol. 15, no. 1 , pp. 52-60, 1967]. Low values of p or high values of indicate little overlap between p and q.

The combinations of MSE parameters can be analysed and ranked based on the distance between their corresponding sample entropy distributions. Larger Bhattacharyya distance between the two distributions can indicate that the sample entropy values were a better feature candidate for the classification.

Next the extracted features are classified. In particular, one or more classifiers are used to map the extracted features onto subject categories, such as schizophrenia, and affective disorders, or other disorders mentioned herein, for determining mental and physical health.

For example, selection of pattern length m and similarity threshold r can have significant influence on detection of patterns during sample entropy (SampEn) calculation. In an embodiment the following approach for MSE-based classification can be taken:

1 ) For each of the possibilities of MSE calculation, perform preliminary classification using a range of m and r coefficients.

2) Select an approach of MSE calculation, m and r values with the best results according to the preliminary classification.

3) Improve classification using feature selection, for example minimum Redundancy Maximum Relevance (mRMR) algorithm or a Genetic Algorithm.

In order to distinguish between subject categories, traditional activity metrics (M10, LS, RA, IS, IV) and MSE coefficients can be used. To select the best features, classification can be performed on different feature subsets, including manually and automatically selected.

In an embodiment to perform multivariate classification, a Naive Bayes method can be used according to the equation

For binary classification a Support Vector Machine (SVM) method can be used with Gaussian Radial Basis Function (RBF) kernel. SVM classifier attempts to create a hyperplane with a largest distance to nearest points in a feature space to separate target classes. If linear separation in the original feature space is not possible, features can be mapped into higher dimensional space using kernel function, before separation is performed.

To improve feature selection for classification, minimal- redundancy-maximal- relevance (m RMR) criterion can be used to identify the most relevant features. The mRMR method selects features with maximum mutual information with target classes (relevance) and minimal mutual information between features (redundancy) (see Peng et al. [C. Ding and H. Peng, "Minimum redundancy feature selection from microarray gene expression data," in Computational Systems Bioinformatics Conference, International IEEE Computer Society. Los Alamitos, CA, USA: IEEE Computer Society, 2003, p. 523], [H. Peng, F. Long, and C. Ding, "Feature selection based on mutual information: Criteria of Max-Dependency, Max-Relevance, and Min-Redundancy," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226-1238, 2005]) according to the equations 9, 10, 1 1.

where l{x; y) is the mutual information between variables x and y and p(x), p(y) and p(x,y) are probabilistic densities of these variables.

where D is relevance criteria, S is a feature set with features x_t and c is a target class.

where R is redundancy, 5 is a feature set with features x_iand xj . Incremental search can be used to find features which satisfy the above criteria.

In yet a further embodiment, as exemplified in Example 3 below, the present systems and methods can be carried out using a mobile smart device and a mobile network. See, J. Daly et al., A neonatal apnoea monitor for resource-constrained environments. Comp. Cardiology 2012:39-321 -324. (incorporated by reference as of fully set forth herein). The term "telemedicine" refers to the use of Information and

Communication Technologies (ICTs) in order to overcome geographical barriers and increase access to healthcare services. mHealth or Mobile Health - a form of telemedicine - is a term used for the practice of healthcare supported by mobile smart devices such as mobile phones and other wireless computing devices that allow the delivery of healthcare at distance.

A smartphone is a mobile phone offering advanced capabilities. It offers PC-like functionality such as: internet access, e-mail, video cameras, MP3 player, video viewing etc. Smartphones run complex operating systems (OS) that provide a platform for application developers. There are a number of different OS available, the main ones are: Symbian, Android (GNU/Linux based), RIM, iPhone and Microsoft. In one embodiment, because of its open source characteristic an Android based phone may be used.

In an embodiment a mobile phone application can record any one or more of accelerometry, audio, body movement, and body position and use those signals (extract features) together with answers from a questionnaire in order to give a probability for the subject to belong to one of the three studied groups (healthy/snorer/OSA) and advise him/her on what action should be taken. FIG. 3 describes one embodiment of a phone application system; different signals are recorded by the phone and features are extracted from those signals. Using those features the subjects can then be classified into one of, for example, the three studied groups: healthy, snorer, OSA. The final classification is the result of the combination of the independent classification from the different signals and answers to the questionnaire (data fusion). Body position can be used for eventually advising the subject to change his/her sleeping position habit or for more specific diagnoses.

Methods and Examples

Example 1

In one example, for circadian pattern estimation, scales of 360 ± 30 (12 ± 1 hours) were used. In this study the signal length is limited to 10 days with 2 minutes activity epochs, i.e. 7200 data points. This allows reliable use of MSE scales from 1 to 10. For other scales, only very well structured activity patterns can be detected reliably, such as circadian rhythms.

Analyzing the first scales in more detail using boxplots (FIG. 4), it can be seen that on the scale 1 schizophrenic subjects are significantly different from the other groups, especially the unemployed control group. However, at higher scales the distributions of these two classes overlap.

For illustration of extremely long-term activity, MSE analysis was applied to a single subject from our sample, where the data were collected continuously for 40 days, so the effect of increased record length on the reliability of Samp En calculation is visible on higher scales.

FIG. 5 illustrates that for 10 days record length the plot of sample entropy (R_SE) values over different scales includes significant variations, making possible only reliable estimation of 12 hour rhythmicity pattern. An increase in the record length allows us to distinguish an 8 hour pattern in a 40 day long record.

FIG. 6 demonstrates median smoothing of an MSE plot taken for a 10 days record with 5, 9 and 15 points window. Key circadian patterns of 12 and 8 hours can be observed on all plots.

FIG. 7 presents MSE plots taken over four different non-overlapping 10 day periods from a single 43 days recording. We can clearly see significant differences between these periods.

Potential classification accuracy of all approaches for MSE calculation was analyzed using all scales as features and for the ranges of m = [1..7] and r =

[0.1..0.7]. Classification was performed using Naive Bayes method that fits multivariate normal densities with a diagonal covariance matrix estimate ('diagquadratic' classifier of Matlab). Training and testing were performed on all points of input data to find the best values of m and r, where feature distributions are the most different between samples and can be separated with Naive Bayes method.

Naive Bayes classification assumes that all features are independent within each class. This assumption is not true in our case, but these classifiers are known to work well even when independence assumption is not valid and thus applicable for estimation of possible classification accuracy.

FIG. 8 shows the non-parametric distributions of the standard activity metrics. Note that although no single metric provides separation between classes, the use of all of these (plus SampEn) increases class separation.

Activity metrics, as we can see from the visual inspection of the FIG. 9, can be used as rather strong predictors for classification.

Example 2

In this study the performance of the proposed OSA classifiers is assessed on clinical data. A SVM was used to classify the OSA patients from non-OSA patients and the results were compared to the diagnosis made by Respiratory Medicine physicians. FIG. 10 shows the separation between the two classes of data (OSA vs. non-OSA) using three features from the calculated sample entropy (fl_SE) values. The plots in the results section that use scale factors between 1 to 30 as the best R_SE candidates were found for scale factors in this region and also for better visualization.

The separation between the distributions of non-OSA and OSA patients was analyzed using boxplots (FIG. 11) and the Bhattacharyya Distance A_B (see also, Section III- C, Appendix B). The measurements of the separation between the classes are

demonstrated in FIGS. 12 and 13. A summary of the sample entropy parameters providing the best separation between the two OSA and non-OSA groups is presented in Tables I and II. In order to find the features providing the maximum classification accuracy from the ones listed in Tables I and II, all combinations of k from total number of features were used to train and validate the classifier ((£), k≤n = 10).

As the SVM classifier with an RBF kernel over-performed the linear SVM, the former was chosen as the candidate classifier (See Section lll-D, Appendix B). For each combination of features, RBF σ values of 1 to 10 were tested and σ = 1 provided the best classification results. The results indicated that a maximum classification accuracy of 90% is possible over several combinations of features. Table III summarizes the best classification results for the Boxplot and Bhattacharyya Distance feature selection methods (see Section III-C4, Appendix B) and over different (£) combinations of scale factors taken from Tables I and II.

FIG. 13 demonstrates the Bhattacharyya Distance between the R_SE (SampEn) values for the OSA and non-OSA patients.

An upper body segmentation technique was explored on videos recorded during sleep monitoring. The technique successfully segments head and torso in almost all frames with the patient's back view and most frames with the frontal view. However, the method fails for some other viewing angles particularly the ones more than 30° away from the frontal view. FIG. 14 shows an example of a segmented area in one of the frames. This illustrates that more body-part specific metrics can be derived from the data.

The conventional methods of studying sleep for detection of sleep disorders involve analyzing multiple physiologic signals recorded over a long period of time (usually overnight). Sleep specialists spend hours reviewing such studies trying to find correlations between different signals such as pulse oximetry, sound, and a patient's movements.

Moreover, recording of all PSG signals involves connecting various electrodes and chest straps to the patient. The design of a robust automated system for the detection of sleep disorders from the video recording of sleep may make a significant difference in the efficiency of analyses of sleep studies. Additionally, this automated system with fewer or no electrodes attached to the patient can greatly increase the convenience and create a more natural sleep environment. In comparison with the movement signal from an actimeter, motion detected from videos includes more detailed movements from all body parts whereas actimeters tend to be more sensitive to some positions and movement directions [M. Elbaz, G. Roue, F. Lofaso, and M. Salva, "Utility of actigraphy in the diagnosis of obstructive sleep apnea," Sleep, vol. 25, no. 5, pp. 527-531 , 2002].

The results of classification of OSA using PSG videos alone are explained in Section IV, Appendix B. A good separation between the two classes of OSA and non-OSA patients was achieved using sample entropy values of specific scale factors (τ). This is visually demonstrated in FIGS. 10 and 1 1. The SVM classifier achieved a maximum accuracy of 90% on both the training data and the validation data. Out of scale factors in the range of 1 to 60, the ones between 10 to 30 seem to give the best separation between the two classes. Since the OSA episodes happen repeatedly and last for more than 10 seconds, the 10 to 30 scale factors can be related to the most common durations of OSA events (data sampling rate after filtering was = 1 Hz).

Out of the ten OSA patients two had sample entropy values close to non-OSA patients (illustrated in FIG. 10) and were therefore more difficult to classify. Interestingly, these patients were diagnosed with minor supine OSA referring to a minor OSA condition that only happens when the person is in the supine position. This indicates that it is possible to detect the severity of OSA using a probabilistic classifier.

The results of upper body segmentation of IBME normal data indicate successful segmentation of head and torso in almost all frames with the patient's back view and most frames with the frontal view. However, to cover more viewing angles improvements are required. Promising upper body segmentation leads to the possibility of extracting more features such as respiratory related features from the videos.

Example 3

In an embodiment, we applied our present system and process for the detection of obstructive sleep apnea (OSA). See, E. Gederi & G. D. Clifford, Fusion of Image and Signal Processing for the Detection of Obstructive Sleep Apnea, IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI2012), pp 890-893, which is expressly incorporated by reference as if fully set forth herein. For the classification of OSA and non-OSA patients, a support vector machine (SVM) with a radial basis function (RBF) kernel was used. C. Cortes and V. Vapnik, "Support-vector networks," Machine learning, vol. 20, no. 3, pp. 273-297, 1995. The features used in the classification were the sample entropy values that were calculated for each patient's motion signal calculated. The kernel parameter σ was set to 0.5, and the cost parameter C was set to 2. The data was divided into training and validation sets using the k-fold cross-validation technique with k = 5.

Therefore, in each of five iterations, a different subset of the data was used for the validation and the remaining four were used for training the classifier. The mean of the classification performance of all iterations on the validation data was calculated as the overall performance of the classifier.

The probability estimates of the SVM classifier were calculated as described in H. Lin, C. Lin, and R. Weng, "A note on platts probabilistic outputs for support vector machines," Machine Learning, vol. 68, no. 3, pp. 267-276, 2007, to generate an OSA severity score. Since the value of AHI was not available for all the patients, the SVM severity scores were compared with the oxygen desaturation index (ODI) defined as the number of > 4% blood oxygen desaturations per hour of sleep. ODI has a high correlation with apnea of more than 10 seconds. J. Stradling and J. Crosby, "Predictors and prevalence of obstructive sleep apnoea and snoring in 1001 middle aged men." Thorax, vol. 46, no. 2, p. 85, 1991. The values of ODI were automatically generated by the PSG system. As one patient diagnosed with OSAS had no OSA severity category (mild, moderate, severe) assigned by the clinicians, the patient was excluded from the study. From the remaining 10 OSA patients, two were diagnosed with mild supine OSA, three with moderate OSA, and five with severe OSA.

Nine of the 1200 sets of calculated sample entropy values which provided the best separation between the OSA and non-OSA patients were selected using the Bhattacharyya distance. An overview of the overlap between the OSA and non-OSA sample entropy distributions for m = 3, r = 0.2 and r e{1 30} is demonstrated in FIG. 15 using boxplots.

In order to find the features with the maximum classification accuracy from these nine sets, all combinations of c from total number of features were used to train and validate the classifier

. The results indicated that a maximum classification accuracy of 90% is

possible over different combinations of less than five features (Table IV). An example list of features (sample entropy parameters) providing the highest classification accuracy is presented in Table V, however, it was possible to obtain the same results using other combinations of features. FIG. 16 gives a visual representation of the detected OSA severity scores and the automatically generated ODI against the clinically diagnosed OSA categories.

The promising separation between the two groups of OSA and non-OSA patients using sample entropy values of specific scale factors ( τ ) resulted in a maximum

classification accuracy of 90% on both training and validation data. As it is illustrated in FIG. 16, the detected OSA severities showed a distinct separation between patients with moderate to severe OSA and patients without OSA. Additionally, for patients with moderate to severe OSA, the detected severity score was a better indication of the level of OSA comparing to the ODI. However, the two patients with minor supine OSA who were misclassified lie close to the non-OSA patients. Taking the patient's sleeping position into consideration can resolve this type of misclassification. In general, a combination of other apnea related features and physiological signals with signals derived from the video recordings may enhance the classification performance.

Example 4

This example involved an embodiment of a mobile phone application of our present disclosure. J. Behar et al. An automated OSA screening application for smart phones. Submitted to IEEE, Transactions on Information Technology in Biomedicine (attached hereto as Appendix E).

A. Choice of the hardware and general set up

There are a number of different mobile phone Operating Systems (OS) available, with the main ones being: Symbian, Android, RIM, iOS and Windows 7 mobile (and Windows Phone 8). An application was developed under the Android OS (V 2.2 on a HTC Wildfire) because of its open source licensing and recent exponential growth [Canalys, "Google's Android becomes the world's leading smart phone platform."

http://www.canalys.com, January 2011 ].

The audio file recorded by the mobile phone should be of the best possible quality in order to preserve all the features of the signal with their potential associated diagnostic information. There are multiple parameters in the audio acquisition workflow that have an impact on the audio quality: frequency response of the phone's audio card, audio media format (encoder), the type of headset, the location of the headset.

The sound card (audio analog-to-digital converter) of the phone preferably will have a relatively flat frequency response in the frequency band of interest (20Hz - 5.5kHz) so that little distortion appears in the record. GSMArena provides audio quality data on mobile phones.

B. Location of phone during sleep

Where on the body the phone should be worn can be a function of various criteria, for example: 1 ) Where is the best actigraphic signal? 2) Where is it best to record body position? 3) How convenient is it for the patient? Additionally we might want to alter the location of the phone to suit the microphone placement.

There are a number of possibilities for placing the mobile phone on- or off-body. Placing the phone on the arm, especially the front of the arm, may be the most comfortable location for the patients. However, according to Ozeke et al. [O. Ozeke, O. Erturk, M.

Gungor, S. Hizel, D. Aydin, M. Celenk, H. Dincer, G. Ilicin, F. Ozgen, and C. Ozer, "Influence of the right-versus left-sided sleeping position on the apnea-hypopnea index in patients with sleep apnea," Sleep and Breathing, vol. 16, no. 3, pp. 1-4, 2011], the left side sleeping position had a statistically higher apnoea hypopnea index (AHI) score than the right side sleeping position. Therefore the arm on which the phone is placed should be taken into account. Positioning the phone on the chest can record both gross body movements as well as respiratory movements. As the highest AHI usually occurs when the patient is sleeping in the trunk supine position [E. Van Kesteren, J. van Maanen, A. Hilgevoord, D. Laman, and N. de Vries, "Quantitative effects of trunk and head position on the apnea hypopnea index in obstructive sleep apnea," Sleep vol. 34, no. 8, pp. 1075-81 , 2011], recording the actigraphy of the chest may lead to a better understanding of when apnoeas and hypopnoeas occur. Alternatively the phone can be placed on the head, as this would record both head movement and head position. This may be useful as the torso may be supine but the head may be right/left. Van Kesteren et al. [E. Van Kesteren, J. van Maanen, A. Hilgevoord, D. Laman, and N. de Vries, "Quantitative effects of trunk and head position on the apnea hypopnea index in obstructive sleep apnea," Sleep vol. 34, no. 8, pp. 1075-81 , 2011], suggested that two position sensors, one placed on the head and one on the trunk, should be considered for sleep recordings. The phone may be placed on the bed frame for general whole body actigraphy. It can even be placed remotely to video record the sleep and convert the movements in video data to actigraphy [E. Gederi and G.D. Clifford, "Fusion of image and signal processing for the detection of obstructive sleep apnea," in Proceedings of the IEEE=EMBS International Conference on Biomedical and Health Informatics (BHI2012), IEEE, Jan. 2012].

For this app, the arm position was chosen as it is the most comfortable position for the user.

C. Actigraphy assessment

Mobile phone accelerometers vary in their precision and sensitivity. The accelerometer measures the acceleration applied to the device

. Conceptually, this is done by measuring forces applied to the sensor · wnere N is tne

number of external forces applied to the system. Gravity always influences the measured acceleration; if we split the left hand side in two: This phenomenon

can be observed by leaving the phone on the desk

where the accelerometer gives a value of magnitude g=9.8ms^'2.

To access the proper acceleration of the device the contribution of the force of gravity needs to be removed. This can be achieved by applying a high-pass filter to the raw time series. Conversely a low pass filter can be used to isolate the force of gravity. A simple first order low pass filter has been implemented in Java for filtering the raw data on the phone. Below is the discrete model used for the filter:

Where x[i] is our input signal (total gravitation), y[i] is our output signal (gravity), a€ [0;l] is the filter rate and /^' the time index, a was set equal to 0.10. The previous equation can be rewritten in terms of acceleration (Acc) at time t:

where At corresponds to the time between two samples. To keep only the high frequencies of the signal:

D. Body position assessment

The Android Orientation Sensor' has been used for recording body position. The space can be mapped (see FIG. 17) into the different sleeping positions using pitch and roll angles (azimuth was not used). This figure assumes that the phone is worn on the left upper arm and is oriented parallel to the coronal plane.

E. Signal structure and feature extraction

MultiScale Entropy (MSE) analysis was used for determining signal structure and feature extraction. It involved a method of measuring the complexity of a finite length time series over different spatio-temporal scales [M. Costa, A. Goldberger, and C. Peng, "Multiscale entropy analysis of biological signals," Physical Review E, vol. 71 , no. 2, p. 021906, 2005]. As such, it provided an excellent method for describing the regularity in a time series and therefore is particularly useful for identifying the rhythms indicative of poor and healthy sleep. Sample entropy seeks matching patterns throughout a time series to calculate its degree of regularity. It takes two parameters for example; the pattern length m, and the similarity criterion r which is the tolerance for accepting pattern matches. Two patterns of length m can match if every point in the first pattern is within distance rfrom the corresponding point in the second pattern. The distance between two vectors is defined as the maximum absolute difference between the components of those two vectors.

F. Data fusion

Based on the work of Roebuck et al. [A. Roebuck and G.D. Clifford, "Multiscale entropy applied to audio data for classifying obstructive sleep apnoea patients," in Am Thor Soc 2012 Conference, May 2012], Gederi et a/.[E. Gederi and G.D. Clifford, "Fusion of image and signal processing for the detection of obstructive sleep apnea," in Proceedings of the IEEE=EMBS International Conference on Biomedical and Health Informatics (BHI2012), IEEE, Jan. 2012] and Higgins er a/. [N. Higgins, supervised by G.D. Clifford, "The Detection of Obstructive Sleep Apnoea Using a Mobile Phone," MSc. Thesis, University of Oxford, Trinity term 2012], sample entropy (calculated over varying time scale) was chosen as the features to quantify irregularity in the audio and actigraphy recordings, with extremes of entropy indicating pathological physiology. Once the features have been derived - from audio, actigraphy, body position and the questionnaire - the user's data are classified using a trained support vector machine (SVM). The implementation draws upon the multi-platform open-source software library LIBSVM [C. Chang and C. Lin, "LIBSVM: A library for support vector machines," ACM Transactions on Intelligent Systems and Technology, vol. 2, pp. 27:1-27:27, 2011. Software available at http://www.csie.ntu.edu.tw/cjlin/libsvm], as Android uses standard Java libraries. Development of the classifier took place in MATLAB (using LIBSVM) to generate a structure containing the SVM type, kernel type, number of classes, total number of support vectors, offset, class labels and the support vectors (see A. Roebuck and G.D. Clifford, "Multiscale entropy applied to audio data for classifying obstructive sleep apnoea patients," in Am Thor Soc 2012 Conference, May 2012, for details). This structure was then saved as a text file to make it readable by LIBSVM on the phone and allow forward propagation of the classifier. MSE was implemented using Physionet's Open-source C- version [M. Costa, "Multiscale entropy analysis." http://phvsionet.org/phsiotools/mse/, August 2004],

FIG. 18 illustrates a system 100 for carrying out the methods for determining mental health according to the present disclosure. As indicated in FIG. 18, the system 100 generally comprises a signal/data acquisition device or system 102 and a processing device or system 104 that are coupled such that data can be acquired and sent from the data acquisition system 102 to the processing system 104. The processing system 104 may comprise, for example, a hand-held device, a portable device, a computer, server, dedicated processing system, or other system, as can be appreciated. The hand-held device can be, for example, a smart mobile phone. The processing system 104 may include various input devices such as a keyboard, microphone, mouse, touch screen or other device, as can be appreciated. By way of example, the system 100 can comprise a stand-alone device or part of a network, such as a local area network (LAN) or wide area network (WAN).

The signal/data acquisition system 102 is configured to acquire data concerning the activity of a subject. The signal/data acquisition system 102 can include any one or more of the devices mentioned above for acquiring data on the subject, in particular activity data. For example, the device(s) can include any one or more of an acceleration measuring device, audio and/or video equipment for monitoring the activity of a subject, a pulse oximeter, and ECG or a smartphone or other smart mobile device. The signal/data acquisition system 102 can also include, for example, an oximeter such as a NellcorOxiMax N-600x Pulse Oximeter, and ECG recorder such as a Welch Allen CP-50 ECG recorder, a blood pressure monitor such as an Omron R6 Blood Pressure Monitor, and/or a respiration rate monitor such as a Masimo Rainbow Acoustic Respiration Rate (RRa) system.

As described below, the processing system 104, in particular the software provided on the processing system, is configured to receive the data acquired by the signal/data acquisition system 102 and evaluate that data to determine the acceptability of the physiological signals collected. Notably, although the signal/data acquisition system 102 and the processing system 104 are illustrated as separate components in FIG. 18, the two components and/or one or more of their respective functionalities can be integrated into a single system or device, if desired. FIG. 19 is a block diagram illustrating an architecture for the processing system 104 shown in Fig. 18 according to one or more embodiments herein. The processing system 104 of FIG. 19 can comprise a processor 200, memory 202, a user interface 204, and at least one I/O device 206, each of which is connected to a local interface 208. The local interface 208 may be, for example, a data bus with a control/address bus as can be appreciated.

The processor 200 can include a central processing unit (CPU) or a semiconductor- based microprocessor in the form of a microchip. The memory 202 can include any one of a combination of volatile memory elements (e.g., RAM) and nonvolatile memory elements (e.g., hard disk, ROM, tape, etc.).

The user interface 204 comprises the components with which a user interacts with the processing system 104 and therefore may comprise, for example, a keyboard, mouse, and a display, such as a liquid crystal display (LCD) monitor. The user interface can also comprise, for example, a touch screen that serves both input and output functions. The one or more I/O devices 206 are adapted to facilitate communications with other devices or systems and may include one or more communication components such as a

modulator/demodulator (e.g., modem), wireless (e.g., radio frequency (RF)) transceiver, network card, etc.

The memory 202 comprises various software programs including an operating system 210 and the present system 212 for collecting and recording data and filtering the signals/data acquired by the signal/data acquisition system 102, and processing the data to extract features associated with a subject's mental health and then identifying the most relevant features and classifying the extracted features according to the methods described herein. The operating system 210 controls the execution of these programs as well as other programs and provides scheduling, input-output control, file and data management, memory management, and communication control and related services.

The components stored in the memory 202 may be executable by the processor 200. In this respect, the term "executable" refers to a program file that is in a form that can ultimately be run by the processor 200. It can include any one or more of the conventional devices used to acquire, collect and store the data. Exemplary devices include executable programs that may be, for example, a compiled program that can be translated into machine code in a format that can be loaded into a random access portion of the memory 202 and run by the processor 200, or source code that may be expressed in proper format such as object code that is capable of being loaded into a random access portion of the memory 202 and executed by the processor 200, etc. An executable program may be stored in any portion or component of the memory 200 including, for example, random access memory, read-only memory, a hard drive, compact disk (CD), floppy disk, or other memory components.

The memory 202 is defined herein as both volatile and nonvolatile memory and data storage components. Volatile components are those that do not retain data values upon loss of power. Nonvolatile components are those that retain data upon a loss of power. Thus, the memory 202 may comprise, for example, random access memory (RAM), readonly memory (ROM), hard disk drives, floppy disks accessed via an associated floppy disk drive, compact discs accessed via a compact disc drive, magnetic tapes accessed via an appropriate tape drive, and/or other memory components, or a combination of any two or more of these memory components. In addition, the RAM may comprise, for example, static random access memory (SRAM), dynamic random access memory (DRAM), or magnetic random access memory (MRAM) and other such devices. The ROM may comprise, for example, a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or other like memory device.

In addition, the processor 200 may represent multiple processors and the memory 202 may represent multiple memories that operate in parallel. In such a case, the local interface 208 may be an appropriate network that facilitates communication between any two of the multiple processors, between any processor and any one of the memories, or between any two of the memories, etc. The processor 200 may be of electrical or optical construction, or of some other construction as can be appreciated by those with ordinary skill in the art.

The operating system 210 is executed to control the allocation and usage of hardware resources such as the memory, processing time and peripheral devices in the processing system 104. In this manner, the operating system 210 serves as the foundation on which applications depend as is generally known by those with ordinary skill in the art.

Various programs (i.e. logic) have been described herein. Those programs can be stored on any computer-readable medium for use by or in connection with any computer- related system or method. In the context of this document, a computer-readable medium is an electronic, magnetic, optical, or other physical device or means that contains or stores a computer program for use by or in connection with a computer-related system or method. Those programs can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer- based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions.

It should be emphasized that the above-described embodiments of the present systems and methods are merely possible examples of implementations set forth for a clear understanding of the principles thereof. Many variations and modifications may be made to the above-described embodiment(s) of the invention without departing substantially from the spirit and principles of the invention. For example, the present system and method can be applied to any combination of signals and signal quality measures or indeed to physiological parameters alone or signal quality measures alone. All such modifications and variations are intended to be included herein within the scope of this disclosure and the present invention and protected by the following claims.

Further, in other embodiments, more apnea related features can be evaluated and used for the classification. These features can include physiologic signals such as ECG and pulse oximetry. However, as the number of features increases, the role of a good feature selection technique becomes more important. Therefore, feature selection methods such as genetic algorithms and minimum redundancy maximum relevance (MRMR) may be used for feature classification.

Claims

What is claimed:

1. A method for classifying a disorder in the health of a subject, comprising: a) acquiring activity data of the subject from a device for monitoring the subject;

b) pre-processing the data;

c) identifying irregular pre-processed data and extracting features from the irregular data relevant to the health of the subject; and

d) classifying the extracted features by mapping the extracted features onto categories representative of the disorder in the health of the subject.

2. The method of claim 1 , wherein the step of classifying the extracted features includes use of at least one of a Support Vector Machine (SVM), a linear classifier, or a multi-variate classifier to classify entropies.

3. The method of claim 1 , wherein the activity data includes one or more of acceleration, motion, heart rate, respiration rate, blood oxygen saturation, sweat, temperature, skin resistance, posture, location, audio, noise, video, or use of a monitoring device.

4. The method of any of claims 1-3, wherein the irregular data includes one or more of an irregularity in physiological or physical activity including circadian rhythm, heart rate, skin conductance, pulse oximetry, pulse rate, respiration rate, sound, posture, motion, acceleration, location or other behavioural measures.

5. The method of any of claims 1 -4, wherein the device for acquiring activity data is selected from one or more of an acceleration or motion measuring device, an audio microphone, a video recorder, a finger photoplethysmogram, a pulse oximeter, a nasal cannula, a nasal cannula in conjunction with a microphone, an electrocardiogram (ECG), a smart mobile device, or a Hotler-type recorder.

6. The method of claim 1 , wherein the disorder includes one or more of sleep disorders, bipolar disorders, depression, schizophrenia, affective disorders, cardiovascular disease, chronic obstructive pulmonary disease, diabetes, other neurological disorders, or deterioration during post-operative recovery or post discharge from hospital.

7. The method of claim 1 , wherein identifying irregular data and extracting features from the irregular data includes use of multi-scale entropy.

8. The method of claim 1 , wherein the device for acquiring the data is a mobile smart device including one or more sensors selected from the group of an accelerometer, a gyroscope, a microphone and a camera.

9. The method of claim 1 , wherein identifying irregular data and extracting features from the irregular data includes use of multi-scale entropy to identify and extract behavioral complexity features and classifying the extracted features involves use of one or more of a Support Vector Machine (SVM), a linear classifier, or a multi-variate classifier to classify entropies .

10. The method of claim 9, wherein the disorder includes one or more of sleep disorders, bipolar disorders, depression, schizophrenia, affective disorders, cardiovascular disease, chronic obstructive pulmonary disease, diabetes, other neurological disorders, or deterioration during post-operative recovery or post discharge from hospital.

1 1. The method of claim 10, wherein the step of classifying the extracted features includes classifying the severity of a disorder.

12. The method of claim 10, wherein the activity data includes one or more of acceleration, motion, heart rate, respiration rate, blood oxygen saturation, sweat, temperature, skin resistance, posture, location, audio, noise, video or use of a monitoring device.

13. The method of claim 12, further including providing a device for monitoring activity of the subject from which the activity data is acquired.

14. The method of claim 1 , further including providing a device for monitoring physical or physiological activity of the subject from which the activity data is acquired.

15. A system for classifying a disorder in the health of a subject, comprising: at least one computing device; and

at least one application executable in the computing device, the at least one application comprising:

a) logic that receives activity data of the subject from a device for monitoring the subject;

b) pre-processes the data;

c) analyses the pre-processed data and extracts features from the irregular data relevant to the health of the subject; and

d) classifies the extracted features by mapping the extracted features onto categories representative of the disorder in the health of the subject.