METHOD OF MONITORING HEART RATE VARIABILITY AND THE USE OF THAT METHOD IN THE PREDICTION OF FALLS AND OTHER APPLICATIONS

The invention relates to methods involving the monitoring of a subject's Heart Rate Variability. The methods may be useful in obtaining an indication of the likelihood of a subject to have a fall, or in the prediction of a change in blood pressure, whether due to the subject standing up (Orthostatic Hypotension) or otherwise. The invention also relates to systems, apparatus and devices which embody this method.

Falls represent one of the most common problems of later life. The annual incidence ranges between 35-50% and increases with age, reaching 66% per year among the healthy elderly. Falls reduce overall well-being, mobility and quality of life of the elderly and of those who care for them in the family. The mean and median costs for a fall are about EUR 9,000 and EUR 1 1 ,000. In view of the ageing population, this equates to millions of Euros in the coming years.

Falls are caused by complex and dynamic interactions between intrinsic (subject- based) and extrinsic (environmental) factors. Over 400 risk factors have been identified and their prioritization remains unclear. Moreover, the applicability, sensitivity and particularly the specificity of subject-specific assessment of fall risks remains imprecise.

Several studies have investigated the independent capability of several

technologies to prevent falls, including posturography, balance/gate, trunk accelerations, sock pressure sensors, bed/chair alarms and other indoor ambient sensors. However, recent systematic reviews have highlighted that these technologies suffer from numerous limitations, including the fact that the

occurrence of false alarms is too high to maintain full attention of the nursing staff. Moreover, these approaches require the use of additive sensors (i.e. pressure matrices or wearable accelerometers) having no other direct benefits for the elderly health and resulting in additional costs. There is thus a need for an improved method of predicting and so preventing falls that addresses the shortcomings of known approaches.

In one aspect, the invention provides a method of obtaining an indication of the likelihood of a subject to have a fall, the method comprising the step:

(a) determining, from an Autonomous Nervous System (ANS) state

measurement obtained from the subject or a value derived therefrom, an indication of the likelihood of the subject to have a fall,

wherein the ANS state measurement or value derived therefrom is one which is significantly associated with the likelihood of subjects having falls.

In general, the method of the invention will be computer-implemented, by which is meant that the method involves the use of a data processor, which may be embodied in any form of computer, eg a PC, a tablet computer, a smartphone, a microprocessor or microcontroller, or any other form of electronic device. Such a device may be a stand alone device, or it may be connected (by a wired or wireless connection) to a network (e.g. Local Area Network, Wide Area Network, intranet or internet). Thus, in a further aspect, the invention provides a system or apparatus or device comprising at least one processing means arranged to carry out the steps of a method of the invention.

The processing means may, for example, be one or more computing devices (of any suitable form, including those listed above) and at least one application executable in the one or more computing devices. The at least one application may comprise logic to carry out the steps of a method of the invention.

In a further, related aspect, the invention provides a data carrier bearing software comprising instructions for configuring a processor to carry out the steps of a method of the invention. The invention may be of greatest utility in relation to the prediction and prevention of falls in humans, particularly elderly humans. However, the subject may be any animal, preferably a mammal, most preferably a human. In some embodiments, the subject may be one with heart disease, e.g. one with coronary heart disease, hypertension, coronary atrial fibrillation, bradycardia, tachycardia, arterial disease or congestive heart failure.

In some embodiments of the invention, a human subject is over 50, over 60, over 70, over 80 or over 90 years old.

In some embodiments of the invention, the method additionally comprises the step of obtaining one or more ANS measurements from the subject. As used herein, the term "control" relates to an individual or group of individuals of the same species as the subject in question. For example, if the subject in question is a human, the control will also be a human.

The control may be a healthy subject, e.g. one with no history of heart or circulatory disorders.

As used herein, the term "an indication of the likelihood of the subject to have a fall" is intended to mean an indication of whether the subject is likely to have a fall within a limit period following the taking of the ANS measurement. In some embodiments, the prediction may be a prediction of an imminent fall, and the future time period may, for example, be the next 60, 30, 20, 10, 5 or 2 minutes. In other embodiments, the assessment may be of a general propensity of the subject to experience a fall, the future time period being, for example, the next 1 , 2, 3, 4, 5, 6 or 7 days, or longer, e.g. the coming weeks or months.

In particular, the ANS state measurement or value is one or more features associated with the subject's Heart Rate Variability (HRV), or the subject's Heart Rate (HR) and HRV. Heart Rate (HR), also commonly called pulse rate, is the number of times the heart beats per minute. Heart Rate Variability (HRV) is the variation of the heart rate (or pulse rate) across the mean value of the heart rate during its registration.

In other words, HRV is the difference between the instantaneous HR and the mean measured HR, as a time series or function of time.

The heart is a spontaneous organ and it beats autonomously. The variation of heart rate is controlled by our ANS that will increase or decrease its speed according to internal or external circumstances. Therefore, HRV is considered a good estimator of ANS activities or states.

The pulse rate can be found by hand in different body parts and particularly well at:

wrists

· inside of the elbow

side of the neck

top of the foot

However, to have an accurate HRV analysis, the HR needs to be registered with a reliable approach and device. This can be done in various ways, including the following:

1 ) Detecting the peak of the excerpts of the electrocardiogram (ECG) signal, which is known to correspond in time with the heartbeat. These excerpts are known as "QRS" complexes and the peaks are called "R". the distance from one peak to the next one is called RR and measured in milliseconds (ms).

2) Detecting the sound of the heart during a bead (systole or diastole) using a microphone. 3) Detecting the peak of blood pressure in any part of the body (e.g. wrist or neck), which peak corresponds to the moment in which the heart is ejecting the blood in our body (therefore few milliseconds after the heartbeat).

4) Measuring the peak of velocity of blood flow in any part of the body, which increases during an heartbeat.

5) Measuring the peak of % of oxygen in capillaries (this is called SpO2) or other vessels.

6) Detecting the peak of the red component of an image of our skin. Since the colour our skin is due to the quantity of blood circulating locally, and since this quantity transiently increases during an heartbeat, detecting the moments in which the red colour increases, which equate to detecting R peaks (heartbeats).

ECG is the currently preferred method. However, other methods such as photoplethysmography (PPG), blood pressure variation (BPV), phonocardiography (PCG) may be used, though PPG acquired at the fingertip may not be suitable, as it is unable to detect rapid variations in HRV and may also suffer from artefacts that interfere with the measurement.

Once the beats are detected, it is possible to calculate the time distance between consecutive beats (hereinafter called RR, or NN if specifically relating to normal beats). This time series will be the inverse of HR in any time (if HR=60 beats per minutes - or 1 beat per second - the RR interval is 1000 ms).

Features associated with the subject's HR or HRV may be selected from those features set out in Tables A-C. The features of Table A are time-domain features, those of Table B are frequency-domain features, and those of Table C are nonlinear domain features. Table A

Time domain HRV features

Feature Units Description and interpretation

RRMEAN ms The mean of RR intervals (also referred to as NN intervals)

RRMAX ms The maximum RR interval

RRMIN ms The minimum RR interval

RRMED ms The median RR interval

H RMEAN 1 /min The mean of heart rate (or pulse rate)

SDHR 1 /min Standard deviation of instantaneous heart rate (or pulse rate)

values

SDNN ms Standard deviation of all NN intervals

SDANN ms Standard deviation of the averages of NN intervals in all 5 min segments of the entire recording

RMSSD ms The square root of the mean of the sum of the squares of

differences between adjacent NN intervals

SDNN index ms Mean of the standard deviations of all NN intervals for all 5 min segments of the entire recording

SDSD ms Standard deviation of differences between adjacent NN intervals

NN50 Number of pairs of adjacent NN intervals differing by more than 50 ms in the entire recording. Three variants are possible counting all such NN intervals pairs or only pairs in which the first or the second interval is longer

pNN50 — NN50 count divided by the total number of all NN intervals

HRVtri HRV Triangular Index: Number of normal RR intervals divided by the height of the histogram of all the normal RR intervals measured on discrete scale with bins of 1 /128s (7.8125ms)

TINN ms Baseline width of the minimum square difference of triangular

interpolation of the highest peak of the histogram of all normal R-R intervals A sub-set of the above features is: RRMEAN, H R _ME AN, SDHR, SDNN, SDANN, RMSSD, SDNN index, SDSD, NN50, pNN50, HRVtri and TINN.

5 Table B

Frequency Domain HRV features

0 These frequency-domain HRV measurements are based on an estimation of the power spectral density (PSD), preferably by using a Welch periodogram, or a Auto-Regressive (AR) method or a Lomb-Scargle periodogram. These features are in this document labelled with subscripts AR, WE or LS (i.e. LF _AR or LF _WE or

^LF LS)-5

A sub-set of these features is: ULF, VLF, LF, HF, LFnu, HFnu, LF/HF, VLFpf, LFpf and HFpf. Table C

Non-linear domain HRV features

5 Further HRV measurements which may be used in the context of the present invention may be found in Melillo et al. BioMedical Engineering OnLine 201 1 , 10:96 (http://www.biomedical-engineering-online.eom/content/10/1 /96) and Melillo et al. BMC Cardiovascular Disorders 2012, 12:105

(http://www.biomedcentral.com/1471 -2261/12/105), the contents of which are specifically incorporated herein by reference.

Where numerical values for any of the features set out in Tables A-C are quoted herein, it should be understood that the units associated with those values are those listed in the Tables.

In some embodiments, the ANS state measurement is RRMEAN, SDNN, RMSSD, NN50, pNN50, HRVtri, TINN, LF, HF, LF/HF, SD _r SD ₂ , En(0.2), SampEn, D ₂ , a _v a ₂ , --MEAN, LMAX, RPREC, RPDET and ShEn, or a value derived from one or more such measurements.

In some embodiments, the ANS state measurement is RR _M!N , VLF _WE , TP _WE , LF _LS ^{T P} LS' ^L MAX' '"MEAN' ^{RPD IV} > RPDET, or ShEn, or a value derived from one or more such measurements.

In some embodiments, the ANS state measurement is RR _MAX , L _MAX , a ₂ , pNN50, SDNN, VLF _LS or RPREC, or a value derived from one or more such

measurements. In some embodiments, the ANS state measurement is RR _M!N or l_ _MAX , or a value derived from one or more such measurements. In particular, an increase in one or more of RR _M!N , L _MAX , L _MEAN and ShEn relative to a control subject is indicative of the subject having an increased likelihood of having a fall. In some embodiments, the ANS state measurement is VLF _WE , TP _WE , LF _LS , RP _LS , RPDIV or RPDET, or a value derived from one or more such measurements. In particular, a decrease in one or more of VLF _WE , TP _WE , LF _LS , RP _LS , RPDIV and

RPDET relative to a control subject is indicative of the subject having an increased likelihood of having a fall. In some embodiments, the ANS state measurement is a ₂ , pNN50 or RR _MAX , or a value derived from one or more such measurements. In particular, if

<¾ < 0.947 and/or

pNN50 < 26.7 and/or

RR _MAX >2265.6, then the subject may have an increased likelihood of having a fall.

The measurement may be of a ₂ , pNN50 and RR _MAX , and the subject may have an increased likelihood of having a fall if a ₂ < 0.947 and pNN50 < 26.7 and RR _MAX >2265.6.

In some embodiments, the ANS state measurement is l_ _MAX or SDNN, or a value derived from one or more such measurements. In particular, if

^L MAX > ^{2 1 79 and/0 r}

SDNN < 40.4 then the subject may have an increased likelihood of having a fall.

The measurement may be of l_ _MAX and SDNN, and the subject may have an increased likelihood of having a fall if

L _JAV > 2179 and SDNN < 40.4. In some embodiments, the ANS state measurement is VLF _LS , L _MAX or RPREC, or a value derived from one or more such measurements. In particular, if

VLF _LS > 15.1 and/or

l_ _MAX > 2177 and/or

RPREC < 0.514, then the subject may have an increased likelihood of having a fall. The measurement may be of VLF _LS , l_ _MAX and RPREC, and the subject may have an increased likelihood of having a fall if

VLF _LS > 15.1 and l_ _MAX > 2177 and RPREC < 0.514. In some embodiments, the ANS state measurement is a time-domain HRV measurement, or a value derived from one or more such measurements.

Preferably, the time-domain HRV measurement is selected from average of all RR intervals, standard deviation of all NN intervals (SDNN), square root of the mean of the sum of the squares of differences between adjacent NN intervals (RMSSD), number and percentage of differences between adjacent NN intervals that are longer than 50 ms (NN50 and pNN50), standard deviation of the averages of NN intervals in all 5-min segments, mean of the standard deviations of NN intervals in all 5-min segments, maximum of RR intervals, minimum of RR intervals, median of RR intervals, HRV triangular index, i.e. the proportion of all accepted RR intervals to their modal measurement at a discrete scale of 1 /128 s bins, and triangular interpolation of RR interval histogram, i.e. the baseline width of the distribution measured as a base of a triangle, approximating the NN interval distribution by using the minimum square difference, or a value derived from one or more thereof. The ANS state measurement preferably includes at least one non-linear feature, or a value derived therefrom, in particular one or more features set out in Table C, and in particular one or more non-linear features selected from SD _r SD ₂ , En(0.2), D ₂ , a _v a ₂ , LMEAN, L _MA X, RPREC, RPDET and ShEn.

The invention provides a number of predictions of the likelihood of having a fall.

In one manner of implementation, the invention may be used to obtain an indication of the propensity of a subject to experience a fall within a relatively lengthy forthcoming period, or in other words the likelihood of the subject falling within that period. The period may be a period of hours, e.g. the next 1 hour, 6 hours, 12 hours or 24 hours, or it may be a period of days, e.g. the next 1 day, or the next 2, 3, 4, 5, 6 or 7 days. Alternatively, the method may be used for estimation of the likelihood of a fall in a longer timeframe, e.g. the next 1 , 2, 3 or 4 weeks, or the next 1 , 2, 3, 4, 5 or 6 months. In such a manner of implementation, the method of the invention may be carried out just once, or intermittently at intervals, with a view to identifying subjects that are at risk of falls, so that appropriate measures can be put in place to prevent a fall or to mitigate the consequences of a fall, should a fall occur. Alternatively, the subject may be continuously monitored, so that such measures can be taken in the event that the measured features of the subject's HR and HRV indicate that the subject has developed a risk of falling.

As set out above: a) an increase in one or more of RR _M!N , L _MAX , L _MEAN and ShEn may be indicative of the subject having an increased likelihood of having a fall; and/or b) a decrease in one or more of VLF _WE , TP _WE , LF _LS , RP _LS , RPDIV and RPDET may be indicative of the subject having an increased likelihood of having a fall.

In further specific embodiments: c) an increase in two, or three, or all of RR _M!N , L _MAX , L _MEAN and ShEn may be indicative of the subject having an increased likelihood of having a fall; d) an increase in two, three, four, five or all of VLF _WE , TP _WE , LF _LS , RP _LS , RPDIV and RPDET may be indicative of the subject having an increased likelihood of having a fall.

The increase or decrease of the measured feature may be an increase or decrease relative to a corresponding measurement performed on a control (i.e. a difference between the value measured for the subject and a normal value measured on a healthy control), or it may be an increase or decrease relative to a previous value measured for the subject (i.e. a change that indicates that the subject has developed a risk of falling). Specific models that have been developed for implementation in this manner are the following:

Model 1 :

IF

(0(2 < 0.947 and pNN50 < 26.7 and RRMAX >2265.6)

THEN the Odds Risk (OR) of falling in the next week is at least 5.1 2 (CI 95% 1 .42- 18.41 ; p<0.01 )

Model 2:

Rule 1 :

IF

('α ₂ ≤ 0.947' and 'pNN50 < 26.7' and 'RRMAX > 2265.6')

THEN the risk of falling is higher in the next week

[OR of falling in the next week is 5.12 (CI95% 1 .42-18.41 ; p<0.01 )]

Model 3:

Rule2:

IF ('Ι_ΜΑΧ> 2179' and SDNN< 40.4')

THEN the risk of falling is higher in the next week

[OR of falling in the next week is 3.66 (0.79-17.02; 0.08)] Model 4:

Rule 3:

IF

('VLFLS > 15.1 ' and 'LMAX> 2177' and 'RPREC < 0.514') THEN the risk of falling is higher in the next week

[OR of falling in the next week is 1 1 .16 (1 .21 -102.6;<0.01 )]

Model 5:

IF

(Rule 1 is true and Rule 2 is true)

THEN the risk of falling is higher in the next week

[OR of falling in the next week is 4.98 (1 .80-13.79; <0.01 )]

Model 6:

IF

(Rule 1 is true and Rule 3 is true)

THEN the risk of falling is higher in the next week

[OR of falling in the next week is 7.09 (2.31 -21 .76; <0.01 )]

Model 7:

IF

(Rule 2 is true and Rule 3 is true)

THEN the risk of falling is higher in the next week

[OR of falling in the next week is 2.72 (0.65-1 1 .36; 0.16)] Model 8:

IF

(Rule 1 is true and Rule 2 is true)

THEN the risk of falling is higher in the next week [OR of falling of falling in the next week is 4.32 (1 .61 -1 1 .56; <0.01 )]

In another mode of implementation, the method of the invention may be used to obtain an indication of the propensity of a subject to experience a fall within a relatively short forthcoming period, or in other words the likelihood of the subject falling within that period. The period may be a period of minutes, e.g. the next 1 , 2, 3, 4, or 5 minutes. The risk of a fall being experienced on such a timescale may be linked to a behaviour of the subject or a transient metabolic condition of the subject that may lead to a variation in the subject's blood pressure. One

particularly common such behaviour is a change of posture, e.g. occurring when the subject stands up from a seated or prone position.

An orthostatic change, namely a change in body posture from sitting to standing, causes specific changes in heart rate and blood pressure as a compensatory reaction of the body. In the few minutes after standing, there is a redistribution of the blood volume and a pooling of blood in the lower extremities due to

gravitational forces. As a consequence, the venous return to the heart falls and the cardiac filling pressure is reduced, diminishing the stroke volume and cardiac output. Therefore, there is a drop in blood pressure (ΔΒΡ) due to the change of position, a condition referred to as Standing Hypotension (SH) or Orthostatic Hypotension (OH).

The prediction of this ΔΒΡ is particularly relevant as at least 30% of the indoor falls of older subjects happen when they are rising from a bed or chair, such falls in the majority of the cases being due to the ΔΒΡ caused by the activity of standing. Therefore, a model that could predict this ΔΒΡ could help to predict a significant percentage of these falls.

To avoid dizziness or fainting due to the limited blood supply to the brain, the blood pressure in the large arteries decreases to compensate and regulate. The objective of this regulatory mechanism is to achieve normal blood pressure as quickly as possible while providing an adequate supply of blood to the vital organs. Healthy subjects respond with an autonomic adjustment, which increases vascular tone, heart rate and cardiac contractility, and stabilizes arterial pressure. In particular, the sympathetic outflow to the heart and blood vessels increases and the cardiac vagus nerve activity decreases. In healthy subjects, during standing, the contraction of the lower body skeletal muscles prevents excessive pooling and augments the venous return to the heart.

Therefore, the dynamic of the blood pressure and particularly the capability to restore homeostasis after standing is strongly dependent on the status of the ANS. Measurement of features associated with HRV in the time immediately prior to standing provides an indication of the status of the ANS before standing, which in turn enables the prediction of the magnitude of the ΔΒΡ a few minutes after standing, and hence the risk of a fall occurring. In this manner of implementation, the method of the invention will generally be carried out continuously, and may be carried out in conjuction with the monitoring of the subject's behaviour or metabolic condition, so that an event that may give rise to a ΔΒΡ can be identified, and the data gathered in the time leading up to that event analysed in order to predict the magnitude of the ΔΒΡ following that event, and hence the risk of that event will cause a fall. Thus, accelerometers or other motion sensors may be used to detect a change in posture of the subject, as in rising from a seated or prone position to standing, or blood glucose sensors may be used to identify a fall in blood glucose level for a diabetic subject. If such an event is detected, the status of the ANS in the period immediately preceding that event, typically measured in the period of 2 to 5 minutes prior to the event, is used to predict the ΔΒΡ following that event.

Thus, in one particular aspect, the invention provides a method of predicting a fall, which method involves

a) simultaneously

a1 ) monitoring a subject's Heart Rate Variability (HRV) by determining one or more parameters associated with HRV; and a2) monitoring the subject for the occurrence of an event that potentially may cause a change in the subject's blood pressure; and

b) if such an event is detected, correlating the one or more parameters associated with HRV measured during a pre-determined time interval prior to said event, or a value derived from said one or more parameters, with the likelihood of the subject experiencing a fall.

The pre-determined time interval is preferably a period of several minutes, e.g. between about 1 and about 5 minutes, prior to the detected event. The pre- determined time interval may, for instance, commence up to 10 minutes prior to the detected event, more commonly 5 minutes prior to that event, and may end immediately prior to the event, or a short time before that event, e.g. 30 seconds, 1 minute or 2 minutes prior to the event. The one or more parameters preferably include at least one time domain parameter (e.g. one or more of those listed in Table A) and/or at least one frequency domain parameter (e.g. one or more of those listed in Table B). The one or more parameters preferably include at least one non-linear parameter (e.g. one or more of those listed in Table C).

The parameters associated with HRV may be related in a linear manner to the ΔΒΡ following the detected event.

The parameters associated with HRV may include RMSSD, NN50, TINN, HF and RPDET. The relationship between ΔΒΡ (in immHg) and those parameters may be expressed as:

ΔΒΡ = A1 + B1 ^* RMSSD + C1 ^* NN50 + D1 ^* TINN + E1 ^* HF + F1 ^* RPDET where A1 is from -20 to -30, B1 is from 0.35 to 0.55, C1 is from -0.04 to -0.06, D1 is from -0.015 to -0.025, E1 is from -0.005 to -0.015 and F1 is from 0.24 to 0.36. The values of A1 , B1 , C1 , D1 , E1 and F1 may vary within those ranges, depending on numerous factors, including the subject's age, medical condition (including diabetes, hypertension, heart failure or other cardiovascular conditions), medical history, genetic factors, time of day, medication or combination of medications. The subject may be considered to be at risk of an imminent fall if (A1 + B1 ^* RMSSD + C1 ^* NN50 + D1 ^* TINN + E1 ^* HF + F1 ^* DET) < THRESHOLD where THRESHOLD represents a predicted drop in blood pressure (in mmHg) that gives rise to a risk of falling. The value of THRESHOLD may typically be -20, on the basis that a drop in blood pressure of 20mmHg or more gives rise to a likelihood of a fall.

In other words, the subject may be regarded as being at risk of an imminent fall if

(A1 + B1 ^* RMSSD + C1 ^* NN50 + D1 ^* TINN + E1 ^* HF + F1 ^* RPDET) < -20 In another embodiment, the parameters associated with HRV may include SDNN, pNN50, HF, SD ₂ and RPDET. The relationship between ΔΒΡ (in mmHg) and those parameters may be expressed as:

ΔΒΡ = A2 + B2 ^* SDNN + C2 ^* pNN50 + D2 ^* HF + E2 ^* SD ₂ + F2 ^* RPDET where A2 is from -180 to -260, B2 is from -1 .6 to -2.0, C2 is from 0.18 to 0.28, D2 is from 0.008 to 0.009, E2 is from 0.8 to 1 .2, and F2 is from 1 .9 to 2.8. The values of A2, B2, C2, D2, E2 and F2 may vary within those ranges, depending on numerous factors, including the subject's age, medical condition (including diabetes, hypertension, heart failure or other cardiovascular conditions), medical history, genetic factors, time of day, medication or combination of medications.

The subject may be considered to be at risk of an imminent fall if (A2 + B2 ^* SDNN + C2 ^* pNN50 + D2 ^* HF + E2 ^* SD ₂ + F2 ^* RPDET) < THRESHOLD where THRESHOLD is as defined above, and may typically be -20. In other words, the subject may be regarded as being at risk of an imminent fall if

(A2 + B2 ^* SDNN + C2 ^* pNN50 + D2 ^* HF + E2 ^* SD ₂ + F2 ^* RPDET) < -20 In the event that a risk of the subject falling is identified, the system may be further adapted to issue an alarm notification to the subject or to a person associated with the subject. Typically, such an alarm notification may be an audible or visual alarm given to a care worker or other person with responsibility for the well-being of the subject, so that they can take action to prevent a fall occurring or provide assistance in the event that a fall has occurred.

Specific models that have been developed for implementation in this manner are the following: Model 9:

In this model, ΔΒΡ is estimated as a function of RMSSD, NN50, TINN, HF, and RPDET:

ΔΒΡ = -25.67 + 0.45 ^* RMSSD - 0.05 ^* NN50 - 0.02ΗΊΝΝ - 0.01 ^* HF + 0.3 ^* RPDET

IF

(-25.67 + 0.45 ^* RMSSD - 0.05 ^* NN50 - 0.02ΗΊΝΝ - 0.01 ^* HF + 0.3 ^* RPDET) < -20 THEN

the risk of falling within the next 5 minutes is high.

Model 10:

In this model, ΔΒΡ is estimated as a function of SDNN, pNN50, HF, SD ₂ and RPDET.

The fall in systolic blood pressure ΔΒΡ is linked to the HRV measures as described in the next model: ΔΒΡ = -219.39 - 1 .83 ^* SDNN + 0.23 ^* pNN50 + 8.5 ^* 10 ^"3* HF + 0.97 ^* SD ₂ +

2.34 ^* RPDET

IF

(-219.39-1 .83 ^* SDNN+0.23 ^* pNN50+8.5 ^* 10 ^"3* HF+0.97 ^* SD2+2.34 ^* RPDET) < -20 THEN

the risk of falling within the next 5 minutes is high.

A number of features of HRV may be highly correlated, and so it may be possible to derive other formulae from Models 9 and 10, by changing one feature with another that is correlated with it, and multiplying the associated constant with the correlation coefficient linking the two features. The accuracy of the newly derived model may depend on the p-value of the correlation. As well as being associated with increased risk of falls, a change in blood pressure is implicated in a number of disease states. For instance, such a change may be associated with diabetes, anaemia, atherosclerosis or Parkinson's Disease. A change in blood pressure may also be caused by low blood volume

(hypovolaemia), which in turn may result from bleeding, dehydration, prolonged bed rest, or the use of drugs such as diuretics and vasodilators. A change in blood pressure may also be brought about by administration of medication, e.g. medication for the treatment of hypertension or other cardiovascular disorders, or any of numerous other reasons. The ability to predict a change in blood pressure (ΔΒΡ) may therefore be of utility in other circumstances than the prediction and prevention of falls. According to another aspect, therefore, the invention provides a method of predicting a change in a subject's blood pressure brought about by an event that potentially may cause a change in the subject's blood pressure, which method involves

a) simultaneously

a1 ) monitoring a subject's Heart Rate Variability (HRV) by determining one or more parameters associated with HRV; and

a2) monitoring the subject for the occurrence of said event; and b) if such an event is detected, predicting on the basis of the one or more parameters associated with HRV measured during a pre-determined time interval prior to said event, or a value derived from said one or more parameters, the magnitude of the change in the subject's blood pressure brought about by the event.

As described above, the event may be a behaviour of the subject, such as a change of posture, or it may be a transient metabolic condition. The pre-determined time interval is preferably a period of several minutes, e.g. between about 1 and about 5 minutes prior to the detected event, as described above. The one or more parameters preferably include at least one time domain parameter (e.g. one or more of those listed in Table A) and/or at least one frequency domain parameter (e.g. one or more of those listed in Table B). The one or more parameters preferably include at least one non-linear parameter (e.g. one or more of those listed in Table C).

The parameters associated with HRV may be related in a linear manner to the ΔΒΡ following the detected event.

The parameters associated with HRV may include RMSSD, NN50, TINN, HF and RPDET. The relationship between ΔΒΡ (in mmHg) and those parameters may be expressed as: ΔΒΡ = A1 + B1 ^* RMSSD + C1 ^* NN50 + D1 ^* TINN + E1 ^* HF + F1 ^* RPDET where A1 , B1 , C1 , D1 , E1 and F1 are as defined above.

In another embodiment, the parameters associated with HRV may include STNN, pNN50, HF, SD ₂ and RPDET, and the relationship between ΔΒΡ (in mmHg) and those parameters may be expressed as:

ΔΒΡ = A2 + B2 ^* STNN + C2 ^* pNN50 + D2 ^* HF + E2 ^* SD ₂ + F2 ^* RPDET where A2, B2, C2, D2, E2 and F2 are as defined above.

In general, the features of the invention discussed above in relation to the prediction of falls may also be applicable to the use of the invention to predict ΔΒΡ.

Where the method of the invention is conducted continuously, it may be

implemented in the manner of a Smart Health Monitoring System (SHMS). Signal acquisition may be performed using a multi-sensing wearable device, various forms of which are commercially available, and will be familiar to those skilled in the art. As described above, such a device may be a stand alone device, or (more commonly) it may be connected by a wired or wireless connection to a network (e.g. Local Area Network, Wide Area Network, intranet or internet). Example 3 below provides an illustration of how such an SHMS may be implemented.

The present invention is further illustrated by the following Examples. It should be understood that these Examples are given by way of illustration only. From the above discussion and these Examples, one skilled in the art can ascertain the essential characteristics of this invention, and without departing from the spirit and scope thereof, can make various changes and modifications of the invention to adapt it to various usages and conditions. Thus, various modifications of the invention in addition to those shown and described herein will be apparent to those skilled in the art from the foregoing description. Such modifications are also intended to fall within the scope of the appended claims.

The disclosure of each reference set forth herein is incorporated herein by reference in its entirety. List of other abbreviations:

Figure 1 depicts the experimental protocol used the studies described in

Example 2.

Figure 2 shows values of ΔΒΡ predicted by this model developed in Example 2(A) against measured values.

Figure 3 shows measured values of ΔΒΡ, as described in Example 2(B), together with values predicted by the model developed in Example 2(A). Figure 4 depicts the general system architecture of a Smart Health Monitoring System (SHMS) described in Example 3.

Figure 5 shows a sample screenshot of the results of HRV remote processing for a patient in the study described in Example 3.

EXAMPLE 1

AUTOMATIC PREDICTION OF FALLS VIA HEART RATE VARIABILITY AND DATA MINING IN HYPERTENSIVE PATIENTS

This study presents the results of a cross-sectional study investigating the relationship between HRV patterns and the risk of falling. The hypothesis explored is that a reduced HRV complexity is associated with an increased risk of falling, because it reflects the deteriorating state of the Autonomous Nervous System (ANS).

I. METHODS AND MATERIALS

A. Study design

According to conventions used for retrospective cross-sectional studies, the following was assumed: HRV pattern corresponding to the rules identified with data mining methods was the "risk factor" under investigation; fallers were considered as "cases" and non-fallers as "controls"; patients with a HRV positive to these rules for at least 10% of the day (nominally 24 hours) were considered "exposed" to the risk factor under investigation.

The study analyzed clinical 24h ECG Holter recordings of 168 hypertensive patients (72±8years, 60 female). 47 subjects experienced a fall within the 3 months before or after the registration. B. HRV processing

The series of RR beat intervals were obtained from ECG recordings using open- source software [ Zong et al (Eds), A robust open-source algorithm to detect onset and duration of QRS complexes Computers in Cardiology, 2003: IEEE]. HRV was analyzed concurrently in segments (excerpts) of 30 minutes, excluding those with fewer than 600 valid beats. Standard linear HRV analysis was performed according to International Guidelines [Malik et al, Eur Heart J. 1996;17(3):354-81]. Additionally, non-linear features were computed according to recent literature

[Acharya et al, Med Biol Eng Comput. 2006;44(12): 1031-51]. The HRV analysis was performed using open-source HRV software [J R Design, Evaluation and application of heart rate variability software, 2010; Melillo et al IEEE J Biomed Health Inform. 2013;17(3):727-33\. Standard time-domain HRV measures were calculated:

RRMEAN SDNN RMSSD NN50

pNN50

SDANN RRMAX RRMIN RRMED HRVtri

TINN

The generalized frequency bands in case of short-term HRV recordings are the very low frequency (VLF, 0-0.04 Hz), low frequency (LF, 0.04-0.15 Hz), and high frequency (HF, 0.15-0.4 Hz). The frequency-domain measures extracted from the PSD estimate for each frequency band include:

VLF

LF

HF

TP

LFnu • HFnu

• LF/HF

• VLFpf

• LFpf

· HFpf

The frequency-domain HRV measures rely on the estimation of power spectral density (PSD) computed, in this work, with three different methods: Welch periodogram, Auto-Regressive (AR) method and Lomb-Scargle periodogram. For the Welch's periodogram, the NN interval was first interpolated with cubic spline interpolation at 4 Hz. The interpolated series was then divided into overlapping segments of length 256 points and each segment was Hamming-windowed. The overlap was chosen to be 128 points. AR model order was 16. Welch-based, AR-based and Lomb-Scargle-based measures are identified by the subscripts WE, AR and LS, respectively. For instance, TPWE refers to the estimation of the total power computed by using the Welch periodgram, while LF _L s refers to LF computed by using the Lomb-Scargle periodogram.

Non-linear properties of HRV were analysed by the following methods

Poincare Plot [Melillo et al, Medical and Biological Engineering and

Computing. 2011;49(1):67-74 , which estimates the correlation between successive RR intervals.

Approximate Entropy [Richman et al, Amer. J. Physiol. - Heart and

Circulatory Physiology. 2000;278(6):H2039-H4], which measures the complexity or irregularity of the RR series, with the values of parameters r and m set at respectively 2 and 20% of SDNN.

Correlation Dimension [Carvajal et al, Computer Methods and Programs in Biomedicine. 2005;78(2):133-40], which measures the complexity used for the HRV time series, with a parameter m=10. d) Detrended Fluctuation Analysis [Penzel et al, IEEE Trans Bio Med Eng. 2003;50(10):1143-51], which measures the correlation within the signal, by two parameters: short-term fluctuations (a-i ) and long-term fluctuations (a ₂ ). e) Recurrence Plot [Zbilut et al, Medical Engineering & Physics.

2002;24(1):53-60], which measures the complexity of a time-series, with the values of parameters chosen according to published literature [Niskanen et al, Computer Methods and Programs in Biomedicine. 2004;76(1):73-81 ; Dabire et al, Amer. J. Physiol. - Heart and Circulatory Physiology.

1998;275(4):H1290-H7\.

The following non-linear measures were computed:

. SD ₁

. SD ₂

. En(0.2)

· En(r _max )

. En(rchon)

. D ₂

• a ₁

. a ₂

. ShEn

. RPDET

. RPREC

• LMEAN

• LMAX

. LLE

. LLE (HF)

. LLE (LF)

C. Statistics and data mining methods

Differences between HRV features of faller and non-faller subjects were assessed by repeated measure regression model estimated using Generalized Estimating Equations (GEE) [Zeger et al, Biometrics. 1988:1049-60]. Two well-known and complementary approaches to data mining were used:

divide-and-conquer decision tree algorithms such as C4.5, CART [Breiman et al, Classification and regression trees. Belmont, Calif.: Wadsworth International Group; 1984 , or M5P and covering rule induction algorithms such as RIPPER [Cohen W (Ed), Fast effective rule induction. Twelfth International Conference on Machine Learning; 1995\ or PART. Particularly, a combination of Naive Bayes [Friedman et al, Machine Learning. 1997;29(2):131-63], lift chart and the C4.5 [Quinlan JR. C4.5: programs for machine learning. Morgan Kaufmann; 1993], CART, or RIPPER methods was employed, using the Weka platform for knowledge discovery [Hall et al, ACM SIGKDD explorations newsletter,

2009;11(1):10-8]. Binary splits at internal nodes are made for numerical features, while multi-way splits are done for the categorical ones. All of the features are considered for splits in each node of the tree and a gain ratio is used for evaluating the contribution of each feature. There are some differences between CART and C4.5, when CART is used for classification, including the criterion used for branching in internal nodes and the type of pruning. CART uses the Gini index as impurity measure for determining the split at the internal node. The CART method also uses cost-complexity based sub-tree post-pruning, which is more

conservative than empirically based sub-tree replacement and raising of C4.5 and leads to construction of smaller trees. The general approach is to first prune those sub-trees that, relative to their size, lead to the smallest increase in error on the training data. The pruning is usually performed by an internal cross-validation procedure. The RIPPER {Repeated Incremental Pruning to Produce Error Reduction) algorithm splits the training set into two distinctive sets: a growing set and a pruning set. On the growing set, it greedily constructs rules with perfect coverage. A Naive Bayes (NB) classifier is used to obtain a probabilistic model of the dataset with respect to the health event investigated in this study {fall = 1 ). It outputs posterior probabilities as a result of the classification procedure. Excerpts were ranked according to posterior Bayesian probability for falling and visualized by a lift chart. Highest 10%, 20% and 30% excerpts were then analyzed by C4.5, CART and RIPPER. Additionally, to estimate the maximum potential of the dataset to predict falls based on HRV features, several data mining classification methods were used. This second stage aimed to discover whether it was possible, and to what degree, to obtain a model for falls based on HRV. Sensitivity, specificity, and Area Under the Curve were utilized to assess the best models obtained by these methods.

Random forest (RF) is a decision tree ensemble learner developed by Breiman [Breiman, L, 'Random Forests', Mach Learn, 2001, 45(1), 5-32]. Decision trees that compose the forest are constructed by choosing their splitting attributes from a random subset of k attributes at each internal node. The best split is taken among these randomly chosen attributes and the trees are built without pruning, as opposed to C4.5. The quality of the split at an attribute is determined by its Gini impurity index. RF avoids overfitting due to two sources of randomness - the aforementioned random attribute subset selection and bootstrap training set sampling coupled with majority voting (also referred as bagging), which is shown to reduce the variance of the classifiers.

Rotation forest (RTF) is an ensemble method capable of both classification and regression, depending on the base classifier [Kuncheva et al, Multiple Classifier Systems, Springer, 2007]. By default, rotation forest uses C4.5 decision trees as the base classifiers. The algorithm focuses on presenting transformed data to the classifier by using a projection filter such as principal component analysis (PCA), non-parametric discriminant analysis, random projections, and independent component analysis. The most successful projection filter is the PCA filter.

AdaBoost.MI (AB) is a well-known algorithm for boosting weak classifiers

[Schapire et al, Annals of Statistics, 1998, 26(5), 1651-1686]. The idea of the AB algorithm is to penalize the instances in the training set that are correctly classified by the classifier. In the first step, the algorithm uses the bootstrap method to select the instances for the first training set, by giving equal chance to all the instances. The base classifier is trained and the instances are classified. The instances that are correctly classified receive a penalty to their weight for the next step of training-classification cycle. The algorithm terminates after a predetermined number of iterations. A weight is contributed to each constructed classifier. In the testing phase, each classifier provides a probability estimate for the target class. Each time a target class is selected, its weight is increased, depending on the weight of the classifiers. Finally, voting is performed that selects the target class with the highest weight.

MultiBoost (MB) is regarded as an extension to AdaBoost that combines the AB algorithm with the wagging procedure, which is itself an extension of the basic bagging method [ Webb, G.I., Mach Learn, 2000, 40(2), 159-196].

RUSBoost (RB) is a hybrid approach recently proposed by Seiffert et al. [Seiffert et al, Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE

Transactions on, 2010, 40(1), 185-19] to handle class imbalance. RB relies on the Random Under-Sampling (RUS) technique and AB as boosting algorithm. CART was adopted as weak learner. RUS is one of the most common data level algorithms to deal with unbalanced dataset or rare class problem. RUS randomly discards the majority class samples in order to modify the class distribution, until a desired class distribution (e.g. equal number of instances in the majority and minority classes) is achieved. However, since HRV features have been shown to be correlated, there is the risk that some of the computed features might be redundant and could worsen the classifier performance by increasing the running time and reducing its generalization ability. In order to find the optimal feature space, we adopted the PCA method and we tested the proposed classifier with different number of dimensions.

In order to assess the performance of the classifiers, we adopted the 10-fold person-independent cross-validation [Jha et al, Automatic Facial Expression Recognition Using Extended Ar-Lbp', Wireless Networks and Computational Intelligence, (Springer, 2012)]. In the 10-fold person-independent cross-validation method, subjects are partitioned into two subsets in each round (totally 1 0 rounds): one with 90% subjects for training and the other with 10% subjects for testing. All the excerpts of the same subjects were included in the training or in the testing dataset. Since we were interested in a subject-based classification, for each subject, the proportion of excerpts classified as fallers was computed and considered as an estimate of the probability that the subject belongs to fallers. A subject-based ROC curve analysis was performed: for all the cut-points, true and false positive rates were calculated and we selected as the best cut-point the one that maximizes the true positive rate (TPR), provided a false positive rate (FPR) lower than 20%. The most common measures for binary classification

performance were computed according to the formulae in Table 1 .1 .

Table 1 .1

Binary Classification Performance Measures

TP = number of fallers correctly classified

TN = number of non-fallers correctly classified

FP = number of non-fallers incorrectly classified as fallers

FN = number of fallers incorrectly classified as non-fallers

RF was constructed using an ensemble of 100 random trees with no limit to tree depth. RTF was constructed with an ensemble of 10 C4.5 trees using PCA filter (all dimensions retained). AB was used in combination with C4.5 classifier, the number of iterations was varied between 10 and 200 with steps of 50 iterations, C4.5 decision tree was tested with a variable minimal number of observations in each leaf (2, 5, 10, 20). Confidence factor for pruning was set to default of 0.25. MB was used in combination with C4.5 classifier, the number of iterations was varied between 10 and 200 with steps of 50 iterations, C4.5 decision tree was tested with a variable minimal number of observations in each leaf (2, 5, 10, 20). Confidence factor for pruning was set to default of 0.25. The number of subcommittees was set to 30% of the number of iterations (classifiers), as default. RB was evaluated with PCA dimension varying between 2 and 20. The number of iterations was varied from 20 to 500 with steps of 20 iterations, and CART was tested with a variable minimal number of observations in each leaf (5, 25, 50 and RB' default) and of misclassification cost ratio (from 1 to 20). Post-sampling 50:50 class distribution was adopted. The rules can be read from the C4.5 and CART tree classifier structure once the tree is built.

A set of rules to automatically identify fallers were obtained through five steps that ensure with high certainty that all the relevant and interesting patterns in the data are found. The steps included:

1 . Obtaining an inferable Bayesian probabilistic model for fall occurrence based on excerpts.

2. Ranking of the excerpts based on the Bayesian probability for fall occurrence, 3. Focusing on several subgroups of excerpts that have high probability of fall occurrence.

4. Construction of rules for each such subgroup using several data mining algorithms.

5. Selection of relevant rules for inclusion in the final set of rules.

The most relevant rules were the ones that maximized simultaneously accuracy and coverage. Coverage was defined as the number of excerpts in the dataset that are positive to the rule. Accuracy was defined as the percentage of excerpts that come from fallers and are positive to the rule. For inclusion of a rule in the final rule set, a rule had to have coverage of at least 70 excerpts with an accuracy of at least 50%. These criteria were necessary in order to filter all the less significant rules that were obtained by data mining algorithms. An additional criterion for rule selection was that the rules had to include only features that could be reliably estimated.

II. RESULTS

A significant variation was observed in 10 HRV measures out of the 67

considered, as reported in Table 1 .2.

Table 1 .2

Significant differences in HRV measures between non-fallers and fallers

The frequency-domain features showed a consistently depressed HRV in fallers. Moreover, non-linear features exhibited a significant increase, particularly the increase of L _M AX, indicating a less chaotic behaviour of HRV in fallers. The final rule set obtained from the data mining methods was:

1 . Rule 1 : 'α ₂ < 0.947' and 'pNN50 < 26.7' and 'RRMAX≥ 2265.6'

2. Rule 2: 'L _M AX> 2179' and SDNN < 40.4'

5 3. Rule 3: 'VLF _LS > 15.1 ' and 'Ι_ _ΜΑ χ> 2177' and 'RPREC < 0.514'

Rule 1 was found in 132 excerpts of which 94 were from fallers, therefore with an accuracy of 71 % among excerpts. Rule 2 was found in 104 excerpts of which 58 were from fallers, therefore with an accuracy of 55.8% among excerpts. Rule 30 was found in 90 excerpts of which 59 were from fallers, therefore with an accuracy of 65.6% among excerpts. All three rules consistently indicated a depressed HRV, including a depressed value of features reflecting higher variations in HRV (as pNN50, SDNN), a dominance of lower frequencies (VLF _L s) and a less chaotic state (LMAX)-5

The combinations of these three rules were explored and the results achieved for subjects, not for the HRV excerpts, are given in Table 1 .3.

Table 1 .3

0 Odds Ratio (OR) of falling for subjects presenting one of the HRV patterns

described in the table

^* Statistically significant with p<0.01 Regarding the data mining methods, the performance of the developed classifiers estimated using 10-fold cross-validation technique showed a low sensitivity and high specificity (Table 1 .4).

Table 1 .4

Performance of the classifiers in identifying fallers basing on HRV measures

III. DISCUSSION

This study presented the results of a retrospective cross-sectional study

investigating the OR of falling among those subjects having abnormal HRV patterns. These HRV patterns were defined as a combination of up to three significant rules, which were extracted using established data mining methods. These rules employed seven features, extracted with linear and non-linear methods from time- and frequency-domains, indicating consistently a depressed HRV. These three rules reflected:

1 . a prevalence of lower PSD frequencies (VLF _L s), which are generally

associated with an abnormal prevalence of vagal tone;

2. a depression of those time-domain features reflecting a higher variability of HRV (pNN50 and SDNN,), which is generally associated with an increased cardiovascular risk;

3. a reduction of HRV complexity, (e.g. as shown by increased recurrence plot non-linear feature LMAX), suggesting a reduction of deterministic 'chaos' which is associated with a reduced ability to react to external solicitation and proven to be related to unhealthy conditions. These results were independently confirmed by statistical analysis, which confirmed that fallers had a depressed power spectrum (VLF, LF, and total power) and decreased chaotic behavior of HRV. The odds of falling in subjects having a depressed HRV were significantly higher than in subjects with a non-depressed HRV, suggesting that a depressed HRV is a risk factor for falls.

This study revealed that a depressed HRV could be associated with an increased risk of falls. The statistical analysis on HRV linear and non-linear measures of the current dataset showed that frequency and non-linear measures significantly differed between fallers and non-fallers. Moreover, the statistical analysis suggested that a depressed HRV, particularly at LF, and a less 'chaotic' behavior of HRV, as assessed by Recurrence Plot features, could be associated with an increased risk of falling. Finally, we computed the feature importance according to RF algorithm, and observed that among ten most relevant features, there are frequency domain features expressed in normalized units, non-linear features and geometric linear features. The best performance was achieved by a hybrid data-mining algorithm, RB, integrated with feature extraction based on PCA. This classifier achieved a relatively high specificity and accuracy (80% and 72% respectively), but low sensitivity (51 %). The sensitivity rate achieved is consistent with the fact that at least 42% of falls are due to transient problems, which are related to ANS and CVS states. Since a limited proportion of falls are directly caused by CVS (i.e. syncope), the results presented in this study suggest that ANS/CVS dysfunctions may be responsible for a temporary reduced capability to react to extrinsic risk factors (i.e. reduced reflex velocity) so as to avoid falls. Moreover, this study shows that these dysfunctions are detectable with HRV monitoring. Moreover, the low rate of false positives (1 -SPE=19.8%) suggested that this approach based on HRV analysis could be successfully used in clinical settings, possibly in

combination with other approaches. Furthermore, these findings reinforce the importance of non-linear analysis of HRV, which has been shown to improve the discrimination power between different patho-physiological conditions. Finally, the comparison of the method of the invention, based on HRV, with functional mobility tests and computer-based tests to discriminate between fallers and non-fallers, reported in Table 1 .5, showed higher OR and/or relative risk values for the method of the invention ""Abnormal HRV").

Table 1 .5

Comparison with other tests for faller identification

¹ Tiedemann et al, Age Ageing, 2008;37(4):430-5

2Schoene et al, Age Ageing. 2014;43(2):285-9

The proposed method showed higher performance than all the functional tests, which had relative risk (RR) ranging from 1 .3 to 2.3 and sensitivity and specificity scores ranging from 1 1 % to 78%, and 28% to 93%, respectively. More recently, the Stroop Stepping Test (SST), using low-cost computer gaming technology, has been proposed to discriminate between older fallers and non-fallers, but the authors provided only the odds ratio (1 .7), which is lower than the one proposed here and reported in Table 1 .5.

Also, the proposed method of the invention is clinically feasible, since it only requires a ECG recording, which is often performed in cardiovascular patients though wearable devices. For instance, the proposed method does not require the use of other technologies, such as wearable accelerometers or pressure matrices, which are not used in everyday clinical practice, not having direct benefits for hypertensive outpatients. For that reason, the proposed method could be used widely in outpatient settings to identify high-risk patients who need further assessment and could benefit from fall prevention programs or fall detection systems. In particular, the HRV analysis and the automatic classification could be obtained through a web cloud-based platform, the ECG could be easily acquired by a commercial wearable device (e.g. Bioharness BH3 manufactured by Zephyr Ltd) and an ad hoc developed client application, and the physician could see the acquired signal and the processing results (i.e. HRV analysis and automatic classification) using a web browser.

Finally, our findings have been obtained in a population of hypertensive patients, in which HRV is already known to be depressed compared to healthy people. This suggests that depressed HRV could be a more relevant risk factor for falls in people free of cardiovascular disease.

IV. CONCLUSIONS

The current study demonstrates the viability of an automated method based on HRV analysis to identify fallers among elderly suffering from cardiovascular disease. Specifically, this study shows that the risk of falling in elderly people with a depressed HRV is five times higher than in those without a depressed HRV (Odds Ratio 5.12, CI 95% 1 .42-18.41 , p<0.01 ). The classifier presented achieved a satisfactory overall diagnostic accuracy and specificity, showing better performances than several functional tests proposed in literature for fall risk assessment. Since the proposed method requires only ECG recording, which is often performed in cardiac patients, it could be an inexpensive and clinically feasible tool for identifying older hypertensive subjects in need of further medical assessment. The accuracy and the sensitivity achieved suggest that HRV based classification would be a valuable complementary adding to other multidisciplinary approaches already in use to predict and prevent falls.

EXAMPLE 2(A)

DEVELOPMENT OF A MODEL USING SHORT TERM HEART RATE VARIABILITY TO PREDICT BLOOD PRESSURE DROPS DUE TO STANDING The current study proposes a mathematical model to predict the lowering of blood pressure (ΔΒΡ) due to standing in healthy subjects, based on the HRV features extracted from an ECG recorded, through the use of wearable sensors, in the five minutes prior to standing. I. METHODS

A. Study population

This study was conducted on a group of 10 healthy subjects, enrolled in accordance with the following selection criteria, namely that they were:

• not suffering from any pathological cardiovascular conditions, neurological or psychiatric disorders or other severe diseases;

• not taking any medication at the time of the study;

• not professional athletes or high-level sport participants;

• had not taken any caffeine or alcohol in the 12 hours prior to the

measurements.

B. Experimental protocol

The protocol was defined to maximize the repeatability and reproducibility of the experiments and aimed to simulate as far as possible the real life action of standing from a bed.

During the tests, an ECG signal was monitored using a one-lead wearable electrocardiogram sensor, the BH3-M1 (Zhephyr Ltd), attached to lightweight patches (33 grams) using standard ECG electrodes. In addition, a three-lead ECG was recorded using a wearable biomedical amplifier, the Nexus 10 (MindMedia Ltd) in order to benchmark the signals acquired with the BH3-M1 .

The BP was measured with a digital sphygmomanometer, the M2 basic (OMRON Ltd), with the left arm comfortably positioned on a horizontal surface and the cuff positioned at the level of the heart at about 2 cm from the inner side of the elbow.

The experiments were carried out in a quiet room, with dimmed lighting and a comfortable temperature of about 23 °C. All the experiments were carried out at the same time of the morning to minimize the circadian effects on the HRV.

The protocol was composed of three phases (sitting, lying and standing) as described below (see Figure 1 ):

• Phase 1 (Sitting): the volunteers were invited to sit in a comfortable

position for a baseline recording of BP and an ECG (>2 minutes).

• Phase 2 (Lying): the volunteers were invited to lie down in a supine

position at about 45 cm from the ground for 10 consecutive minutes (5 minutes of resting + 5 minutes for the recordings): systolic and diastolic BP was recorded four times (with a 60 second interval) in the final 5 minutes before standing; an ECG was recorded during these final 5 minutes.

• Phase 3 (Standing): the volunteers were invited to stand up actively

(without any help) and to stay in an upright position for 5 minutes. The subjects were all trained to stand up in a uniform manner: tilting the trunk and simultaneous twisting the body to the left; putting on the floor first the left and then the right foot; resting for five seconds; and finally standing up without using their hands to help. Once standing, systolic and diastolic BP was recorded four times (with a 60 second interval) and an ECG was recorded continuously for 5 minutes.

After standing the subjects were asked to report any symptoms of vertigo or dizziness and, if this was present, its magnitude (moderate or severe). Phases 2 and 3 were repeated four times after 5 minutes of resting. The whole protocol lasted approximately 100 minutes for each subject.

C. Signal registrations

During Phases 2 and 3, HRV excerpts of 5 minutes length were extracted from the ECG recordings.

D. Signal feature extraction

The ECG measurements were pre-processed by using Kubios [Niskanen et al, Computer methods and programs in biomedicine 2004, 76( 1):73-81; Tarvainen et al, IEEE transactions on bio-medical engineering 2002, 49(2):172-175 , a Matlab- based software package for event-related bio-signal analysis developed by the University of Kuopio, Finland. Kubios is an advanced computer program to extract and analyse HRV.

Each ECG record is cleaned from power line interference, and muscle and movement artefacts, by using a cubic spline interpolation method [Daskalov et al, Medical engineering & physics 1997, 19(4):375-379; Mateo et al, IEEE

transactions on bio-medical engineering 2000, 47(8):985-996\. The distance between two successive heart beats in a normal rhythm determines the RR interval, also called the NN interval. Since artefacts in the RR interval (i.e. the intervals between consecutive R peaks in the ECG signals) time series may interfere with the analysis of these signals, we adopted the Kubios Artefact removal RR filter with a maximum threshold of 5%.

The filtered signals were then analysed. A standard linear HRV analysis was performed according to the guidelines of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology [Electrophysiology TFotESoCtNASoP: Circulation 1996, 93(5): 1043-1065]. Additionally, non-linear features were computed according to the literature [Melillo et al, Non-linear analysis research in biomedical engineering. Focus on non-linear analysis research, Nova Science Publishers, 2013; Acharya et al, Medical & biological engineering & computing 2006, 44(12). -1031-1051]. E. Raw Dataset

The HRV measurements were associated with the BP measurements as described in Figure 1 and collected in a database. According to the protocol, the following data were recorded for each subject:

· 1 HRV excerpt of 2 minutes, 1 systolic and 1 diastolic BP measurement during Phase 1 (sitting).

• 1 HRV excerpt of 5 minutes, 4 systolic and 4 diastolic BP measurements during Phase 2 (lying), which were repeated four times.

• 1 HRV excerpt of 5 minutes, 4 systolic and 4 diastolic BP measurements during Phase 3 (standing), which were repeated four times.

Therefore, the database contained nine instances for each subject. Each instance / in the database was constituted by the following 30 pieces of information:

• sub_id: a number value from 1 to 10 to univocally identify each subject; · HRVjd: a number value indicating one of the three HRV excerpts related respectively to the three phases (Phase 1 : sitting, Phase 2: lying or Phase 3:standing);

• test_id: a number value from 1 to 4 indicating the experiment repetition;

• symptom_id: a number value form 0 to 2 to indicate if the subject feels any manifestation of vertigo/dizziness: 0 = no vertigo; 1 = moderate vertigo; 2 = major vertigo;

• SYS_BP/ ^' : the 9 values of the Systolic Blood pressure measured;

• DIA_BP/ ^' : the 9 values of the Diastolic Blood pressure measured;

• /: a vector containing the 22 HRV measurements reported in Table 1 ;

Therefore, each instance ' was defined as follows: i = sub _id ; HRV _U ; test^ ; symptom _id ; SYS _BPi ; DIA _BPi ; f F. Predictive dataset

In order to develop and test a model to predict the magnitude of the BP drop, a second dataset was generated. The new dataset, hereinafter referred to as the predictive dataset, contained for each subject four instances defined as follows: i = sub _id ; HRV _id = 2 ; test _id ; symptom^ ; ASYS _BP ; ADIA _BP ; /I

where:

(2.3) and

In other words, the predictive database contained the HRV measurements extracted from the 5 minute ECGs recorded in Phase 2 (while lying) associated with the drop of systolic and diastolic BP detected in Phase 3 (after standing). The BP drop (ΔΒΡ) was calculated as the difference between the BP value measured during the 2 ^nd minute after the standing action (BP7) and the mean of the BP measurements taken in Phase 2. Since Phases 2 and 3 were repeated four times, four instances were generated for each subject and the final complete dataset was composed of 40 instances (4 instances ^* 10 subjects).

The BP7 measurement was used because it has been demonstrated that the BP drop is more frequent (observed in about 46% of patients) within 3 minutes after the standing action. Since the BP measurement device reported a nominal error (NE) of ±3 immHg, the measurement error of the ΔΒΡ was estimated as follows:

NE

BP _emr = NE + - _r

^ ⁿ (2.5) where n is the number of repeated measurements. Therefore, an error of ±4.5 immHg was estimated on the measured ΔΒΡ.

G. Predictive model

A robust multi-linear regression [Huber PJ: Frontmatter. Robust statistics, John

Wiley & Sons, Inc; 1981, i-xi] was used to develop a mathematical model to predict the ΔΒΡ by using the subset f of HRV features.

As stated above, the hypothesis was that the drop of BP after standing could be predicted by using the HRV measurements extracted from the ECGs registered in the 5 minutes before standing. Therefore, we modelled the ΔΒΡ as a combination of n HRV features (Equation 2.6):

BP = c ₀ +c _l +... + c _n + _{(2 6)} where:

• Co was the intercept and c-i.. _n determined the contribution of the independent variable f-i ... _n .

• ε was a random variable normally distributed.

The model was developed by using the robustfit function of the MATLAB version R2013a [18,19].

H. Feature selection and performance evaluation

The best subset f of HRV features was selected using the so-called exhaustive search method [Jain et al, IEEE Transactions 2000, 22(1):4-37\, investigating all the possible combinations of k out of N features (with k from 1 to n). Since the number of HRV measurements was relatively high considering the number of subjects, we limited the value of n to a maximum number of 5.

The best subset was selected as the one that minimizes the Regression Standard Error a _es t, defined as follows: where: Y was the actual value (the observed BP drop), Y' was the predicted value (the predicted BP drop), and N was the number of instances utilised to develop the model. The numerator was the sum of the squared differences between the actual scores and the predicted scores. Therefore, this error represented the average distance between the real BP drop values and the regression line. Consequently, Oest reported on how wrong the regression model was on average using the same units of the predicted value. Smaller a _es t values were considered better because they indicated that the prediction was closer to the fitted line. The a _es t was estimated using the leave-one-subject-out approach. Therefore, the robust regression was performed ten times, each time using nine subjects to develop the model and one to test it. Consequently, 10 values of a _es t were computed and the average of these 10 values was used to choose the best feature combination. For this combination of features, the final model was then fitted on the whole dataset (all the 10 subjects).

Moreover, the percentage of correctly predicted values (%CP), i.e. those with an error lower than the measurement error of the BP drop (4.5 immHg), was

computed as follows: (2.8) where Θ represents the Heaviside function [Legua et al, Proceedings of the international conference on Computational Science and Its Applications, Part I, Perugia, Italy 2008, 1212-1221, Abs.].

Finally the rate of false positives and negatives was calculated according to the following conventions:

• "false negative" - the cases where the ΔΒΡ was underestimated with an error above 5 mmHg (i.e. predicted ΔΒΡ < measured ΔΒΡ - 5mmHg)

• "false positive" - the cases where the ΔΒΡ was overestimated with an error above 5 mmHg (i.e. predicted ΔΒΡ > measured ΔΒΡ + 5mmHg)

II. RESULTS

A group of 10 volunteers, 7 women and 3 men, with a median age of 30.4 years (range 23-43 years), was enrolled. All the subjects signed informed consents and met the inclusion criteria. Each subject underwent 33 BP measurements: one in Phase 1 ; 16 in Phase 2 (four times 4 measurements); and 16 in Phase 3 (four times 4 measurements). For each subject 42 minutes of ECG were recorded: two minutes in Phase 1 , 20 minutes in Phase 2 (four times 5 minutes); and 20 minutes in Phase 3 (four times 5 minutes). The final predictive dataset contained 40 instances as the protocol was repeated 4 times for each subject. Using these instances the best model identified was the one developed using the following five features: RMSSD, NN50, TINN, HF, and RPDET. The model is reported in Equation 2.9:

ΔΒΡ = -25.67 + 0.45 * (RMSSD) - 0.05 * NN50

(2.9)

- 0.02 * (ΉΝΝ) - 0.01 * (HF ) + 0.3 * (RPDET )

The regression standard error calculated on the testing (leaving one out estimate), the training and the whole dataset was respectively 5.22 mmHg, 4.30mmHg and 4.29mmHg.

Figure 2 reports the values of ΔΒΡ predicted by this model (crosses) against the measured values (circles) and the measurement error (vertical error bars). The vertical dashed lines divide the four instances of each subject. The

percentage of correctly predicted values (error below 4.5mmHg) was 80%, indicating that in 32 of 40 experiments, the ΔΒΡ was predicted with an error below the measurement error of the sphygmomanometer.

As shown in Figure 2, in 3 measurements out of 40, the measured ΔΒΡ was underestimated with an error above 5 immHg, resulting in a false negative rate of 7.5%: subject 2, fourth measurement, subject 6, fourth measurement, and subject 8, fourth measurement. Finally, in 4 measurements the ΔΒΡ was overestimated, resulting in a false positive rate of 10%: subject 6, first measurement, subject 7, first measurement, subject 8, second measurement, and subject 10, third measurement. Another model identified was one developed using the following five features: SDNN, pNN50, HF, SD ₂ and RPDET. This model is reported in Equation 2.10:

ΔΒΡ = -219.39 - 1.83(SDNN) +

+ 0.23(p N50) + 8.5 * lCT ³ (HF) + ( ² - ^{1 0} )

+ 0.97(SD ₂ ) + 2.34(RPDET) The regression standard error calculated on the testing, the training and the whole dataset were respectively 5.34mmHg, 3.88mmHg and 3.87mmHg.

III. DISCUSSION

This study presents a model to predict systolic BP drop due to standing that relies on HRV measurements extracted from 5 minute ECGs recorded before standing. The proposed model is based on the hypothesis that the magnitude of BP drop in the few minutes after standing is correlated with HRV features extracted from ECG recordings taken in the 5 minutes before standing. According to the mathematical model proposed here, an increased ΔΒΡ appears to be related to:

· a reduced overall short-term HRV, as assessed by TINN and RMSS. • an increased value of HF and/or NN50, which is strongly correlated to HF.

• an increased determinism (as evaluated by RPDET), and, therefore, a decline of non-linear heartbeat dynamic.

As HF is an indicator of parasympathetic function, while the overall variability is considered a marker of both sympathetic and parasympathetic tones, the combined association with a reduced overall variability and increased HF suggests that an increased ΔΒΡ could be associated with a reduction in the sympathetic tone. Mechanistically, it is assumed that the upright position gravity causes venous pooling in the lower extremities, resulting in a reduction in systemic blood pressure. This should result in a baroreflex activation, which leads to an increased sympathetic tone. Sympathetically mediated peripheral vasoconstriction and tachycardia should then occur in order to maintain the Blood Pressure value. In the case of a decreased sympathetic tone, this mechanism could fail to maintain the blood pressure value and result in higher values of ΔΒΡ.

The findings of the current study are in agreement with the results of Example 1 , which found a significant association between a depressed HRV and the risk of falling in the elderly.

These findings are intriguing because they suggest that a reduced overall short- term HRV might represent a predictive parameter for SH, which is one of the main causes of indoor falls. Previous studies aiming to prevent falls used wearable accelerometers, pressure sensors, ambient sensors, or a combination of these three technologies, which had no other direct benefit for later life problems. In contrast, the model proposed in this study is based on features extracted from HRV, which has been associated with a number of other health outcomes.

Therefore, the clinical implications of these findings are potentially relevant, since these parameters are based on simple and non-invasive measurements.

Moreover, recent systematic reviews, investigating the independent capability of different technologies to prevent falls, have highlighted their limitations, in particular, the rate of false alarms (16%), which is too high to maintain the full attention of the nursing staff. The model of the current study is estimated to achieve a lower false positive rate (10%) and could be enhanced by the addition of other sensor information, for example, accelerometric signals and breath rates, which are already acquired by the adopted wearable device.

IV. CONCLUSIONS

This study suggests that the lowering of blood pressure in the few minutes after standing can be predicted by monitoring HRV features recorded in the 5 minutes before standing. Particularly, more significant blood pressure decreases were observed in subjects presenting a depressed and less chaotic HRV pattern. This could reflect a transient depressed autonomic response causing a slower recovery of homeostasis and adjustment of blood pressure.

EXAMPLE 2(B)

VALIDATION OF MODEL USING SHORT TERM HEART RATE VARIABILITY TO PREDICT BLOOD PRESSURE DROPS DUE TO STANDING

This study concerns the testing of the model developed in Example 2(A) in a different research centre and by an independent team of researchers.

I. METHODS

Five healthy subjects were enrolled in the study, using the same inclusion criteria as are set out in Example 2(A). The experimental protocol was as for

Example 2(A), and ΔΒΡ was measured and HRV data was generated as for Example 2(A).

For each subject, ΔΒΡ was predicted on the basis of the HRV data, using

Equation 2.9. The measured and predicted ΔΒΡ values were recorded for each experiment (20 experiments in all: 4 for each of the 5 subjects). II. RESULTS

Figure 3 shows the measured values of ΔΒΡ after standing (circles), together with the values predicted by Equation 2.9 (crosses) with the measurement errors (vertical bars).

III. DISCUSSION

The results shown a correct prediction of 80% of the validation set, confirming the results obtained with the training/testing dataset. This value was calculated as the % of prediction that differ from the measured ΔΒΡ less than the measurement error (4.5 immHg) of the sphygmomanometer. These results confirm that it is possible to predict ΔΒΡ due to postural hypotension using HRV.

EXAMPLE 3

DESIGN AND PRELIMINARY VALIDATION OF A SMART HEALTH MONITORING SYSTEM (SHMS) FOR FALL RISK ASSESSMENT IN

HYPERTENSIVE PATIENTS

This study describes the design and the preliminary validation of the SHMS developed in the framework of the EU-funded research project "Smart health and artificial intelligence for Risk Estimation" (SHARE). Results of the automatic classifiers, developed in a controlled environment, are described in Example 1 . In this Example, the design of the overall system and the preliminary validation in real life scenario will be presented. I. MATERIALS AND METHODS

A. System Architecture

The general system architecture of the SHARE system, integrating recording device and a Cloud infrastructure, could be described according to a three-layer model, as shown in Figure 4. The user-base layer has two types of

interface: web-base and client-base. On the web base, users (mainly physicians) can connect to the server via HTTPS protocol and need only a web browser to log into the system, and access data; on the client base, two different software interfaces were developed for signal acquisition: ShareAPP for Android Smartphones and ShareLogs for Windows PC, to be used by patients and physicians, respectively. The function-base layer consists of a Web interface and Web Services to provide all the functionalities of the system. The data-base layer consisted of a relational database to store all the signals and data recorded.

In the architecture design, the following considerations were considered:

a) user comfort, data transmission and power consumption, data quality: these criteria motivated the choice of a commercial wearable device, already tested in real-life scenario, with a long battery and no electrodes, which are not well accepted;

b) security and privacy (secure data transmission): the cryptographic protocol Transport Layer Security was adopted for all the network communications containing sensible data (i.e. health data associated with the name of the subjects);

c) area of movement: in order to not limit area of movement, the wearable device is required to have a long-term storage, so that there is no need for an always-on connection of the device with the smartphone and of the smartphone with the server;

d) scalability: a scalable framework was set up to support the processing of multiple data streams for concurrent application services, by adopting a Cloud environment providing storage and Virtual Machine (VM)-based approach for computational process.

In the following paragraphs, each component of the systemis described in more detail.

ECG and actigraphy were registered using the Bioharness™ (vers. 3 BH3, Zephyr Technology), a state-of-art commercial wearable multi-sensing device, which enables long-term recordings of several biomedical signals and data. The monitoring device acts as a data logger and/or transmitter, has a memory of up to 480 hours and battery life of up to 24 hours. ShareAPP (referred as Share Cardio Health in Google Play) is an Android application, developed to provide a user interface for patients, for real-time transmission of the data acquired by BH3 though a smartphone. The App was designed in order to minimize user interaction. Moreover, it enables the physician to submit a daily questionnaire to the patients. The ShareAPP was implemented by using the BioHarness API and Example Android Project available in the BioHarness Blueetooth Developer Kit provided by Zephyr Techonology

(http://zephyranywhere.com/zephyr-labs/development-tools/ ). ShareLogs is an ad hoc developed Windows standalone application for the upload of all the acquired signals by BH3. It has been designed in order to follow the physician's usual workflow: at each planned visit, the physician uploads all the data stored on the BH3. This avoids the risk of losing the data that were not stored in real-time (e.g. because the mobile device was offline or out of Bluetooth coverage area or network problems). The ShareLogs was developed by using the BioHarness Bluetooth Logging System Interface available in the BioHarness Blueetooth Developer Kit provided by Zephyr Techonology

(http://zephyranywhere.com/zephyr-labs/development-tools/ ). The Web Interface consists of a Content Management System (CMS) to show all the public information on the Project and a Restricted Area reserved to the system users (i.e. physicians, researchers, patients). The CMS relies on Wordpress while the Restricted Area application was developed in ASP.NET (C#) by using Visual Studio 2013 and MySQL.

The Web Services represent the software interface to store the data, acquired by BH3 and transmitted though the ad hoc applications. Moreover, they provided the most advanced functionalities of the system, including the remote processing and data mining of HRV.

The remote processing, particularly, the HRV analysis, was performed by adapting and integrating public libraries, in particular from the physionet.org website and from an open-source Matlab-based HRV software toolkit [Ramshur J. Design, Evaluation and application of Heart rate variability software. 2010], in agreement with the GNU free software redistribution license. The integration in the web server required some reediting of the library functions, mostly regarding the input and the output data format. The reediting was performed using MATLAB version R2013a (The MathWorks Inc., Natick, MA). The processing output was validated comparing the server output and the values obtained by a stand-alone HRV software [Niskanen et al, Computer Methods and Programs in Biomedicine.

2004;76(1):73-81]. Finally, the Web integration was achieved compiling the MATLAB functions as .NET objects.

The ECG recordings could be analyzed concurrently in segments of user-specified length. HRV analysis was performed as described in Example 1 . Likewise, data- mining was carried out as described in Example 1 . B. Validation procedure

The platform validations was based on observational clinical trials. The results of a trial performed to develop and test the remote processing and data-mining modules are described in Example 1 . A study was also designed to assess the compliance of the patients, based on the following measures: the ratio between the real length of the recordings and the expected length; the number of non-adherences due to patients failing to use the device for the monitoring period; the number of calls from health operator to technical assistance and those from patients due to technical reasons.

II. RESULTS

The SHARE system for Remote Processing of Heart Rate Variability and Data- mining has been designed and deployed through a web portal. Figure 5 shows a sample screenshot of the results of HRV remote processing from the ECG of one enrolled patient. Figure 5a shows time-domain HRV results, Figure 5b reports frequency-domain HRV results, while Figure 5c shows non-linear HRV results. Figure 5d reports the results of data-mining methods for cardiovascular risk assessment. The data-mining modules have been trained, tested and validated in a

retrospective study. For faller identification, among 168 hypertensive patients (including 57 female and 1 1 1 male, age 71 ± 8 years), 47 subjects experienced one fall within three months from the recording. No statistically significant differences were assessed between the two groups (fallers and non-fallers subjects) in terms of clinical and demographic parameters, except smoking habits. The best algorithm for fall risk assessment achieved an overall accuracy of 72% with an AUC of 67.6% (sensitivity rate: 51 %; specificity rate: 80%), as described in Example 1 . Finally, 36 patients (aged 60 ± 3 years; 14 women) have been enrolled in the prospective clinical trial conducted to test the compliance with the SHMS in a real environment. All the patients were instructed to wear the wearable devices and invited to use them for a week. Moreover, as for clinical Holter recordings, they were asked to fill a brief report. The preliminary results of the patient compliance are reported in Table 3.2. Only one patients failed to complete the expected monitoring period and performed a call to the health operator for technical reason. The other patients did not report any particular problem in using the provided devices, although four of them reported that the BH3 they received worked for a shorter time than expected. A shorter battery duration was confirmed for these devices. The wearable device recorded the signals for about 80.4% of the expected time, meaning a data loss of 19.6%, mainly due to short battery duration of some devices. Table 3.2

Performance measurement of the SHARE platform (Patient and Physician

Modules)

III. CONCLUSIONS AND DISCUSSIONS

The design and preliminary validation of a SHMS with advanced remote

processing and data-mining functionalities were described in this study. As regards signal acquisition, the SHMS relies on a commercial multi-sensing wearable device. The most advanced functionalities, i.e. ECG processing and automatic classification, were provided by the centralized cloud-based structure of the system, and the users, i.e. physicians, need only to have a Web browser running in a personal computer and a network connection to access these services. The system can be updated and new features can be easily added without interfering with the medical users. The addition or incorporation of a new tool in the GUI can be a quite simple task, by adding a link to the function that runs under MATLAB and performs the corresponding processing. This fact makes the system into an open structure that can easily incorporate new tools as soon as they are developed, and therefore have an immediate presence in the support of clinical diagnosis.

The performances achieved by the classifiers suggest that a clinical decision support tool, processing tele-monitoring data, could contribute to a quicker and possibly more accurate clinical assessment of the patients. The relationship between HRV and risk of falls has been demonstrated in Examples 1 and 2. Moreover, the system appeared to be well accepted by almost all the patients (95%) with a limited amount of data lost (<20%), thus showing that the electrocardiographic data could be reliably acquired by the system.