Biological age estimation using an eHealth system based on wearable sensors

The population of all countries of the European Union (EU) has significantly changed over the years. The rise of life expectations, the increasing health care cost, the caregiver burden and the importance given by people to independent living, justify the development of new assisted life technologies aiming at a safer and self-contained aging. Moreover, the only way to forecast and delay incoming disabilities is the adoption of preventing measures for these people who have higher risk (Cesari et al. 2013). The EU is focused and sensible to these issues linked to the aging society, sharpening the necessity of fast and accurate screening tests for those age-related diseases, some of which are particularly hard to be spot through today’s diagnostic modes. Anyway, it is tricky to define how much someone is getting older compared to her/his contemporaries, because life expectation changes from community to community reflecting a mix of genetics, epigenetics, environmental and random factors. Aging is indeed a complex biological process, characterized by the progressive cell and organs deterioration that in turn causes the decline of biological functions (Weidner and Wagner 2014). The most effective mean to reduce disease burden and to control costs is to delay this progression by extending the health span, the period of human life during which the individual is healthy and free from serious diseases, hence the urgent need for biological measures of age.

Indeed, the human body has two different ages, the chronological age (CA) and the biological age (BA). The former depends on the birth date while the latter represents the age based on the biological quality and functioning of tissues, apparatus and organs of an individual. The CA is not a reliable indicator of the physiological decline, such that people with the same CA could show significant differences in their health status (Levine 2012). For this reason we need to take into consideration the BA to accurately forecast the course of diseases and the mortality rate. Therefore, knowing a precise BA index for each person is extremely important to be able to identify aging sources, to estimate rejuvenation therapies and to beforehand intervene. The BA is commonly associated to the “real” human age because it reflects the health status and the psycho-physical deterioration of a body. It is capable of detecting individuals at risk and of forecasting disabilities or even death, reason why this concept is so dear to the scientific community (Karasik et al. 2005).

Over the time several studies have tried to find a method to estimate the BA, but still today we lack a universally trusted and accepted golden standard. Nowadays, main methodologies employed to esteem the BA are based on the measurement of biomarkers that are strictly linked to the CA. The World Health Organization (WHO) defines the term biomarker in a broad sense as a word that “includes almost any measurement reflecting an interaction between a biological system and an environmental agent, which may be chemical, physical or biological”. Nevertheless, the complexity of the aging process has made impossible the identification of a unique biomarker able to give an accurate measure of the biological aging rate.

Scientific literature provides several possible methods and algorithms that exploit biomarkers to estimate the BA, such as the multiple linear regression (MLR) and the principal components analysis (PCA). Recent studies added new molecular biomarkers based on the DNA and RNA to the classical phenotypic biomarkers usually adopted in scientific research. Among these new biomarkers, we find as a trustworthy predictor of BA, the DNAmAge (also said epigenetic clock) that measures the age of the DNA’s methylation. Nowadays, the most powerful predictors of the BA are two of these molecular measures, namely the Horvath’s clock and the Hannum’s clock (Jylhävä et al. 2017).

Marioni et al. (2016) has been the first to identify association between blood DNAmAge measurements and health status during the aging. In their study they found the cross association of the Horvath’s clock with the worsening of the psycho-physical conditions of a subject (Marioni et al. 2015). Breitling et al. (2016) has detected a direct association of the Horvath’s clock with the increase of the psycho-physical deficits owned by a person. The correlation between BA and deficits accumulation has fostered the definition of two alternative methods to estimate the BA (Searle et al. 2008). These methodologies take into account the psycho-physical conditions of a subject and estimate the BA thanks to the analysis of owned deficits and the associated assessment of the frailty. Frailty is commonly described as a state of higher vulnerability to adverse events (Searle et al. 2008) and, in the geriatric medical practice, it is evaluated through deficits count.

Even though the DNAmAge methodology is extremely accurate, it is invasive and expensive too. A non-invasive approach to estimate the BA in elderly people is the one built upon the frailty evaluation, that is commonly made using very complex procedures due to the necessity of continuous monitoring of the psycho-physical conditions of the subject. Anyway, technological progress can support in dealing with the just mentioned issue. Thanks to the information and communication technologies (ICTs) and to the development of wearable and non-invasive sensors, it is possible to reach the 24-h monitoring of subjects still allowing them to do things of daily life without annoyance. Moreover, in this prospective scenario, the doctor would be able to remotely supervise the patient accessing to the gathered data and checking for any variations through the cloud service (Manas et al. 2018). This kind of application is included in the ambient assisted living (AAL) solutions, a flourishing field of research and development. The AAL is born due to the global aging of the population and to the necessity to help elderly persons in living as much as possible in their favourite environment, increasing their self-sufficiency, monitoring their actions and providing assistance (Calvaresi et al. 2017).

Our research project arises in this context, with the aim of finding a non-invasive methodology for the BA evaluation in elderly people. In particular, we propose a complete eHealth system subject based on commercial devices, wearable sensors, smartphone application, and cloud services, usable in both outpatient and home monitoring. Below we are going to present the characteristics of the developed eHealth system, that supports the frailty assessment procedure thus allowing the BA evaluation.

It is not easy to understand directly the BA of a subject, as previously pointed out, reason why largely diffused methodologies to compute a reliable measurement of BA start from the frailty estimation (Karasik et al. 2005). Conceptually, frailty is a diagnosable state, typical of elderly people, characterized by an increased vulnerability, caused by age-related declining that undermines daily capabilities. Specifically, frailty could even be present without any other detected pathological condition (Cesari et al. 2013). Several studies have theoretically addressed this topic, ending up that frailty is a tangible geriatric syndrome, characterized by low reserves and poor resistance to stressors, associated with an higher risk of bad outcomes for the subject’s future health (Ensrud et al. 2018).

In literature, the most exploited methodology of BA estimation among the frailty-based approaches, is the one grounding on the frailty index (Mitnitski et al. 2001). The frailty index “q” can be quantified after a complete geriatric evaluation of the patient, because it depends on the count of current deficits through multiple domains (cognition, mood, mobility, chronic diseases and so on). In particular, it is computed as the ratio between current deficits of a subject respect to the total number of deficits that have been taken into account during the geriatric evaluation (Searle et al. 2008). The study carried out by Rockwood and Mitnitski (2007) estimates the biological age of a group of Canadian over-65 people, exploiting the Frailty Index computed considering a total amount of 20 deficits. Starting from the average values of the frailty index and using the linear regression, Mitnitski obtained a mathematical model and extracted from it the log-linear behaviour of the frailty index in relation to the CA of the subject (Mitnitski et al. 2002). The mathematical model described by Mitnitski, depicted in Fig. 1, matches the following equation:

$$\begin{aligned} \ln (q) = -\, 4.23 + 0.03 \cdot CA. \end{aligned}$$

(1)

From the mathematical model, it is possible to observe that the increasing count of deficits that turns into an higher Frailty Index for older subjects (higher CA) (Fig. 1). Applying an inverse regression to the Mitnitski’s model, we can derive the personal biological age (PBA) employing the following equation:

$$\begin{aligned} PBA=126 + 26.09 \cdot \ln (q), \end{aligned}$$

(2)

where PBA represents the average value for which certain deficits are present in the reference group of people ageing well.

Open image in new window

The frailty index methodology, although precise and accurate for its detailed risk scale, has some drawbacks, such as the large number of deficits to be taken into account (at least 30/40), the lack of an unequivocal definition of them (Bergman et al. 2007; Rockwood and Mitnitski 2007) and the dependency of the estimation on a complete geriatric evaluation, something that can be carried out by professional medical staff only (Guaraldi et al. 2017).

The frailty phenotype (FP) has been largely investigated and validated by scientific researchers for its capacity of being reproducible in the clinical practice and because it can be evaluated even without the presence of professional medical staff. Moreover, in agreement with what Cesari et al. (2013) has highlighted, the FP may be more suitable than the frailty index for an immediate identification of non-disabled elders at risk of negative events.

Open image in new window

Our research aims to propose a non-invasive method for biological age evaluation through the FP assessment of a subject. In our previous study (Pierleoni et al. 2018), we have proposed a methodology to estimate the biological age based on a FP assessment carried out according to Fried’s protocol. In this work the data relating to FP assessment were compared with those obtained from the study of Theou et al. (2014) in which the characteristics of frailty have identified, taking into account the frailty phenotype assessment measured of a sample of 27,527 Europeans. However, the study of Theou et al. (2014) only derived the Frailty Phenotype of Europeans without going deeper into the abstraction of a mathematical model useful to estimate the BA of a subject. Otherwise, in our previous work we proposed the application of linear regression techniques to estimate the biological age from data relating to FP assessment based on Fried’s protocol. In fact, accessing a certain amount of FP data regarding a reference sample, it is possible to infer a model similar to the one obtained with the frailty index as depicted in Fig. 2. The model obtained from our previous work (Pierleoni et al. 2018) and represented in Fig. 2 can be exploited to identify the BA starting from the evaluation of the FP. The Fig. 2 shows that the point A (outlined by the dashed orange line) can be found starting from the CA and the natural logarithm of the frailty phenotype of the subject. The BA (outlined by the dashed black line) is determined from the abscissa of the point corresponding to A in the model line. In this study we propose a protocol for the BA evaluation of a subject based on the adoption of an eHealth system able to compute the frailty phenotype of the elderly people and provide the biological age of a subject. Respect our previous study, the proposed protocol includes the use of wearable sensors and a smartphone application for the evaluation of the mobility of an elderly subject and the relative phenotypical criterion. Starting from the data acquired in the proposed protocol, the developed eHealth system is able to provides the FP assessment and the BA estimation of a elderly subject, taking into account the FP measured on a reference sample of different age. The results of this work show the FP assessment and the BA estimation providing a comparison between the results obtained from the proposed eHealth system, the outcome of our previous work (Pierleoni et al. 2018) and those derived analyzing data coming from the study of Theou et al. (2014).

A new method to the BA evaluation made through the frailty phenotype analysis is proposed in this work. It has the main advantage respect to the frailty index of taking into consideration only five simple criteria, that still provide a good groundwork for a standardized screening and for the frailty evaluation. To this end, we developed an eHealth system to allow the frailty phenotype measurement on a reference aged population and the subsequent extraction of a mathematical model able to estimate the biological age. Having frailty phenotype data available, regarding a sample population, and exploiting linear regression techniques, it would be possible to extract a robust model able to estimate the BA of each subject starting from the frailty phenotype assessment, as just described in the previous section.

The eHealth system has been created with the aim of making the frailty phenotype evaluation as much objective as possible, and of supplying a user-friendly and automatic solution for the BA estimation to the customers. The method of administration, the results significance and the adoption of efficient supporting tools are essential components to create a reliable method of the BA estimation that starts from the frailty phenotype evaluation. For this reason, a complete eHealth system has been developed, that is able to take in from different sources the information useful for the frailty phenotype evaluation, to efficiently store data and to allow information analysis thanks to the online upload. Indeed, exploiting cloud computing techniques, the system will enable the medical staff to remotely monitor patients, define personalized treatments and diagnose diseases before they appear.

Open image in new window

The choice of developing a cloud application comes from the necessity of having a system able to scale and to be reliable, flexible and apt to interoperability. We used MEAN, a fullstack JavaScript framework for Web applications whose components are MongoDB, Express, AngularJS and NodeJS. The application developed in this work is exploited to store the history of medical reports and the history of data acquired through the evaluation protocol. Managing these data, it is able to automatically compute results for each criterion (weight loss, poor endurance and energy, grip strength, mobility and physical activity) and to determine the frailty state. This architecture has to manage and store a huge amount of clinical data represented in the Health Level 7 (HL7) format, that is the worldwide accepted standard for the clinical data sharing (Orgun and Vu 2006). For this reason, we use MongoDB, a NoSQL open source database that provides support for systems of JSON with document-oriented style. In order to store and analyze HL7 messages in MongoDB, we need a conversion from XML to JSON format, while for the elaboration of data coming from sensors a conversion from CSV to JSON is required.

The system architecture could foster the evolution in other connected areas, e.g. tele-medicine and tele-rehabilitation. Moreover, together with the wearable sensors included in this work whose efficiency has been already tested, it is possible to add to the system sensors such as the environmental ones, for a constant monitoring and for an higher control on the subject at home.

The problems related to the mobility of a subject are very common in elderly people due to age-related physiological and musculo-skeletal system changes and to the higher presence of different diseases which negatively affect the postural stability. Within the medical framework, gait and mobility are commonly assessed through simple questionnaires administered by physicians (Tinetti 1986). These kind of investigations do not require sophisticated equipment and are easy to be done, however they have the great disadvantage of being subjective, assessor’s experience dependent and of lacking specific gait parameters. The only way to objectively and accurately perform a gait analysis today, is to make the assessment inside specialized labs equipped with expensive and complex measurement systems.

The spreading of wearable devices is linked to countless advantages, as for example their potential exploitation outside of the hospital, the low-cost and the possibility to evaluate more than one subject at once. Moreover, thanks to specific algorithms of data fusion, it could be possible to measure and analyze the gait’s kinematics (Aminian 2006). These wearable devices have been exploited in several gait analysis techniques, such as monitoring gait abnormalities and falls-detection thanks to their light weight, smallness, affordability and energy-efficiency (Shin et al. 2007).

The adoption of wearable devices for the acquisition and elaboration of data regarding the mobility is very important for taking care and supporting elderly people (Rockwood and Mitnitski 2007) and therefore also for the evaluation of the frailty phenotype. The use of systems helping the mobility analysis for the frailty phenotype evaluation has been deepened from Schwenk et al. (2015), that for the first time exploited wearable sensors with the final aim of estimate the frailty phenotype. Even though they made an accurate analysis of frailty exploiting wearable sensors, their study do not provide any estimation of the aging process from the results regarding the frailty phenotype. For this reason our aim is to exploit wearable sensors for the FP evaluation and, finally, correlate these results with the biological age. This study is not limited to determining FP but also does intend to provide an estimation of the BA of a subject. In fact, the BA estimation allows to assess the real health state of a subject as it represents the most important indicator of physiological decline.

Based on our previous works (Pierleoni et al. 2014b, 2015), we developed a wireless, small and non-invasive system, to help in the mobility assessment of the frailty phenotype. It is mainly composed of a smartphone and wearable sensors capable of carrying out a mobility analysis of a subject based on Schwenk’s indications. Placing the wearable sensors on subject’s instep through an elastic, the developed system is able to extract useful metrics of the walk, such as stride length, stride speed, percentage of double support and hip angle. The device is made by triaxial gyroscope, accelerometer and magnetometer, a microprocessor and a unit for the storage and transmission of information. Figure 4 depicts the hardware and the block diagram of the developed system. The microprocessor gathers and processes data coming from the sensors, that can be saved in an SD card and conveyed through the Bluetooth interface to a smartphone. A data fusion algorithm, able to precisely estimate the acceleration and the orientation of the body segment on which they are placed, has been implemented in the microprocessor (Pierleoni et al. 2014a) together with a sensing-unit calibration algorithm.

Open image in new window

In order to elaborate the data acquired by the wearable sensors, a specific smartphone application has been developed. Handling acceleration and orientation data collected by the wearable sensor, the smartphone application automatically runs the algorithm implemented for the walk analysis. In particular, through acquired data from the wearable sensors, the developed application estimates three main parameters useful to the mobility assessment: double support, stride speed and stride length (Schwenk et al. 2015). The walked length is obtained through the double trapezoidal integration of the acceleration signal along the direction of the walk. To properly compute the values of speed and distance, two subsequent filtering phases and two integration phases are necessary. Specifically, the acceleration signal is filtered through a Butterworth high-pass filter of the second order. Speed is then computed as the cumulative integral of the acceleration signal adopting the trapezoidal method, and distance applying a second high-pass filter followed again by the trapezoidal integration. The root mean square (RMS) of the acceleration vector is computed and then its peaks can be determined too. The stride detection is allowed by the fact that each peak corresponds to a stride. Once determined the number of stride, the phases of stance and of swing are evaluated thanks to the determination of local minimums. The relevant points are the initial contact, that is where the stance phase begins and corresponds to the minimum before the highest peak found, and the toe-off, beginning of the swing phase that corresponds to the minimum after the highest peak. The stride length is computed as the distance between two subsequent initial contact belonging to the same foot, once identified the stance and swing phases. Stride speed is then computed as the average for each stride. Finally, the percentage of double support is computed taking into consideration the portion of the stance phase while the two feet are on the ground.

The study of Theou et al. has been taken as a reference for the validation of our methodology. Their study is the only one that evaluates the Frailty Phenotype employing a reference population comparable to the one considered by our work. Indeed, we evaluated a sample of Italian elderly people, that is a subset of the European aged population considered by Theou et al. in their research.

Figure 5 depicts the comparison of the obtained results. The blue curve represents the frailty phenotype model emerged from Theou’s work, the red curve shows the results elaborated by our cloud application following Fried’s definition, while the green curve is related to our results adding gait parameters determined thanks to the wearable sensors. It is possible to notice a discrepancy between our results and that from the reference article. This may be caused by the small number of evaluated subjects in this study, or may be related to the better fit embedded in the model found by us that takes into consideration also specific mobility parameters. This last hypothesis should be proven only acquiring more data.

Open image in new window

Open image in new window

As previously highlighted in Sect. 2, from the Frailty Phenotype data processing, we can extract a model to estimate the BA. Figure 6 depicts the models obtained from our results together with the reference model.

This paper presents a non-invasive methodology to establish the biological age of elderly people through the frailty phenotype evaluation. The proposed solution is composed of an eHealth system useful to evaluate the five phenotypic criteria, already present in literature, with the additional employment of sensors. In detail, we developed small and non-invasive wireless wearable sensors for the assessment of the mobility criterion of the frailty phenotype. Thanks to the integration of the gait analysis algorithm into the eHealth system, we could compute three main parameters, that are double support, stride length and stride speed. Moreover, the cloud application built is able to compute the frailty phenotype from the results of the frailty tests and finally estimate the BA. The performed evaluations are then stored in the cloud to provide a time trend of the subject’s aging.

The set method has been validated analyzing data about 15 elderly subjects. Results show that our solution to estimate the BA is comparable with other solutions already existing in literature. However, an extensive acquisition campaign will facilitate a greater data gathering, thus allowing us to better validate our BA estimation model.