MENTAL HEALTH DECLINE PREDICTION

A Smart Sensor for Day to Day Activity Recognition

Thomas Kaegi-Trachsel, Juerg Gutknecht

Native Systems Group, Computer Systems Department, ETH Zurich, Switzerland

Dennis Majoe

MA Systems and Control Limited, Southampton, U.K.

Keywords: Mental health, Depression, Activity sensing, Wearable Smart sensor.

Abstract: The ambulatory activity of a person may be used as one component within an overall wearable sensor

system that predicts the onset of mental health problems. Ergonomic smart sensors that can determine the

total energy expenditure and type of ambulation may provide unique insights to the coping behaviour of

stressed people. Rather than relying on wearable computers, a single smart miniature sensor that is worn

24/7 should perform the complex embedded recognition tasks while meeting difficult battery life, wireless

communications and ergonomic constraints. The development and testing of such a smart sensor is

described which takes into account action timeline variations, as well as action variations both intra

individual and inter individual.

1 INTRODUCTION

This work relates to an activity recognition sensor

developed within the EU research project OPTIMI.

The project’s aim is to provide on-line predictive

tools for the early identification and intervention

during the onset of a mental illness, in particular

depression, following the inadequate coping with

day to day stress.

Second only to depressed mood itself, tiredness,

low energy and listlessness are the most common

symptoms associated with depression (S. M. Stahl,

2002). Energy loss in depression is the factor that

correlates most strongly with lost productivity and

lack of social functioning.

The Psychological Counselling Centre, which

acts as the primary support for students at the ETH

Zurich and University of Zurich, recognize that

students who are significantly traumatized by

examination stress and border on mild depression

will disengage from social activities, sports and any

events requiring physical involvement. Group

activities are avoided, and the student will prefer to

stay at home and do nothing.

The exact relationship between activity and

depression is not clear. In a study on 956 Japanese

men diagnosed with metabolic syndrome (T.

Takeuchi, 2009 ) it was shown that the deposition of

fat around the waist line, was a predictor for the

onset of depression. By studying the behaviour of

the control group the conclusion was that a healthy

lifestyle that involves regular exercise alleviates

depression while conversely a sedentary lifestyle

may increase the visceral fat of individuals with

metabolic syndrome, thereby increasing the risk of

depression. In other related work (A. Berlin, 2006) it

was shown that an enforced reduction in daily

exercise resulted in symptoms of depression.

It is very difficult to find conclusive research that

depression causes a loss of energy and activity.

However what seems to be widely agreed is that

activity and sport has a positive effect in reducing

depression and a reduction in exercise is associated

with the onset and ongoing depression. The quantity

and type of exercise that one performs per week that

might predict the onset of depression has not been

studied to our knowledge and should be determined

as one outcome in the OPTIMI project.

In the field of physical activity research, see the

Compendium of Physical Activity (B. Ainsworth,

2000), there exists an alternative to measuring

energy in Calories or Joules. For example one MET

219

Kaegi-Trachsel T., Gutknecht J. and Majoe D..

MENTAL HEALTH DECLINE PREDICTION - A Smart Sensor for Day to Day Activity Recognition.

DOI: 10.5220/0003150002190227

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2011), pages 219-227

ISBN: 978-989-8425-34-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

(for 1 unit of metabolic rate) corresponds to the

typical energy consumption when at rest. Running at

17Km/h has been found on average to correspond to

around 10.0 METs.

Having studied the Compendium, the activities

that are most likely to reduce as the onset of

depression occurs are mainly ambulatory activities

such as Bicycling, Sports, Dancing, Walking,

Running, Stairs and Hill climbing. The table below

shows some of the MET values for each activity.

Table 1: METs for different activities.

Code METs Activity Description

01015 8.0 Bicycling

03031 4.5 Dance: Disco, Folk, Country

15711 8.0 Sport: e.g. Volleyball, Gym

17151 2.0 Walking, less than 2.0 mph,

17231 8.0 Walking, 5.0 mph

15734 10.0 Running, sprint, athletics

Rather than compute actual energy expenditure

in calories, it is convenient to try and count the MET

minutes product used by a person per day. Once one

can identify what ambulatory activity is being

performed then the total time spent during each

activity provides a simple way to measure and

compare people’s daily behavior.

In other related research it has been shown that

there is a relationship between a person’s gait or

walking style, and their depressive symptoms (J.

Michalak, 2009). However the primary impact of

depression is on the upper body posture and the

speed of walking, not on the leg posture or the action

cycle of the legs. Therefore we do not expect

depression to affect ambulation patterns.

In summary, as part of a multiple sensor solution,

an activity sensor that can recognize the different

ambulatory activities is being developed and this

paper reports on the success to date.

Section 2 describes the OPTIMI architecture

within which the activity sensor operates. Section 3

describes the sensor hardware while section 4

explains the recognition approach and section 5

discusses the results obtained during preliminary

trials.

2 SENSOR ARCHITECTURE

The OPTIMI project incorporates a number of smart

wearable sensors. The following summarizes the

sensors and their target function:

• The Activity Sensor (ACT) for ambulatory

activity recognition, described in this paper

• ECG for heart rate derived stress indicators

• EEG to derive affective status (sensitivity)

• Sleep Quality, restlessness and insomnia

• Sub Dermal Cortisol

• Speech Analysis to estimate depression

In many wearable sensor applications, data from

the sensor is streamed to a signal processing

computer such as a wearable PDA. This approach

can result in a 90% wastage of the available sensor

battery energy in wireless communications and

sensors that are worn 24/7 would require frequent

recharging.

Instead significant energy conservation can be

achieved by making the sensor itself do the signal

processing using highly optimized signal processing

in which the large amounts of real time information

are processed and converted to time stamped results

comprising only a few hundred bytes per day. When

this data is communicated wirelessly not only is

their little energy spent on communications but in

addition raw data privacy and data security are

enhanced.

Therefore the policy in OPTIMI has been to

process data at source as far as possible and to

derive a much smaller encrypted set of time stamped

data reflecting the state of the user. As a result the

activity and ECG sensors process data locally all day

long and store the results locally. At one time during

the day, the user is invited to update their daily diary

hosted on a Home PC as well as use the speech and

EEG sensors. While these tasks are being carried out

the ECG and ACT data is wirelessly downloaded.

Figure 1: The OPTIMI sensor and Home PC Architecture.

3 SENSOR HARDWARE

The ACT hardware is based on the nRF24LE1 from

Nordic. This micro-controller was chosen despite its

limited processing capability, based on its very low

cost, very small footprint, suitable ADC and flash

HEALTHINF 2011 - International Conference on Health Informatics

220

EEPROM resources and the integration of a 2.4Ghz

RF transceiver.

The nRF24LE1 is combined with an Analog

Devices ADXL325 three axis accelerometer for the

purpose of measuring accelerations of the user’s

lower leg; the sensor being placed just above the

ankle.

The microcontroller samples the three axes at a

rate of 50 or 100Hz and performs activity

recognition processing on this data. Following each

sample and activity identification the sensor logs the

result in the EEPROM, time stamps it and continues

to the next activity recognition.

Due to the fact that the RF end of the

microcontroller consumes a large proportion of the

power, the RF stage is seldom switched on during

normal operation. However every 5 seconds, the RF

stage is switched on for 20 milliseconds and the

device listens for a command packet from the users

Home PC. In the event one is received the sensor

begins an authentication handshake and subsequent

interchange of relevant data.

As the device is to be worn continuously three

very important design constraints have been

demanded:

• The battery life must be as long as possible

• The device must be small and hermetically

sealed to survive swimming and bathing

• The device must therefore be non contact

charged

The device is powered by a 20mAh lithium

polymer battery with the intention being to last

ideally 72 hours between charge times. Besides

reducing any unnecessary usage of the RF stage, as

previously mentioned, the design includes several

energy saving methods.

For example the ADXL325 device has been

chosen for its wider G range and very low cost

however it is slightly more power hungry than other

devices available and is the main power-sink after

the microcontroller. To reduce power applied in

sensing the accelerometer and the microcontroller

ADC are only switched on before and switched off

immediately after the 3 axes are sampled. Allowing

for switch on times delay, this saves about 50% of

the power compared to leaving them on at all times

i.e. saving roughly 0.15mA on average.

Similarly, it is planned that the activity

recognition task should itself be managed depending

on time and observed activity. That is, by using the

local timer and sleep function, the sensor will set to

conduct activity recognition between around 10%

and up to 70% of the time, depending on the

observed activity seen by the accelerometer.

This is achieved by implementing a small

dedicated code to check the accelerometer outputs

and to decide if there is some activity or not. If there

is no activity, then the device goes to sleep for 100

milliseconds before repeating the activity check. As

soon as activity is detected the main recognition task

is engaged and repeated up to a maximum of 70% of

the time, the rest of the time the sensor is set to sleep

mode.

In this way if the user is continuously active,

such as engaged in a long walk, the energy

maximum usage is 70% of worst case. Meanwhile if

the user is sleeping, then the sensor is largely

inactive apart from if the user moves from time to

time.

3.1 Ergonomics Constraints

The small size of the sensor, Figure 2.0, has been

achieved using an electronic PCB design equal in

size to the dimensions of the LiPo battery used. The

assembled device has a very low profile with no on-

off switch or flash connectors.

Figure 2: The ACT sensor, 20mm x 12 mm x 3mm, the

flash connector on the right is snapped off pre

encapsulation.

The hermetic sealing requirement implies that

the sensor must have a contactless charging method

and that the sensor is totally encapsulated.

To achieve this, the ACT sensor incorporates a

LiPo battery charging circuit as well as the

rectification stage for an inductive voltage power

input. The microcontroller, powered by the battery,

is able to sense when the rectified inductive input

voltage rises above the input threshold required by

the charging circuit. When this occurs the

microcontroller goes into a deep sleep effectively

switching itself off and appearing as a tiny load

shunt across the battery. This allows the charging

circuit to charge the LiPo battery without any

interference, until the battery voltage has reached the

fully charged level.

Meanwhile the inductive loop coil which is

driven at 10KHz by a standard off the shelf

MENTAL HEALTH DECLINE PREDICTION - A Smart Sensor for Day to Day Activity Recognition

221

inductive power supply, is integrated into the

strapping system used to attach the sensor to the

users lower leg.

Figure 3.0 shows a device under inductive

charge. The electronics is sealed within a two part

epoxy resin (ALH Systems Ltd., U.K.) that has been

chosen to provide maximum water resistance,

maintain hardness to over 80 degrees C., and above

all to be extremely toxic and irritant free when worn

against the skin. Further the color and molding of

the resin is chosen to create a device that could be

worn as a fashionable accessory.

Figure 3: The ACT sensor in encapsulation and on

inductive charge.

4 RECOGNITION APPROACH

Activity recognition for health purposes has become

a much researched topic. The measurement of

activity has major implications for diseases such as

obesity and is important for monitoring and assisting

the elderly and disabled. Several research works

have attempted to detect user activity using motion

sensing [6, 7, 8].

Having reviewed the state of the art, two major

design issues were considered important in this

work. Firstly, how to cope with variations in the

time taken to complete an activity and secondly

whether a sensor should be trained for its own user

or if a generic user independent method could be

created.

The recognition approach itself requires a feature

extraction followed by a machine learning of the

classification parameters and essentially two

methods exist namely discriminative and generative.

In many of the discriminative approaches, time is an

important dimension. For example most high end

pedometers and more ambitious work (U. Maurer,

2006) use time domain statistics exclusively to

extract the features. Similarly frequency domain

power spectral density may also be used, (M. Lee,

2009). In these cases how fast or slow the action is

performed factors into the type of activity

classification. This can result in different

interpretations of whether a walk activity is a fast

walk a slow walk or a slow run.

In the generative approach, the feature extraction

must generate a sequence of observations and these

could be treated as time invariant; just a list of

codes. If one can achieve total time independence

then how fast or how slow the person performs the

activity does not factor into the recognition problem.

In order to provide a generic classifier that can

recognize anyone’s activity based on a general

training, it was felt that a probabilistic approach

should be used that would include a wide range of

user variations.

Combining these two considerations, what was

decided was that a probabilistic generative approach

that excluded time during the feature extraction

process was necessary. Both feature extraction and

learning classifier should combine to detect the

activity at a general level rather than to detect the

specific users activity cycle. In this way the user

simply applies the sensor to their leg on a plug and

play basis and does not need to engage into a

specific one to one training procedure.

In a similar approach to (J. Suutala, 2007) we

have used a simpler k-means approach, rather than

SVM approach, to derive a time independent

observation feature space followed by a Hidden

Markov Model approach for the observation

sequence processing and classification.

4.1 Feature Extraction

As discussed above, the time dimension is

preferably excluded from our analysis. By doing so

we attempt to detect walking, irrespective of

whether the walk cycle is a slow walk, a medium

walk or a brisk walk cycle.

When one studies the ambulatory cycles of

walking, running, climbing and so on it becomes

clear that while there are several similarities, there

are also differences. Specifically the sequence of

rotations and jerks of the leg are different

Figure 4: The walk cycle versus the run cycle showing leg

rotation difference particularly during the flight phase.

HEALTHINF 2011 - International Conference on Health Informatics

222

irrespective, to a large degree of the time over which

they occur.

The legs rotational posture can be sensed by the

3 axis accelerometer which detects leg inclination

relative to the gravitational field. The leg jerk

accelerations are then superimposed on the gravity

signal.

Except for the case of dance and sports, where

the leg often moves in the Z direction, from left to

right, for walking and other ambulation, the main

accelerations occur in the sagittal XY plane.

Figure 5: X vs Y plots of the accelerometer axes for

running, stairs and walking.

In Figure 5.0 above the X and Y axis readings

are shown for walk, run and climbing stairs. The

readings are shown as a line plot of X against Y.

One can see that the readings map out three totally

different looking plots and these plots tend to be

largely time invariant, looking roughly the same

irrespective of the time taken to do the motion.

Therefore these X vs Y plots are the base for the

time invariant approach where the objective is to

define key points in the plot that can be used as

features.

If one looks at the related time plots for this data,

see Figure 6.0 below, one can see that in the time

domain, it is possible to quickly identify by eye the

specific parts of the action cycle corresponding to a

feature of the activity. An example in the run cycle

is the point of “flight” where we can predict both of

the user’s legs are off the ground, or in the stair

climb, where the leg is lifted to the next stair as

opposed to pushed forward in the case of the walk

cycle.

Figure 6: Annotation of the run and walk cycle indicating

primary sequence points.

Instead of using time references to mark these

points of interest we alternatively use X, Y

coordinates in the 2D XY space. These reference

points allow us to quantize the raw data and assign it

a code thereby defining an observation. Each

observation code becomes an input to the final

HMM classifier that will recognize the activity.

To automatically create the points for a specific

activity we apply a k-means algorithm to the XY

data generated during that activity cycle. This results

in a much smaller number of points that represent

the overall plot by a set of region points or zones of

interest. Such a technique was successfully used in

previous work conducted by the authors in a Tai Chi

activity recognition where points were mapped in a

3D space, (D. Majoe, 2009).

These zones are calculated as the centroids of the

standard k-means algorithm, which was modified to

accentuate the weighting associated with each data

point. Since the most interesting data points in the

activity cycle are those where the leg is moving, so

the weighting of a data point was calculated as a

function of the square of its velocity. In addition, to

de-accentuate the high number of XY values that

occur when the leg is stationary and vertical, a

“quiet” point, the weighting is increased as a

MENTAL HEALTH DECLINE PREDICTION - A Smart Sensor for Day to Day Activity Recognition

223

function of the square of the distance away from the

“quiet” point.

As a final measure, one or two of the centroids

calculated can be slightly moved by hand to further

accentuate a specific zone of interest. The centroid

distribution for an example walk cycle is shown in

Figure 7.0 .

Figure 7: K-means Centroids for the walk cycle.

Having calculated the M centroids, in XY space,

the sequence of observations for the HMM is

calculated as that sequence of centroids closest to

the flight path of the activity signal as it moves over

the XY space. In this work the number of centroids

M was finally set at 15, giving 15 possible

observation values.

4.2 HMM Recognition

The HMM recognition method has been well

described and documented (L. R. Rabiner, 1989).

The method allows the modeling of a state change

process which associates to each state an emission

probability for a given sequence observation

occurring. Therefore the XY space derived activity

observations, coming from the feature extraction,

may be associated with a probability of occurring at

a specific stage of the state change process. If an

observation does or does not occur at a point in the

sequence finally corresponds to a higher or lower

probability being estimated.

Since one has no clear idea of what the hidden

state model should be in the case of the various

ambulation activities, one relies on the Baum Welch

learning algorithm to define the state model and no

restriction is placed on the state transition matrix.

Through our experience with work on activity

recognition, the optimal number of states used is 5

and the observation set is kept to 15. For each

activity to be recognized we use a sliding window of

variable length, between 10 and 20 sequence

observation changes.

In order to train the classifiers for a particular

activity, for example for walking, the following

process is performed:

• Collect 9 walk cycles for volunteer 1

• Generate k-means zones for all data

• Adjust key centroids by hand

• Generate Observation Sequences

• Input Sequences to Baum Welch algorithm

• Save Output matrices

The output matrices reflect all that is needed for

running a classifier for this activity. To make a

classification, the forward probability algorithm is

applied using these matrices on the test observation

sequence data. The forward algorithm generates a

number closer to 1.0 the better the correspondence

of the test sequence to the training set. In order to

create a suite of activity classifiers the above is

repeated for all activities and a bag of classifiers is

used with maximum probability voting logic to

determine which activity has been performed.

To judge inter individual recognition, data from

multiple volunteers is merged to create a general

activity feature extraction centroid reference base.

Since there are differences in the way people walk

and run it was hypothesized that the recognition

rates would be significantly lower than for the

individual trained approach.

Figure 8: Generalized Centroids formed from all the walk

cycles of all Volunteers.

Recognition rates will be acceptable provided

the inter individual and intra-individual activity

patterns are similar and that the feature extraction

and HMM classifiers can make use of these

similarities. The centroid pattern generated by the k-

means algorithm should be similar whether it is for

one person or for multiple people’s activity cycle

data. That is, zones of interest will have a similar

visual pattern for both inter and intra individual

activity cycles.

HEALTHINF 2011 - International Conference on Health Informatics

224

Figure 9: Overlay of Volunteer 1 on the Generalized

Centroids of all 12 Volunteers.

The above plots clearly show that the feature

space generated by using data from all volunteers,

Figure 8.0, closely resembles, when overlaid with

data from a single volunteer, Figure 9.0. The two

sets of centroids demark a similar visual pattern and

this allows us to conclude that the walk cycle

activity is generally speaking the same for all

volunteers. The same was found to be true for other

ambulatory activities.

5 RESULTS

In order to obtain test data a trial, which received

full ethical approval from the Ethics Commission of

ETH Zurich, was conducted in Zurich. Twelve

volunteers were fitted with the sensors and a wide

range of activities were recorded:- Walking,

Running, Jogging, Cycling, Stairs Up, Stairs Down,

Walking up Hill, Walking Downhill, at rest and

playing catch ball.

The primary aim of this trial was to check the

sensor hardware and verify the base algorithms. So

far we have determined initial recognition accuracies

for walking, running and stairs up for inter and intra

individual recognition.

5.1 Sensor Hardware

The sensors were used in two ways. At first they

were used simply as data recorders, sampling the

activity cycles as the volunteers walked around the

Zurich city centre and then downloading the data

wirelessly to a Net-Book afterwards. In order to

perform most of the algorithm development, the

recognition analysis was then done offline on

desktop PCs.

Following the training of the classifiers and once

the emission, state and transition matrices had been

obtained for each specific activity, the HMM

Classifier based on the forward algorithm was run

on the ACT sensor embedded as an application. This

was done to assess the computational load in a real

time situation.

The feature extraction and recognition on the

ACT device is performed as follows:

• Sample Motion Data for 5 seconds at 100 Hz.

• Generate Sequences for each activity

• Isolate the first occurring “quiet time”

observation

• Apply forward Algorithm from this point

• Majority vote for highest output classifier

• Store results and loop back to sampling

The HMM computation performed on the sensor

is restricted to the forward algorithm for each of the

activity classifiers of interest. Acting as a dedicated

activity recognition device, the ACT is very easily

reprogrammed with a new activity recognition task

by simply uploading different HMM matrices as

data files over the wireless interface.

The low power nRF24LE1 running at 16Mhz has

two main challenges, to perform the feature

extraction front end and generate the sequences as

quickly as possible and then to calculate the forward

probability as fast as possible for each activity.

The processing times measured on the sensor, for

different conditions are as follows:

FEATURE EXTRACTION

Configuration: 15 centroids

Number of samples: 100

Execution time: 459 ms

Number of samples: 50

Execution time: 230 ms

HMM CLASSIFICATION

Configuration: 5 HMM states

Observation types: 15

Sequence length: 20

Execution time: 29 ms

Sequence length: 10

Execution time: 13 ms

These results help us to decide how to structure

the overall quasi real time recognition algorithm.

Given these values, it should be possible to sample

data for 2 seconds at 50 Hz (encompassing all types

of cycle) and follow this with 0.5 seconds of

processing for the feature extraction followed by 10

classification activity types (walk, run, cycle etc.)

taking up another 300ms. So at maximum we

sample for 2 seconds and process off line for 1

second and this confirms that the use of the

MENTAL HEALTH DECLINE PREDICTION - A Smart Sensor for Day to Day Activity Recognition

225

nRF24LE1 is acceptably powerful for this

recognition processing task.

Figure 10: Trial volunteer walking on a road and sensor

located against right lower leg.

5.2 Early Recognition Results

So far the feature extraction and HMM classification

have been tested with data from the 12 volunteers.

This is probably insufficient data for a robust and

full training of the classifications, however what was

required was a proof of concept feedback at this

early stage of the work. In particular we wanted to

know if one can apply a generalized recognition

approach where individuals do not need to train their

own sensors.

Table 2: Confusion Matrix showing the percentage of

detections for a given activity (row) for each classifier

(column).

Bike Run Stairs Walk

Bike 45 0 44 10

Run 0 90 10 0

Stairs 1 6 90 3

Walk 0 0 22 78

The activity data from all volunteers was split

into one third training set and two third testing set.

The trained classifiers were run with the data from

all test cycles. Voting between classifiers was then

performed to select the highest probability output

classifier.

The above results are very encouraging as a

notional target of 80% was initially considered an

acceptable accuracy in assigning MET energy values

to different activities. On investigation the poor

result, 45%, for the Bicycle activity was caused by

vibration noise that had not been filtered out. The

misclassification of the Stairs classifier has been

tentatively explained by the limited training data

though more work is needed to fully explain it.

6 FURTHER WORK

The approach has been evaluated on 12 volunteers.

It is planned to increase the size of the trials to 50

volunteers and to obtain the activity data for

different people (size, age, shoes and clothing), in

different settings and over longer periods. Given this

data the algorithms will be further improved and

tuned.

In order to improve the sensor’s ergonomics,

user acceptance and usage compliance, a usability

study has been started in which various alternative

shapes and strapping methods should lead to higher

24/7 wearability.

In order to map the relationship between

depression and activity, calibration trials with 300

users in three countries are planned to begin in 2011.

Coping strategy trials will begin in 2012.

7 CONCLUSIONS

A smart sensor that performs the task of activity

recognition on a daily basis has been presented. The

hardware design achieves several of the goals of

accuracy, wireless data transmission,

miniaturization, low cost, hermetic packaging with

inductive charging and long battery life through

attention to power management.

The level of activity recognition achieved thus

far is promising. In particular it has been shown that

the sensor at this stage can already offer a high level

of activity recognition accuracy. Most importantly it

has been shown that ambulatory activities can be

generalized and that individual sensor training will

not be necessary.

ACKNOWLEDGEMENTS

Acknowledgements go to the support students at

ETH Zurich and the co-funding received under the

EU IST Framework 7 program for the project

OPTIMI www.optimiproject.eu.

HEALTHINF 2011 - International Conference on Health Informatics

226

REFERENCES

S. M. Stahl, “The Psychopharmacology of Energy and

Fatigue”, Journal of Clinical Psychiatry, Physicians

Postgraduate Press, January 2002, pages 6-31.

T. Takeuchi, M. Nakao, et al, “Association of the

metabolic syndrome with depression and anxiety in

Japanese men: A 1-year cohort study”, Journal of

Diabetes Metabolic Research, Rev. 2009 Nov, Vol

25(8), pages 762-7.

A. Berlin, W. Kop and P. Deuster, “Depressive Mood

Symptoms and Fatigue After Exercise Withdrawal:

The Potential Role of Decreased Fitness”, Journal of

Psychosomatic Medicine, 2006, Vol 68, pages 224-

230.

B. Ainsworth, M. Haskell et al, “Compendium of physical

activities: an update of activity codes and MET

intensities”, Journal of Medical Science Sports

Exercise, 2000, Vol 32, No.9, Suppl., pages. S498-

S516.

J. Michalak, N. Troje et al, “Embodiment of Sadness and

Depression—Gait Patterns Associated With Dysphoric

Mood”, Journal of Psychosomatic Medicine, 2009, Vol

71, pages 580-587.

U. Maurer, A. Smailagic, D. Siewiorek and M. Deisher,

“Activity Recognition and Monitoring Using Multiple

Sensors on Different Body Positions”, Proceedings of

the International Workshop on Wearable and

Implantable Body Sensor Networks, 2006, pages 113 -

116.

M. Lee, J. Kim, K. Kim, I. Lee, S. H. Jee, S. K. Yoo,

“Physical Activity Recognition Using a Single Tri-

Axis Accelerometer”, Proceedings of the World

Congress on Engineering and Computer Science,

2009, Vol 1. WCECS.

J. Suutala, S. Pirttikangas, J. Roning, H. Ichikawa et al,

“Discriminative Temporal Smoothing for Activity

Recognition from Wearable Sensors”, UCS 2007,

LNCS 4836, pages. 182–195.

D. Majoe, L. Widmer and J. Gutknecht, “Enhanced

Motion Interaction for Multimedia Applications”,

Proceedings of the 7th International Conference on

Advances in Mobile Computing & Multimedia, 2009,

Kuala Lumpur, Malaysia.

L. R. Rabiner, “A tutorial on HMM and selected

applications in speech recognition”, In Proc. IEEE,

Vol. 77, No. 2, pp. 257-286, Feb. 1989.

MENTAL HEALTH DECLINE PREDICTION - A Smart Sensor for Day to Day Activity Recognition

227