Pattern identification is a widely known technology, which is used on a daily basis
for both identification and authentication. Examples include biometric identification
(fingerprint or facial), number plate recognition or voice recognition.
However, when we move into the world of medical diagnostics this changes
substantially. This field applies many of the recent innovations and technologies, but
it is more difficult to see cases of pattern recognition applied to diagnostics. In addition,
the cases where they do occur are always supervised by a specialist and performed in
controlled environments. This behaviour is expected, as in this field, a false negative
(failure to identify pathology when it does exists) can be critical and lead to serious
consequences for the patient. This can be mitigated by configuring the algorithm to be safe
against false negatives, however, this will raise the false positive rate, which may increase
the workload of the specialist in the best case scenario or even result in a treatment being
given to a patient who does not need it. This means that, in many cases, validation of the
algorithm’s decision by a specialist is necessary, however, there may be cases where this
validation is not so essential, or where this first identification can be treated as a guideline
to help the specialist. With this objective in mind, this thesis focuses on the development
of an algorithm for the identification of lower body pathologies.
This identification is carried out by means of the way people walk (gait). People’s gait
differs from one person to another, even making biometric identification possible through
its use. however, when the people has a pathology, both physical or psychological, the
gait is affected. This alteration generates a common pattern depending on the type of
pathology. However, this thesis focuses exclusively on the identification of physical
pathologies. Another important aspect in this thesis is that the different algorithms are
created with the idea of portability in mind, avoiding the obligation of the user to carry
out the walks with excessive restrictions (both in terms of clothing and location).
First, different algorithms are developed using different configurations of smartphones
for database acquisition. In particular, configurations using 1, 2 and 4 phones are
used. The phones are placed on the legs using special holders so that they cannot move
freely. Once all the walks have been captured, the first step is to filter the signals to
remove possible noise. The signals are then processed to extract the different gait cycles
(corresponding to two steps) that make up the walks. Once the feature extraction process
is finished, part of the features are used to train different machine learning algorithms,
which are then used to classify the remaining features. However, the evidence obtained
through the experiments with the different configurations and algorithms indicates that it
is not feasible to perform pathology identification using smartphones. This can be mainly
attributed to three factors: the quality of the signals captured by the phones, the unstable
sampling frequency and the lack of synchrony between the phones. Secondly, due to the poor results obtained using smartphones, the capture device is
changed to a professional motion acquisition system. In addition, two types of algorithm
are proposed, one based on neural networks and the other based on the algorithms used
previously. Firstly, the acquisition of a new database is proposed. To facilitate the capture
of the data, a procedure is established, which is proposed to be in an environment of
freedom for the user. Once all the data are available, the preprocessing to be carried out is
similar to that applied previously. The signals are filtered to remove noise and the different
gait cycles that make up the walks are extracted. However, as we have information from
several sensors and several locations for the capture device, instead of using a common
cut-off frequency, we empirically set a cut-off frequency for each signal and position.
Since we already have the data ready, a recurrent neural network is created based on the
literature, so we can have a first approximation to the problem. Given the feasibility of
the neural network, different experiments are carried out with the aim of improving the
performance of the neural network.
Finally, the other algorithm picks up the legacy of what was seen in the first part of the
thesis. As before, this algorithm is based on the parameterisation of the gait cycles for its
subsequent use and employs algorithms based on machine learning. Unlike the use of time
signals, by parameterising the cycles, spurious data can be generated. To eliminate this
data, the dataset undergoes a preparation phase (cleaning and scaling). Once a prepared
dataset has been obtained, it is split in two, one part is used to train the algorithms, which
are used to classify the remaining samples. The results of these experiments validate
the feasibility of this algorithm for pathology detection. Next, different experiments
are carried out with the aim of reducing the amount of information needed to identify
a pathology, without compromising accuracy. As a result of these experiments, it can be
concluded that it is feasible to detect pathologies using only 2 sensors placed on a leg.