By means of the information extracted from the different postural monitoring methods, it is necessary to quantify and estimate the patient’s postural health status. For this purpose, after a first step of monitoring, it is necessary to carry out a second step of applying anomaly detection techniques to detect abnormal postural states. Different techniques have been proposed in the literature to meet this second step (Fig. 6).

The detection of sitting posture anomalies can be approached from two distinct perspectives (Fig. 6). Traditionally, the normal posture of a user has been considered to be that which involves an upright spine, distributing the weight of the body evenly over the seat and backrest. In other words, the terms correct posture and normal posture have often been used indiscriminately, using a general approach without taking into account the physical conditions of each user. This is due to a large extent to the fact that much of the work is oriented towards maintaining correct postural health in office workers, and not to wheelchair users. Following this definition of normality, abnormal posture is considered to be any posture that is not the correct posture, i.e., upright spine.

Intelligent techniques for the detection of sitting postural anomalies from a generalist approach

However, this traditional approach has a number of limitations. Following the traditional approach, there is a single normality, defined by the correct posture. However, this approach does not take into account the intrinsic particularities of each patient’s pathology, which may result in each user having their own postural pattern. On the other hand, the concept of anomaly has been considered taking the correct posture as a reference, when it is possible that a user, due to muscular weakness, will never be able to perform it. Thus, for example, Thus, for example,people who suffer a stroke usually have paralysis on one side of their body. Because of this, he will tend to lean to one side, and will hardly be able to keep the spine straight. Therefore, under this traditional approach, this person will be under an abnormal postural state constantly.

Therefore, this work considers it interesting to define anomalies based on a new approach. In this approach, the individualized sitting postural pattern is characterized for each user. Subsequently, changes or alterations in these postural patterns that may be indicative that the user’s functional status has changed are sought. These changes in the postural pattern will be considered as postural anomalies. Thus, following the previous example, this person with stroke will characterize his normality by this chest tilt. Under this new approach, any posture that deviates from this lateral tilt will be considered abnormal. Thus, it is possible to detect either a recovery, when he/she manages to keep the back upright, or a worsening, when this tilt becomes more pronounced.

Throughout this section the techniques for the detection of anomalies used to date in the field of postural diagnosis following a classical or generalist approach will be shown (section Traditional approach through generalized techniques), as well as the need to treat the problem from a new individual approach, showing the most common techniques used in other fields (section New and individualized anomaly detection approach). For these techniques, quantitative aspects such as the percentage of success, false positive and false negative rate or computational cost have been analyzed. In addition, other aspects have been taken into account, such as the prior knowledge about anomalies necessary for model training, as well as the effort for data labeling.

As mentioned above, the sitting posture anomaly detection has usually been carried out in a generalist way by detecting all those postures that are not considered to be correct. For this purpose, most techniques have been based on the generation of classification models that allow the classification of different common postures. For this purpose, the techniques used are usually supervised, with the availability of samples labeled as normal and anomalous.

To monitor these anomalies, three types of techniques can be distinguished (Fig. 7): rule-based techniques, statistical techniques and intelligent techniques.

Diagram of the anomaly detection techniques used in the traditional approach

The first group of techniques, i.e., rule-based techniques, are based on the assumption that prior knowledge of the postural behavior of users is available. In this way, logical reasoning can be applied to impose the use of rules or syllogisms that allow the postural state of a given user [49, 63, 74].

The main advantage offered by this type of techniques is their simplicity of application. In general, they are based on a low number of rules, being therefore the speed of inference high, allowing them to be implemented in systems with limited computational and time requirements. On the other hand, the classification process is totally transparent and explainable, being formed by logical and understandable rules, thus allowing the user to understand at all times the decision-making process of the model.

Furthermore, this type of rules have a number of shortcomings that may make them unsuitable for detecting sitting posture anomalies. Moreover, as they are designed, expert knowledge is necessary for the definition of the different logical relationships that characterize the postural behavior of the users. In the absence of this expert knowledge, the model no longer has the desired performance. In addition, these techniques have a limitation in terms of complexity. In general, they are not able to capture complex relationships, so the models developed are limited, capturing a limited number of postures, and leaving out the monitoring of postures characteristic of people with low mobility such as thoracic rotation. Finally, these rules can be developed for a specific case, but the developed model does not necessarily have to be generalizable to the rest of the users.

Among the different techniques, using discriminant thresholds is the most common due to its ease of understanding and implementation [63]. This is a widely used algorithm based on establishing a predefined threshold above which two activities or postures can be separated. The characteristics that can be taken into account to perform this discrimination are diverse.

In case of using wearable sensors, the usual approach when using thresholding is to define as a threshold the degree of inclination with respect to the vertical axis [49]. If a certain threshold is exceeded, the posture adopted is considered inappropriate, whereas, if it does not reach that figure, the posture is considered correct. In case the monitoring system is through the use of pressure sensors, the preset threshold is useful to prevent ulcers, locating those points where the specified pressure (the established threshold) is exceeded [63].

In addition to making use of thresholds, these can also be combined with the application of different rules or decisions, thus adding an additional condition to define the postures. Thus, as an example, in [74] it is established that a person is in an upright posture, when the x-axis data taken from an accelerometer is zero, and the acceleration on the z-axis is coincident with gravity. In sedentary sitting, a stepwise binary discrimination process is followed. At each stage, one passes to a subsequent stage or not on the basis of prior verification of the fulfillment of a certain rule. In this way, up to 6 different postures are differentiated [65].

In general, rule-based techniques are computationally inexpensive but lack sufficient discrimination to distinguish postures with a high degree of complexity. These complex postures precisely define the postural behavior of individuals with low mobility. Thus, this type of techniques are restricted for use in binary classification applications, where the result is an all or nothing, i.e., the correct posture or not. For this reason, different sets of techniques that allow a more complex study of postural anomalies are addressed in the literature.

The second group of techniques is composed of the so-called statistical techniques [13,14,15,16,17,18,19,20,21,22,23]. This set of techniques is characterized by the use of statistical and probability methods to study whether a new sample belongs to a certain class or another, based on the relationships and patterns found in the training data.

Among the main advantages of this set of techniques is, as in the previous case, the interpretability of the results. As with rule-based techniques, a process of understanding the decision-making process of the model can be carried out, thus being able to understand the logical reasoning followed by the model. Not only that, statistical models usually provide probability estimates for each classification class, so it is also possible to know the degree of confidence of the prediction. The possibility of using them in multiclass problems, added to a reasonable training time of the models, makes that this type of techniques has been frequently studied in the literature.

The main limitations of this type of techniques are that they work under different assumptions. Among them, that the training (and new) data follow a statistical distribution, which is the one that allows the estimation to be performed later. Knowing the distribution that the data follow (if they do) again requires knowledge of the database by an expert. Another assumption that is often made is the independence of the input variables. This independence does not need to be true, especially in the face of an increase in the dimensionality of the problem. These techniques are not able to capture the relative importance of the input features and may be sensitive to irrelevant ones. This results in limited performance if the data are complex.

Among the classifiers based on statistical techniques, the Naive-Bayes classifier or Naive Bayesian classifier is in first place [13, 14]. This classifier is characterized by using Bayes Theorem to calculate probabilities. This theorem assumes that the features of a class or object are independent of each other, and therefore, each of them contributes independently to the probability of one class or another. Among the advantages of this type of methods is their simplicity and the small number of parameters required for their implementation. On the contrary, it assumes an independence of variables that in practice is not necessarily true, which can lead to a decrease in the precision obtained. This model has proved to be successful when the number of monitored postures is small [14], however, the results are affected when the number of postures increases significantly [28].

The use of K-nearest neighbor (KNN)-based classification systems have also been proposed in the literature [15,16,17,18,19]. This classifier is based on the construction of a multidimensional feature space where it is assumed that those data belonging to the same class have similar characteristics, and are therefore grouped into nearby clusters. Thus, once the classifier is trained and the feature space is created, to determine the class of a new data, the classes of the K-nearest neighbors are checked. The class to which a larger number of neighbors belong will be the one that is finally assigned. It is a simple algorithm, but it returns high effectiveness ranges. Thus, it has been used in [20] to classify postures that involve a back movement such as tilt or rotation, considered as anomalous. Likewise, in [21] KNN is used for the classification of different postures that involve a change in both the back and legs, mainly oriented to the detection of the postural state in office workers. Nevertheless, since it relies on classifying a posture based on proximity to others in the database, it is necessary to have a diverse and balanced database for all classes. Otherwise, the algorithm will tend to be biased towards those postures that are more frequent in the database. However, a priori, the postures that are not correct and therefore, anomalous based on this approach, will be the most frequent, since they will be all those that are not considered correct. This is why it is not always possible to meet this objective of having a balanced database. Moreover, it is highly dependent on the size of the database, since it has to measure the distance between the new point and all the existing data. Therefore, although it is a technique that does not return bad results, it can be inefficient for high dimensional databases.

Another algorithm used for the determination of sitting posture anomalies based on statistical methods is logistic regression [15, 18]. Logistic regression is based on predicting the probability that a point belongs to one of two mutually exclusive categories. In the postural domain, it is calculated whether a point belongs to a particular posture or not. For this, a logistic function whose coefficients are adjusted during the training phase of the system is used. To carry out a multiclass regression, a logistic probability is calculated independently for each postural class independently and the new data is assigned to the one that returns a higher probability, i.e. to the posture that it is more likely to be. One of the main advantages of this method is that, unlike other statistical methods, it makes no prior assumptions about the distribution of the data, and is capable of handling unbalanced training data sets. Nonetheless, this method does assume that there is a linearity relationship between the different variables, so nonlinear problems are not suitable for this technique. Likewise, it has problems extracting patterns from complex, high-dimensional data, giving lower hit results.

Finally, although it is not a classification technique itself, the use of the principal component analysis (PCA) technique as a preliminary step to the use of other algorithms should also be highlighted [22, 23]. In general, when collecting measurement data, more information is get than is necessary for subsequent posture classification. In addition, many of the features or variables are correlated with each other, resulting in redundancy in the data. As has been shown, one of the main limitations of statistical models is that they have difficulty in dealing with the redundancy of the data and the high dimensionality of the data. Principal component analysis is an unsupervised dimensionality reduction technique that seeks to simplify the input data to a classification model used later. This algorithm does not serve by itself to make postural diagnoses, but as a preliminary step and in order to facilitate the use of a subsequent classifier.

The last of the groups of anomaly detection techniques used to date are intelligent techniques or those based on using artificial intelligence. Specifically, Machine Learning techniques are used, a subfield of artificial intelligence that gives devices the ability to ‘learn’ without being explicitly programmed for this purpose.

Among the main advantages of this type of techniques is the fact that they are able to capture complex relationships in the input data, thus allowing more accurate anomaly detection. In addition, these techniques can generalize patterns from training data and apply these patterns to unseen data. On the other hand, they are more adaptable to changes in the input data, being able to retrain the model without the need for large adjustments. Finally, they can detect correlations in the input data, often not evident to humans, and can automatically learn the relative importance of the input features, eliminating or ignoring those that are redundant [75].

Among the main disadvantages of this set of techniques is the need for large amounts of data compared to previous techniques in order to characterize the existing patterns in the data. This increase in the number of training data is directly related to an increase in training time, as well as in the computational resources required to carry it out [76]. The difficulty of choosing the hyperparameters of the models requires a more complex adjustment and configuration than the previous techniques, requiring expert knowledge to do so. Finally, unlike the previous models, this set of techniques usually presents problems when interpreting the results, commonly referred to as black box models, in which the decision-making process is unknown to the user [77, 78].

One of the techniques used in the field of bioengineering for the determination of anomalies are hierarchical methods. These algorithms are based on decision-making, where the responses lead to different nodes where a new decision-making process occurs. This continuous decision making procedure is continued until a point is reached where the answer to the decision turns out to be one of the searched classes. The decision trees [5, 15, 17, 24] is an algorithm that allows to automate this successive decision making process. In this way, the algorithm decides, by means of a previous training in which an analysis of the input data and classification classes is performed, which is the tree configuration that maximizes the probabilities of performing a successful classification. Among the major advantages of using this technique is the fact that it is computationally efficient and does not require prior scaling of the data, saving additional data preprocessing time. In addition, this set of techniques allows the identification of those features that are most relevant for decision making, gaining in interpretability of the results. However, they may have limitations in case the relationships between the different postural variables are on a global scale, since the decision making is based on local divisions of the features. Moreover, to achieve high accuracies, the number of postures to be classified is highly limited. This is reflected in [5], where decision trees are used to differentiate whether a person is sitting, lying or walking. Although high accuracies (98.6%) are achieved, they are limited to monitoring only three actions. Similarly, in [24] decision trees are used for fall detection.

Another hierarchical method used is the Random Forest technique [25,26,27,28]. This is a variant of decision trees. While in decision trees, the entire feature set is used during training to build the model, in Random Forest, a part of the global feature set is used to build a decision tree and obtain a classification result. This process is repeated, creating different trees, each consisting of a different data set, and obtaining a classified class from each of them. The class that is most repeated among the responses of the different trees is the result returned by the Random Forest algorithm. This procedure allows to achieve greater precision, as well as a greater capacity for generalization. As in the previous cases, despite achieving high percentages of precision in the model (over 90%), the number of positions is still limited (maximum of 7). All this, at the cost of losing interpretability in the results, which can be vital in order to be able to provide relevant postural information to health specialists. In addition, it requires a higher computational and memory cost, so it may not be suitable for systems with limited resources.

Another technique to highlight is the Support Vector Machines (SVM) [15, 19, 29,30,31,32]. The SVM algorithm is a supervised method capable of classifying samples through the use of a separator. While this technique falls under the category of linear separators, it’s important to note that the separation between classes doesn’t necessarily have to be linear. The algorithm of this classifier is based on the search for a separation hyperplane that is equidistant from the closest points of each class. It is also sought that there is the maximum possible margin between classes. Once the model has been trained and the optimal hyperplane has been found, the position of the new data with respect to it is evaluated to decide to which class it belongs. In general, this technique is widely used in biomedical applications, since it is effective for small data sets. In addition, it has a high generalization capability. However, it can give problems in high dimensionality data, especially in those problems with many classes. This is especially detrimental in the case where the number of postures to be studied is high. Thus, in [30], despite the fact that a high number of postures can be classified, studying a total of 12, the accuracy of the classifier is reduced (hit results below 80%). Similarly, in [29], four load cells are used for the classification of 6 common postures. However, both works are oriented to office workers. This population presents a different postural problem from that of people with low mobility, with postures different from those of wheelchair users, such as displacements on the seat or leg movements.

Finally, techniques based on neural networks are used, which have proven to be one of the classification techniques with the greatest potential in the field of bioengineering applications [79], having grown considerably in recent years. The network is composed of layers formed by interconnected layers, where each neuron transmits the input information modified by a function of its own and multiplied by the specific weight to the successive layers. In this way, a structure with the capacity to process large amounts of information and to identify trends and classify postures based on them is achieved. This system is based on automatic learning by means of a previous training, where the different specific weights of each of the neurons are weighted.

There are several typologies of neural networks, which are classified according to the structure of the layers or the interneuronal connections, among other aspects. One of the most widely used neural network models are the Multilayer FeedFordward or Multilayer Perceptron (MLP) networks [66, 68, 80,81,82]. This type of networks are characterized by having an input layer and an output layer, together with an undetermined number of hidden layers. Each of the neurons in one layer is interconnected with all the neurons in the next layer. This means that the number of parameters in this type of network can reach very high levels. This is a supervised learning method that uses the Back-Propagation algorithm for training. This algorithm, based on the gradient descent method, seeks to modify the weights of each neuron to reduce the global error, starting from the last layer and continuing with the preceding layers. Thus, in [75, 80, 81], use this type of networks are used to perform recognition of activities, such as climbing stairs, or running, as well as various postures. However, these works focus more on activity recognition and do not put their focus on postural diagnosis in seated position of wheelchair users.

Another type of networks used for postural classification are convolutional neural networks (CNN) [76,77,78, 83]. This type of networks, first perform a process of extraction and reduction of the dimensionality of the features and then go on to perform the classification based on these features. These networks are focused on the field of artificial vision and image processing. Thus, they are used with data extracted using depth cameras for anthropometric scanning of the human body [76], as well as with data extracted using pressure mats for postural diagnosis [77, 78]. After all, as explained in [83], pressure distribution signals extracted from pressure mats can be treated as images. Nevertheless, there are small variations that differentiate it from images extracted through cameras. Among these differences is the fact that the sensors are not isolated and therefore the pixels depend on the force exerted on the surroundings. However, these studies, despite achieving superior performance with respect to previous methods, also in part because the input data set is larger, the computational requirements grow in the same way. In addition, they require a large number of input data for the model to be effective, and they have difficulty in detecting small details within the overall image. Therefore, they can present problems in detecting certain anomalies, especially if they are similar to another set of studied postures.

Most classifiers used in the literature focus on the classification of static postures and activities, which require a period of time for the data to stabilize. However, transitional states between different postures can be a challenge when detecting anomalous postures, since they can be mixed. That is why, some authors [84, 85] decide to make use of fuzzy logic techniques for the optimization of transient or switching postures. Fuzzy logic arises from the idea of emulating human thinking, where sometimes not everything is true or false, and there are middle terms whose degree of definition is often imprecise. In this way, and derived to data processing, fuzzy logic aims to emulate a decision-making tool where the input information is imprecise or ambiguous. The features are described by means of membership functions, which are subjected to rules to infer an output class. Thus, although fuzzy logic is not used as the sole method of diagnosis, the use of these techniques allows the increase of the percentage of success in combination with other techniques, by eliminating the existing inaccuracies in the diagnosis between positions, or transition positions.

In conclusion, postural classifiers has been widely used in the literature for the detection of postural anomalies following a traditional generalist approach. Techniques for the development of these classifiers can be classified into rule-based, statistical and intelligent techniques. While intelligent techniques have high performance and the ability to classify a larger number of postures, they have the counterpart of increased computational requirements and lack of interpretability of decision making. This is why other authors prefer to make use of statistical or rule-based techniques, based on expert knowledge of the database. The complete set of advantages and disadvantages, as well as the above cited references are summarized in Table 2.

Thus, it can be seen how the works that make use of techniques based on machine learning stand out above the rest due to the advantages they offer. However, there are still shortcomings in these works to be taken into account as future lines of research. Firstly, it should be noted that the classification techniques used to date are mainly oriented to a population different from people with low mobility. For this reason, there is a lack of works developed in clinical settings as well as an approach from health perspective to detect wheelchair common sitting postures.

Moreover, this approach is based on the generalizability of classification models. Considering that the wheelchair population is very varied in physical complexions, it is necessary that the analysis around these models further deepen this aspect. The works achieved so far generally yield good results, but they are not typically analyzed in individuals with diverse physical builds. Therefore, it is necessary to study more extensively how the results may be affected for the diverse population.

The vast majority of studies address postural anomaly detection as a common problem for all users, without taking into account the pathology of each patient. However, each user exhibits a unique and characteristic postural pattern. Within the same pathology, different individuals may display varying behaviors based on their physical characteristics [