Search

Feature selection is a challenging problem that occurs in the high-dimensional data analysis of many major applications. It addresses the curse of dimensionality by determining a small set of features to represent high-dimensional data without significant or noticeable loss of information. The purpose of this study is to develop and investigate a new unsupervised feature selection method which uses the k-influence space concept and subspace learning to map features onto a weighted graph and rank them by importance according to the PageRank graph centrality measure. The graph design in this method promotes feature relevance, downgrades redundancy, and is robust to outliers and cluster imbalances. In K-Means classification experiments using the ASU feature selection testing datasets, the method produces better accuracy and normalized mutual information results than state-of-the-art unsupervised feature selection algorithms. In a further evaluation, using a dataset of over 14,000 tweets, conventional classification of features selected by the method gave better sentiment analysis results than deep learning feature selection and classification.

The analysis of knee kinematic data, which come in the form of a small sample of discrete curves that describe repeated measurements of the temporal variation of each of the knee three fundamental angles of rotation during a subject walking cycle, can inform knee pathology classification because, in general, different pathologies have different kinematic data patterns. However, high data dimensionality and the scarcity of reference data, which characterize this type of application, challenge classification and make it prone to error, a problem Duda and Hart refer to as the curse of dimensionality. The purpose of this study is to investigate a sample-based classifier which evaluates data proximity by the two-sample Hotelling T 2 statistic. This classifier uses the whole sample of an individual&rsquo, s measurements for a better support to classification, and the Hotelling T 2 hypothesis testing made applicable by dimensionality reduction. This method was able to discriminate between femero-rotulian (FR) and femero-tibial (FT) knee osteoarthritis kinematic data with an accuracy of 88.1 % , outperforming significantly current state-of-the-art methods which addressed similar problems. Extended to the much harder three-class problem involving pathology categories FR and FT, as well as category FR-FT which represents the incidence of both diseases FR and FT in a same individual, the scheme was able to reach a performance that justifies its further use and investigation in this and other similar applications.

Modeling parallel robots is a famous problem of research especially the Forward Kinematic Model (FKM). It is very difficult to solve it compared to serial manipulators and it is also hard to obtain its analytic solution. Most researchers have resorted to the numerical methods. But, these have a lot of problems : the divergence caused by the bad choice of the initial condition and providing more than one feasible solution. To solve these problems, an analytical method is proposed in this article. In this paper, the FKM of the parallel robot Par4 is determined using the Modified Denavit Hartenberg Method (MDHM). This method facilitates greatly the forward kinematic problem of parallel robot. The MDHM is powerful, useful and can be easily used for the closed loop robots. It is very accurate and provides a unique solution. In the best of our knowledge, the MDHM isn’t applied to the parallel robot Par4 to give a relationship between its operational and articular variables. After a short presentation of the robot Par4, the FKM is determined using the MDHM method. The results of this method will be compared to the existing numerical method like the iterative Newton method (INM).

Feature selection has become an indispensable part of intelligent systems, especially with the proliferation of high dimensional data. It identifies the subset of discriminative features leading to better learning performances, i.e., higher learning accuracy, lower computational cost and significant model interpretability. This paper proposes a new efficient unsupervised feature selection method based on graph centrality and subspace learning called UGFS for ‘Unsupervised Graph-based Feature Selection’. The method maps features on an affinity graph where the relationships (edges) between feature nodes are defined by means of data points subspace preference. Feature importance score is then computed on the entire graph using a centrality measure. For this purpose, we investigated the Google’s PageRank method originally introduced to rank web-pages. The proposed feature selection method has been evaluated using classification and redundancy rates measured on the selected feature subsets. Comparisons with the well-known unsupervised feature selection methods, on gene/expression benchmark datasets, demonstrate the validity and the efficiency of the proposed method.

Driving fatigue detection has been the subject of several studies which relied on a set of features describing driver’s facial expressions, driving behaviors, and bio-signals. The purpose of this study is to improve driving fatigue detection by identifying the adequate set of features which accurately characterize fatigued drivers. The considered features are derived from non-intrusive sensors; they are related to the changes in driving behavior and visual facial expressions. The relevance is first investigated by several feature selection methods. Second, a meta-analysis was performed to investigate method agreement about the relevance of each feature in driving fatigue recognition. Support vector machine and DBSCAN classifiers were used to detect fatigue by means of the identified features. Experimental analyses are performed on a real-world database, collected through the computer vision system “FaceLab” and car sensors, from 66 senior drivers when driving an instrumented car on a highway. Results provide a list of the most discriminative features, which enhances the classification average accuracy to 89.13%.

This paper has two objectives: the first is to generate two binary flags to indicate useful frames permitting the measurement of cardiac and respiratory rates from Ballistocardiogram (BCG) signals—in fact, human body activities during measurements can disturb the BCG signal content, leading to difficulties in vital sign measurement, the second objective is to achieve refined BCG signal segmentation according to these activities. The proposed framework makes use of two approaches: an unsupervised classification based on the Gaussian Mixture Model (GMM) and a supervised classification based on K-Nearest Neighbors (KNN). Both of these approaches consider two spectral features, namely the Spectral Flatness Measure (SFM) and Spectral Centroid (SC), determined during the feature extraction step. Unsupervised classification is used to explore the content of the BCG signals, justifying the existence of different classes and permitting the definition of useful hyper-parameters for effective segmentation. In contrast, the considered supervised classification approach aims to determine if the BCG signal content allows the measurement of the heart rate (HR) and the respiratory rate (RR) or not. Furthermore, two levels of supervised classification are used to classify human-body activities into many realistic classes from the BCG signal (e.g., coughing, holding breath, air expiration, movement, et al.). The first one considers frame-by-frame classification, while the second one, aiming to boost the segmentation performance, transforms the frame-by-frame SFM and SC features into temporal series which track the temporal variation of the measures of the BCG signal. The proposed approach constitutes a novelty in this field and represents a powerful method to segment BCG signals according to human body activities, resulting in an accuracy of 94.6%.

The purpose of this study is (1) to provide EEG feature complexity analysis in seizure prediction by inter-ictal and pre-ital data classification and, (2) to assess the between-subject variability of the considered features. In the past several decades, there has been a sustained interest in predicting epilepsy seizure using EEG data. Most methods classify features extracted from EEG, which they assume are characteristic of the presence of an epilepsy episode, for instance, by distinguishing a pre-ictal interval of data (which is in a given window just before the onset of a seizure) from inter-ictal (which is in preceding windows following the seizure). To evaluate the difficulty of this classification problem independently of the classification model, we investigate the complexity of an exhaustive list of 88 features using various complexity metrics, i.e., the Fisher discriminant ratio, the volume of overlap, and the individual feature efficiency. Complexity measurements on real and synthetic data testbeds reveal that that seizure prediction by pre-ictal/inter-ictal feature distinction is a problem of significant complexity. It shows that several features are clearly useful, without decidedly identifying an optimal set.

Epileptic seizures are caused by abnormal electrical activity of brain cells, frequently accompanied by a short-lived loss of control or awareness. Epileptic seizures differ depending on their origin in the brain. They can be categorized as either focal or generalized in onset. The identification of seizure category is essential in brain surgery and in selecting medications that could help bring seizures under control. It is not always feasible to find out exactly if the seizure was generalized or focal without a thorough analysis of the continuous prolonged electroencephalographic (EEG) waveforms. In this study, we propose an automatic classification method based on Hjorth parameters measured in EEG. 1497 EEG signals from the Temple University Hospital Seizure Corpus (v.1.5.1) are used. Hjorth parameters (activity, complexity, and mobility) are extracted from these EEG records. To address class imbalance, data was rebalanced by Synthetic Minority Over Sampling (SMOTE). We also investigated the impact of changing the window length on the random forest classifier. For comparison, cost-sensitive learning has been applied by providing more weight to the minority class (generalized seizure) directly in the classifier. The performance of the proposed method was compared using accuracy, recall, and precision measures. Our method achieved a highest accuracy rate of 92.3% with a recall of 92.7% and precision of 91.8% using Hjorth parameters extracted from 10 s windows and rebalanced using SMOTE. A slight variation in performance measures occurred according to window size.

Human activity recognition (HAR) by wearable sensor devices embedded in the Internet of things (IOT) can play a significant role in remote health monitoring and emergency notification to provide healthcare of higher standards. The purpose of this study is to investigate a human activity recognition method of accrued decision accuracy and speed of execution to be applicable in healthcare. This method classifies wearable sensor acceleration time series data of human movement using an efficient classifier combination of feature engineering-based and feature learning-based data representation. Leave-one-subject-out cross-validation of the method with data acquired from 44 subjects wearing a single waist-worn accelerometer on a smart textile, and engaged in a variety of 10 activities, yielded an average recognition rate of 90%, performing significantly better than individual classifiers. The method easily accommodates functional and computational parallelization to bring execution time significantly down.

A large amount of data including joint kinematics, joint kinetics, clinical and functional measurements constitutes the clinical gait analysis basis which is a process whereby quantitative gait information are collected to aid in clinical decision-making. Therefore, better understanding the relationship between the biomechanical and clinical data for the knee osteoarthritis (OA) patient is for a relevant importance. It’s the purpose of this paper, which aims to analyze and visualize the correlation structure between biomechanical characteristics and clinical symptoms, and thus to provide an additional knowledge from the coupling of these parameters that will be useful for the pathology assessment of knee-joint disease in the end-staged knee OA patients. We perform two multivariate statistical approaches, first, a Canonical Correlation Analysis (CCA) to assess the multivariate association and, second, a graphical- based representation of the multivariate correlation to better understand the association between these multivariate data. Results show the usefulness of using such multivariate approaches to highlight association and specific correlation structure between the features and to extract meaningful information.

The purpose of this study is to perform data pre-processing to identify, for a subject, robust representative knee kinematic patterns from a set of three-dimensional (3D) curves, recorded during gait. In this study, we identify a representative gait of a given subject using the within-subject variability evaluation for outliers’ removal and reliable curves’ selection. The proposed pre-processing method encompasses steps of gait events detection, normalization, outlier detection and cycles’ selection. In order to measure the reliability of the subjects’ curves before and after pre-processing, we computed the intra-class correlation (ICC) estimates and their 95% confidence intervals for knee kinematics of a multicentric dataset of 226 asymptomatic subjects and knee osteoarthrosis (OA). The experimental results show that the proposed pre-processing method allows to identify representative knee kinematic patterns.

Ballistocardiogram signals describe the mechanical activity of the heart. It can be measured by an intelligent mattress in a totally unobtrusive way during periods of rest in bed or sitting on a chair. The BCG signals are highly vulnerable to artefacts such as noise and movement making useful information like respiratory activities difficult to extract. The purpose of this study is to investigate a classification method to distinguish between seven types of respiratory activities such as normal breathing, cough and hold breath. We propose a feature selection method based on a spectral analysis namely spectral flatness measure (SFM) and spectral centroid (SC). The classification is carried out using the nearest neighbor classifier. The proposed method is able to discriminate between the seven classes with the accuracy of 94% which shows its usefulness in context of Telemedicine.

One of the main motivations for classifying knee kinematic signals, namely the variation during a locomotion gait cycle of the angles the knee makes with respect to the three-dimensional (3D) planes of flexion/extension, abduction/adduction, and internal/external rotation, is to assist diagnosis of knee pathologies. These signals are informative but high dimensional, and highly variable, which has posed difficulties that have been addressed by machine learning algorithms. The purpose of this study is to investigate classification of knee kinematic signals through the entire gait using deep neural networks. The signals are first pre-processed to identify representative patterns, which are then used for deep learning of discriminative classifiers. This paper describes an efficient means of distinguishing between knee osteoarthrisis patients and asymptomatic participants, and our methods and experiments which validate it.

Unobtrusive monitoring of the heart rate (HR) is essential for improving medical intervention. A new generation of mattress-based fiber optic sensor (FOS) is emerging for HR monitoring. The use of this FOS mattress for medical diagnosis requires appropriate advanced signal processing algorithms. In our study, we aim to weigh the performances of a novel and cheaper microbend FOS mattress by applying ballistocardiogram and HR extraction algorithms. Therefore, our study targets comparing four types of HR extraction algorithms on the FOS mattress, namely MODWT, CEEMDAN, cepstrum and clustering. The goal is to select, based on their accuracy and computational speed, the most suitable one for online or offline application purposes. Results of applying these four chosen algorithms on the FOS mattress show that the cepstrum is the most accurate algorithm with a mean absolute error (MAE) of 4.62 ± 1.68 BPM. However, the cepstrum is more appropriate for offline monitoring with a runtime of 662.9 ms for a 10-second signal segment. The results also show that the Maximal Overlap Direct Wavelet Transform (MODWT) is more efficient with a runtime of 4.1 milliseconds for online purposes, but with a slightly bigger MAE (6.87 ± 1.94 BPM). Both methods proved to be as efficient on the new mattress technology as past intelligent mattresses.

The purpose of this study is to characterize driving fatigue using a set of facial features. These features are derived from facial expression and measure eyes and head behaviors, such as PERCLOS, blink frequency and their duration, micro-sleep, head nodding and face position. We investigated feature ranking methods to identify relevant features characterizing driving fatigue. Supervised and unsupervised classification techniques were used to evaluate the identified feature effectiveness. Experimental results are performed on a real-world database, collected through the FaceLab system from 66 senior drivers when driving an instrumented car on a highway.

The purpose of this study is to investigate data clustering to determine representative patterns in three-dimensional (3D) knee kinematic data measurements. Kinematic data are high-dimensional vectors to describe the temporal variations of the three fundamental angles of knee rotation during a walking cycle, namely the abduction/adduction angle, with respect to the frontal plane, the flexion/extension angle, with respect to the sagittal plane, and internal/external angle, with respect to the transverse plane. To offset the curse of dimensionality, inherent to high dimensional data pattern analysis, the method reduces dimensionality by isometric mapping without affecting information content. The data thus simplified is then clustered by the DBSCAN algorithm. The method has been tested on a large database of 165 healthy knee kinematic data measurements. Clusters are validated in terms of the silhouette index, the Dunn index, and connectivity. Results show that a two-cluster characterization of the kinematic knee data in each plane is quite effective. A further clinical investigation shows that the men and women knee patterns are balanced between the two clusters and, for 80% of participants, the right and left knees are in the same cluster.

The aim of this study is to develop a human activity classification system based on a wearable intelligent textile and machine learning techniques. Using the Relief-F feature selection algorithm, we identified a set of relevant features collected by the smart textile. Then, the retained features have fed a classifier in order to recognize the underlying activity. In this respect, we test a support vector machine classifier (SVM) and a k-nearest neighbor classifier (KNN). The results show the reliability of the feature selection procedure and indicate that the activities can be recognized with an overall accuracy of more than 96.37 % using the KNN classifier and 95.4 % using the SVM classifier. Since the Hexoskin intelligent textile also allows the collection of physiological data, these experimental results are very promising for practical applications of acquisition of human activities recognition, which will make it possible to study the patient's state of health or to detect physiological abnormalities in real time depending on the physical activity exerted.

The emergence of RGB-D sensors offered new possibilities for addressing complex artificial vision problems efficiently. Human posture recognition is among these computer vision problems, with a wide range of applications such as ambient assisted living and intelligent health care systems. In this context, our paper presents novel methods and ideas to design automatic posture recognition systems using an RGB-D camera. More specifically, we introduce two supervised methods to learn and recognize human postures using the main types of visual data provided by an RGB-D camera. The first method is based on convolutional features extracted from 2D images. Convolutional Neural Networks (CNNs) are trained to recognize human postures using transfer learning on RGB and depth images. Secondly, we propose to model the posture using the body joint configuration in the 3D space. Posture recognition is then performed through SVM classification of 3D skeleton-based features. To evaluate the proposed methods, we created a challenging posture recognition dataset with a considerable variability regarding the acquisition conditions. The experimental results demonstrated comparable performances and high precision for both methods in recognizing human postures, with a slight superiority for the CNN-based method when applied on depth images. Moreover, the two approaches demonstrated a high robustness to several perturbation factors, such as scale and orientation change., IEEE Life Sciences Conference 2018

This paper presents a human posture recognition system based on depth imaging. The proposed system is able to efficiently model human postures by exploiting the depth information captured by an RGB-D camera. Firstly, a skeleton model is used to represent the current pose. Human skeleton configuration is then analyzed in the 3D space to compute joint-based features. Our feature set characterizes the spatial configuration of the body through the 3D joint pairwise distances and the geometrical angles defined by the body segments. Posture recognition is then performed through a supervised classification method. To evaluate the proposed system we created a new challenging dataset with a significant variability regarding the participants and the acquisition conditions. The experimental results demonstrated the high precision of our method in recognizing human postures, while being invariant to several perturbation factors, such as scale and orientation change. Moreover, our system is able to operate efficiently, regardless illumination conditions in an indoor environment, as it is based depth imaging using the infrared sensor of an RGB-D camera.

Variability in surgical strategies for the treatment of adolescent idiopathic scoliosis (AIS) has been demonstrated despite the existence of classifications to guide selection of AIS curves to include in fusion. Decision trees and rule-based algorithms have demonstrated their potential to improve reliability of AIS classification because of their systematic approach and they have also been proposed in algorithms for selection of instrumentation levels in scoliosis. Our working hypothesis is that a rule-based algorithm with a knowledge base extracted from the literature can efficiently output surgical strategies alternatives for a given AIS case. Our objective is to develop a rule-based algorithm based on peer-reviewed literature to output alternative surgical strategies for approach and levels of fusion.A literature search of all English Manuscripts published between 2000 and December 2009 with Pubmed and Google scholar electronic search using the following keywords: "adolescent idiopathic scoliosis" and "surgery" alternatively with "levels of fusion" or "approach". All returned abstracts were screened for contents that could contain rules to include in the knowledge base. A dataset of 1,556 AIS cases treated surgically was used to test the surgical strategy rule-based algorithm (SSRBA) and evaluate how many surgical treatments are covered by the algorithm. The SSRBA was programmed using Matlab. Descriptive statistic was used to evaluate the ability of the rule-based algorithm to cover all treatment alternatives.A SSRBA was successfully developed following Lenke classification's concept that the spine is divided into three curve segments [proximal thoracic (PT), main thoracic (MT) and thoracolumbar/lumbar (TL)]. Each of the 1,556 AIS patients in the dataset was ran through the SSRBA. It proposed an average of 3.78 (±2.06) surgical strategies per case. Overall, the SSRBA is able to match the treatment offered by the surgeon in approach and level of fusion 70 % of the time (with one vertebral level leeway).This study is to the author's knowledge the first attempt at proposing an algorithm to output all surgical alternatives for a given AIS case. It uses a rule-based algorithm with a knowledge base extracted from peer-reviewed literature in an area with great variability. When tested against a database of AIS patients treated surgically, the SSRBA developed has the ability to propose a surgical plan with respect to approach and levels of fusion that match the surgeon's plan in a great majority of cases. Since this SSRBA seems to output multiple valid surgical strategies, it could allow the comparisons of various strategies and the outcomes achieved in similar cases in large databases for a given case and guide surgical treatment.

Virtual reality (VR) has found many uses in physical and motor rehabilitation in recent years. VR offers new possibilities for altering a user's visual self-representation using a self-avatar that replicates his movements accurately, or modifies them by incorporating modulations in real-time. The objective of this study is to quantify the variations in gait kinematics of subjects when the lower-limb movements of their virtual self-avatar are altered. In this study, ten subjects were evaluated. They walked on an instrumented split-belt treadmill while viewing their self-avatar from either a first person perspective or a third person perspective. After a habituation period, the flexion angle of the right hip of the avatar was increased which resulted in an asymmetrical gait pattern. Gait parameters were compared between the different conditions to establish if the subject tends to change his gait in order to follow the self-avatar's gait. Interestingly, the subjects increased their hip extension when presented with their unmodulated self-avatar. Modulation of the avatar did not lead to any consistent changes across the studied subjects. Possible explanations are discussed.

Fall detection is very important to provide adequate interventions for aging people in risk situations. Existing techniques focus on detecting falls using wearable or ambient sensors. However, they do not consider fall orientations. In this paper, we present our novel fall detection system based on smart textiles and machine learning techniques. Using a non-linear support vector machine, we determine the fall orientation which will be helpful to study the impact of a fall according to its orientation. Additionally, we classify falls based on their orientations among 11 classes (moving upstairs, moving downstairs, walking, running, standing, fall forward, fall backward, fall right, fall left, lying, sitting). Results show the reliability of the proposed approach for falls detection (98% of accuracy, 97.5% of sensitivity and 98.5% specificity) and also for fall orientation (98.5% of accuracy).

Background context Variability in classifying and selecting levels of fusion in adolescent idiopathic scoliosis (AIS) has been repeatedly documented. Several computer algorithms have been used to classify AIS based on the geometrical features, but none have attempted to analyze its treatment patterns. Purpose To use self-organizing maps (SOM), a kind of artificial neural networks, to reliably classify AIS cases from a large database. To analyze surgeon's treatment pattern in selecting curve regions to fuse in AIS using Lenke classification and SOM. Study design This is a technical concept article on the possibility and benefits of using neural networks to classify AIS and a retrospective analysis of AIS curve regions selected for fusion. Patient sample A total of 1,776 patients surgically treated for AIS were prospectively enrolled in a multicentric database. Cobb angles were measured on AIS patient spine radiographies, and patients were classified according to Lenke classification. Outcome measures For each patient in the database, surgical approach and levels of fusion selected by the treating surgeon were recorded. Methods A Kohonen SOM was generated using 1,776 surgically treated AIS cases. The quality of the SOM was tested using topological error. Percentages of prediction of fusion based on Lenke classification for each patient in the database and for each node in the SOM were calculated. Lenke curve types, treatment pattern, and kappa statistics for agreement between fusion realized and fusion recommended by Lenke classification were plotted on each node of the map. Results The topographic error for the SOM generated was 0.02, which demonstrates high accuracy. The SOM differentiates clear clusters of curve type nodes on the map. The SOM also shows epicenters for main thoracic, double thoracic, and thoracolumbar/lumbar curve types and transition zones between clusters. When cases are taken individually, Lenke classification predicted curve regions fused by the surgeon in 46% of cases. When those cases are reorganized by the SOM into nodes, Lenke classification predicted the curve regions to fuse in 82% of the nodes. Agreement with Lenke classification principles was high in epicenters for curve types 1, 2, and 5, moderate in cluster for curve types 3, 4, and 6, and low in transition zones between curve types. Conclusions An AIS SOM with high accuracy was successfully generated. Lenke classification principles are followed in 46% of the cases but in 82% of the nodes on the SOM. The SOM highlights the tendency of surgeons to follow Lenke classification principles for similar curves on the SOM. Self-organizing map classification of AIS could be valuable to surgeons because it bypasses the limitations imposed by rigid classification such as cutoff values on Cobb angle to define curve types. It can extract similar cases from large databases to analyze and guide treatment.

The purpose of this article is two-fold : (1) to select a set of bio-mechanical features to characterize arthroplasty candidates and, (2) design a surgical and non-surgical candidate classifier via decision trees. The biomechanical features are generated from 3D knee kinematic patterns, namely, flexion-extension, abduction-adduction, and tibial internal-external rotation measurements taken during gait recordings. The selection of features is done by incremental selection of biomechanical parametes in a classification tree of cross-sectional data. These features are then used to generate decision rules for classification. The effectiveness of the classifier is evaluated by receiver operating characteristic curve analysis, namely, the area under the curve (AUC), sensitivity, and specificity. The classification accuracy is 85% for AUC, 80% for sensitivity, and 90% for specificity. These results demonstrate the effectiveness of the selected biomechanical features and decision tree classifier to perform automatic and objective classification of surgical and non-surgical candidates for arthroplasty.

The pivot shift test is the only clinical test that has been shown to correlate with subjective criteria of knee joint function following rupture of the anterior cruciate ligament. The grade of the pivot shift is important in predicting short- and long-term outcome. However, because this grade is established by a clinician in a subjective manner, the pivot shift's value as a clinical tool is reduced. The purpose of this study was to develop a system that will objectively grade the pivot shift test based on recorded knee joint kinematics. Fifty-six subjects with different degrees of knee joint stability had the pivot shift test performed by one of eight different orthopaedic surgeons while their knee joint kinematics were recorded. A support vector machine based algorithm was used to objectively classify these recordings according to a clinical grade. The grades established by the surgeons were used as the gold standard for the development of the classifier. There was substantial agreement between our classifier and the surgeons in establishing the grade (weighted kappa=0.68). Seventy-one of 107 recordings (66%) were given the same grade and 96% of the time our classifier was within one grade of that given by the surgeons. Moreover, grades 0 and 1 were distinguished from grade 2 to 3 with 86% sensitivity and 90% specificity. Our results show the feasibility of automatically grading the pivot shift in a manner similar to that of an experienced clinician, based on knee joint kinematics.

STUDY DESIGN The assignment of adolescent idiopathic scoliosis (AIS) curves into curve types (1-6), as described by Lenke et al, was evaluated by 12 independent observers using the original description versus a decisional tree algorithm. OBJECTIVE To determine whether a decision tree algorithm can improve classification accuracy using the Lenke classification for AIS. SUMMARY OF BACKGROUND DATA Curve type classification in AIS relies on several parameters to consider, and its relative complexity has lead to conflicting studies that reported fair-to-excellent interobserver reliability. King's classification reliability was shown to be improved using a rule-based automated algorithm. No similar algorithm for Lenke's classification currently exists. METHODS A clinical diagram derived from a decision tree was developed to help clinicians classify AIS curves. Twelve clinicians and research assistants were asked to classify AIS curves using 2 methods: the original Lenke chart alone and the decision tree diagram in addition to the Lenke Chart. Wilcoxon ranking tests were used to evaluate any difference in classification accuracy and speed for both methods. Mann-Whitney tests were used to compare experts' and nonexperts' results. Pearson correlation was calculated to evaluate the relationship between accuracy and time taken to classify. RESULTS.: Use of the decision tree for curve type determination improved classification accuracy from 77.2% to 92.9% (P = 0.005) without requiring more time to classify. This improvement was statistically significant (P < 0.05). A statistically significant correlation between accuracy and time spent classifying when the decision tree is used was also observed (R = 0.62, P = 0.032). CONCLUSION Transfer of a computer algorithm, a decision tree, to a clinical diagram improved both accuracy ofAIS classification. Algorithmic diagrams could prove beneficial to increase classification reliability due to their systematic approach.

This study investigates Bayes classification of online Arabic characters represented by histograms of tangent differences and Gibbs modeling of the class-conditional probability density functions. The parameters of these Gibbs density functions are estimated following the Zhu, Wu, and Mumford constrained maximum entropy formalism, originally introduced for image and shape synthesis. We investigate two partition function estimation methods: one uses the training sample and the other draws from a reference distribution. The efficiency of the corresponding Bayes decision methods, and of a combination of these, is shown in experiments using a database of 9504 freely written samples by 22 scriptors. Comparisons to the nearest neighbor rule method and Kohonen neural network methods are provided.

Objective This article describes a computer-based method for the classification of spine scoliosis severity. This is a first step toward an effective computerized tool to assist general practitionersdiagnosespinescoliosis.Themethodprogresses away from Cobb angles toward pattern and magnitude categorization based upon 3D configurations. Materials and methods The purpose is to classify spine shapes reconstructed from a pair of calibrated X-ray images into one of three categories, namely, normal spine, moderate scoliosis,andseverescoliosis.Thespineshapeisrepresented by the three-dimensional coordinates of a sequence of equidistant points sampled by interpolation on the reconstructed spine shape. Classification is carried out using a selforganizing Kohonen neural network trained using this representation. Results The tests were performed using a database of 174 spine biplane X-rays. The classification accuracy was 97%. Conclusion The results demonstrate that classification of 3DspinedescriptionsbyaKohonenneuralnetworkaffordsa solidbasisforaneffectivetooltoassistcliniciansinassessing scoliosis severity.

The purpose of this study is twofold: (1) to develop a classification method to distinguish between asymptomatic (AS) and knee osteoarthritis (OA) gait patterns using ground reaction force (GRF) measurements, and (2) to investigate OA severity within OA gait patterns. Features were first extracted from the GRF vectors to be used for classification. We investigated a two-level hierarchical classification and analysis method using the nearest neighbor rule. At the first level, the GRF data were classified into two classes: AS and OA. At the second level, the GRF data of OA patients were classified according to the pathology severity. The OA patients were grouped into two OA severity categories according to the Kellgren and Lawrence (KL) scale: KL 1 and KL 2 for one category, and KL 3 and KL 4 for the other. Experiments were conducted using data of 42 cases, 16 AS and 26 pathological. The method discriminated between AS and OA subjects with an accuracy of 38 of 42 cases, and assessed the severity correctly with an accuracy of 20 of 26 cases. These results demonstrated the validity of both, the feature and the classifier, for automatic classification of AS and knee OA gait patterns and for analysis of OA severity.

Personalization of serious games is an important factor for motivating and engaging players. It requires the identification of players' profiles through the analysis of large volume of data including game data. This research study aims at identifying relevant data from an online serious game and the appropriate data mining methods for deduction of players' profiles. Multiple linear regression is applied to analyze the influence of player's characteristics on his performance. Moreover, clustering technique is used, in particular K-means, to extract players' clusters and to identify their common characteristics. The regression models showed that the number of access to the game, completed quests and advantages used contribute significantly to the scores and the gaming duration, while the clustering revealed three forms of players' participation: beginner, intermediate and advanced; who interact with the game according to their experiences.

The purpose of this study is to investigate a new representation of shape and its use in handwritten online character recognition by a Kohonen associative memory. This representation is based on the empirical distribution of features such as tangents and tangent differences at regularly spaced points along the character signal. Recognition is carried out by a Kohonen neural network trained using the representation. In addition to the Euclidean distance traditionally used in the Kohonen training algorithm to measure the similarities among feature vectors, we also investigate the Kullback–Leibler divergence and the Hellinger distance, functions that measure distance between distributions. Furthermore, we perform operations (pruning and filtering) on the trained memory to improve its classification potency. We report on extensive experiments using a database of online Arabic characters produced without constraints by a large number of writers. Comparative results show the pertinence of the representation and the superior performance of the scheme.

The aim of this work is to develop an automatic computer method to distinguish between asymptomatic (AS) and osteoarthritis (OA) knee gait patterns using 3-D ground reaction force (GRF) measurements. GRF features are first extracted from the force vector variations as a function of time and then classified by the nearest neighbor rule. We investigated two different features: the coefficients of a polynomial expansion and the coefficients of a wavelet decomposition. We also analyzed the impact of each GRF component (vertical, anteroposterior, and medial lateral) on classification. The best discrimination rate (91%) was achieved with the wavelet decomposition using the anteroposterior and the medial lateral components. These results demonstrate the validity of the representation and the classifier for automatic classification of AS and OA knee gait patterns. They also highlight the relevance of the anteroposterior and medial lateral force components in gait pattern classification.

Recent studies have shown that the Gibbs density function is a good model for visual patterns and that its parameters can be learned from pattern category training data by a gradient algorithm optimizing a constrained entropy criterion. These studies represented each pattern category by a single density. However, the patterns in a category can be so complex as to require a representation spread over several densities to more accurately account for the shape of their distribution in the feature space. The purpose of the present study is to investigate a representation of visual pattern category by several Gibbs densities using a Kohonen neural structure. In this Gibbs density based Kohonen network, which we call a Gibbsian Kohonen network, each node stores the parameters of a Gibbs density. Collectively, these Gibbs densities represent the pattern category. The parameters are learned by a gradient update rule so that the corresponding Gibbs densities maximize entropy subject to reproducing observed feature statistics of the training patterns. We verified the validity of the method and the efficiency of the ensuing Gibbs density pattern representation on a handwritten character recognition application.

Surgical instrumentation for adolescent idiopathic scoliosis (AIS) is a complex procedure where selection of the appropriate curve segment to fuse, i.e., fusion region, is a challenging decision in scoliosis surgery. Currently, the Lenke classification model is used for fusion region evaluation and surgical planning. Retrospective evaluation of Lenke classification and fusion region results was performed.Using a database of 1,776 surgically treated AIS cases, we investigated a topologically ordered self organizing Kohonen network, trained using Cobb angle measurements, to determine the relationship between the Lenke class and the fusion region selection. Specifically, the purpose was twofold (1) produce two spatially matched maps, one of Lenke classes and the other of fusion regions, and (2) associate these two maps to determine where the Lenke classes correlate with the fused spine regions.Topologically ordered maps obtained using a multi-center database of surgically treated AIS cases, show that the recommended fusion region agrees with the Lenke class except near boundaries between Lenke map classes. Overall agreement was 88%.The Lenke classification and fusion region agree in the majority of adolescent idiopathic scoliosis when reviewed retrospectively. The results indicate the need for spinal fixation instrumentation variation associated with the Lenke classification.

Surgical instrumentation for the Adolescent idiopathic scoliosis (AIS) is a complex procedure involving many difficult decisions. Selection of the appropriate fusion level remains one of the most challenging decisions in scoliosis surgery. Currently, the Lenke classification model is generally followed in surgical planning. The purpose of our study is to investigate a computer aided method for Lenke classification and scoliosis fusion level selection. The method uses a self organizing neural network trained on a large database of surgically treated AIS cases. The neural network produces two maps, one of Lenke classes and the other of fusion levels. These two maps show that the Lenke classes are associated with the the proper fusion level categories everywhere in the map except at the Lenke class transitions. The topological ordering of the Cobb angles in the neural network justifies determining a patient scoliotic treatment instrumentation using directly the fusion level map rather than via the Lenke classification.

Recent studies have shown that the Gibbs density function can model complex patterns and that a constrained maximum entropy formulation affords a powerful means of estimating its parameters from pattern class data. The theory, developed in the context of learning a prior model of natural images, has been applied successfully to the synthesis of textures and shapes, and to pattern classification. The basic parameter estimation algorithm rests on gradient algorithm following the maximization under constraints of an entropy criterion. The purpose of this study is to investigate a Gibbsian Kohonen neural network, a Kohonen network which can learn these constrained maximum entropy Gibbs density parameters for pattern representation and classification. Experiments in classification of handwritten characters verify the validity and efficiency of the method.

The purpose of this study is to investigate handwritten online character recognition by Kohonen neural networks which learn class conditional Gibbs densities from training samples. The characters are represented by histograms (empirical distributions) of features. The Kohonen network learning algorithm implements a gradient ascent which maximizes an entropy criterion under constraints. Using a database of handwritten online Arabic characters produced without constraints by a large number of writers, we conducted extensive experiments which show the advantage of this Gibbsian Kohonen network over other classifiers such as a regular Kohonen neural network and a Gibbsian Bayes classifier.

The purpose of this study is to investigate a new representation of shape and its use in handwritten on-line character recognition by a Kohonen associative memory. This representation is based on the empirical distribution of features such as tangents, and tangent differences at regularly spaced points along the character signal. Recognition is carried out by a Kohonen neural network trained using the representation. In addition to the traditional Hudidean distance, functions such as the Kullback Leibler divergence and the Hellinger distance are investigated to evaluate similarity of feature vectors because these functions provide measures of distance between distributions. We report on extensive experiments using a database of on-line Arabic characters produced without constraints by a large number of writers. Comparative results show the representation relevance and the superior performance of the scheme.

The purpose of this study is to investigate a methodfor high performance on-line Arabic characters recognition.This method is based on the use of Kohonen mapsand their corresponding confusion matrices which serve toprune them of error-causing nodes, and to combine themconsequently. We use two Kohonen maps obtained usingtwo distinct character representations, namely, Fourier descriptorsand tangents extracted along the characters on-linesignals. The two Kohonen maps are then combined usinga majority vote decision rule that, for each character,favors the most reliable map. This combination, withoutadding in any significant way to the process complexity, affordsa much better recognition rate.

The paper investigates a new representation of shape and its use in handwritten on-line character recognition. This representation is based on the empirical distribution of features such as tangents, and tangent differences at distant points along the character signal. Recognition is carried out by a Kohonen associative memory (also called Kohonen self organizing feature map), trained using this representation, and the Hellinger distance, which measures the distance between distributions. We report on extensive experiments that show the pertinence of the representation and the superior performance of the scheme.

Neural networks have been applied to various pattern classification and recognition problems for their learning ability, discrimination power and generalization ability The neural network most referenced in the pattern recognition literature are the multi-layer perceptron, the Kohonen associative memory and the Capenter-Grossberg ART network. The Kohonen memory runs an unsupervised clustering algorithm. It is easily trained and has attractive properties such as topological ordering and good generalization. In this study an on-line system for the recognition of handwriting Arabic characters using a Kohonen network is investigated. The input of the neural network is a feature vector of elliptic Fourier coefficients extracted from the handwritten dynamic representation. Experimental results show that the network successfully recognizes both clearly and roughly written characters with good performance.

The purpose of this study was to identify meaningful gait patterns in knee frontal plane kinematics from a large population of asymptomatic individuals. The proposed method used principal component analysis (PCA). It first reduced the data dimensionality, without loss of relevant information, by projecting the original kinematic data onto a subspace of significant principal components (PCs). This was followed by a discriminant model to separate the individuals' gait into homogeneous groups. Four descriptive gait patterns were identified and validated by clustering silhouette width and statistical hypothesis testing. The first pattern was close to neutral during the stance phase and in adduction during the swing phase (Cluster 1). The second pattern was in abduction during the stance phase and tends into adduction during the swing phase (Cluster 2). The third pattern was close to neutral during the stance phase and in abduction during the swing phase (Cluster 3) and the fourth was in abduction during both the stance and the swing phase (Cluster 4).