Perfect Classification Accuracy Research Articles

Machine learning approach yields epigenetic biomarkers of food allergy: A novel 13-gene signature to diagnose clinical reactivity.

BackgroundCurrent laboratory tests are less than 50% accurate in distinguishing between people who have food allergies (FA) and those who are merely sensitized to foods, resulting in the use of expensive and potentially dangerous Oral Food Challenges. This study presents a purely-computational machine learning approach, conducted using DNA Methylation (DNAm) data, to accurately diagnose food allergies and potentially find epigenetic targets for the disease.Methods and resultsAn unbiased feature-selection pipeline was created that narrowed down 405,000+ potential CpG biomarkers to 18. Machine-learning models that utilized subsets of this 18-feature aggregate achieved perfect classification accuracy on completely hidden test cohorts (on an 8-fold hidden dataset). Ensemble classification was also shown to be effective for this High Dimension Low Sample Size (HDLSS) DNA methylation dataset. The efficacy of these machine learning classifiers and the 18 CpGs was further validated by their high accuracy on a large number of hidden data permutations, where the samples in the training, cross-validation, and hidden sets were repeatedly randomly allocated. The 18-CpG signature mapped to 13 genes, on which biological insights were collected. Notably, many of the FA-discriminating genes found in this study were strongly associated with the immune system, and seven of the 13 genes were previously associated with FA.ConclusionsPrevious studies have also created highly-accurate classifiers for this dataset, using both data-driven and a priori biological insights to construct a 96-CpG signature. This research builds on previous work because it uses a completely computational approach to obtain a perfect classification accuracy while using only 18 highly discriminating CpGs (0.005% of the total available features). In machine learning, simpler models, as used in this study, are generally preferred over more complex ones (other things being equal). Lastly, the completely data-driven methodology presented in this research eliminates the need for a priori biological information and allows for generalizability to other DNAm classification problems.

Clinical assessment of a non-invasive wearable MEMS pressure sensor array for monitoring of arterial pulse waveform, heart rate and detection of atrial fibrillation

There is an unmet clinical need for a low cost and easy to use wearable devices for continuous cardiovascular health monitoring. A flexible and wearable wristband, based on microelectromechanical sensor (MEMS) elements array was developed to support this need. The performance of the device in cardiovascular monitoring was investigated by (i) comparing the arterial pressure waveform recordings to the gold standard, invasive catheter recording (n = 18), (ii) analyzing the ability to detect irregularities of the rhythm (n = 7), and (iii) measuring the heartrate monitoring accuracy (n = 31). Arterial waveforms carry important physiological information and the comparison study revealed that the recordings made with the wearable device and with the gold standard device resulted in almost identical (r = 0.9–0.99) pulse waveforms. The device can measure the heart rhythm and possible irregularities in it. A clustering analysis demonstrates a perfect classification accuracy between atrial fibrillation (AF) and sinus rhythm. The heartrate monitoring study showed near perfect beat-to-beat accuracy (sensitivity = 99.1%, precision = 100%) on healthy subjects. In contrast, beat-to-beat detection from coronary artery disease patients was challenging, but the averaged heartrate was extracted successfully (95% CI: −1.2 to 1.1 bpm). In conclusion, the results indicate that the device could be useful in remote monitoring of cardiovascular diseases and personalized medicine.

Deep neural networks for bot detection

The problem of detecting bots, automated social media accounts governed by software but disguising as human users, has strong implications. For example, bots have been used to sway political elections by distorting online discourse, to manipulate the stock market, or to push anti-vaccine conspiracy theories that may have caused health epidemics. Most techniques proposed to date detect bots at the account level, by processing large amounts of social media posts, and leveraging information from network structure, temporal dynamics, sentiment analysis, etc. In this paper, we propose a deep neural network based on contextual long short-term memory (LSTM) architecture that exploits both content and metadata to detect bots at the tweet level: contextual features are extracted from user metadata and fed as auxiliary input to LSTM deep nets processing the tweet text. Another contribution that we make is proposing a technique based on synthetic minority oversampling to generate a large labeled dataset, suitable for deep nets training, from a minimal amount of labeled data (roughly 3000 examples of sophisticated Twitter bots). We demonstrate that, from just one single tweet, our architecture can achieve high classification accuracy (AUC > 96%) in separating bots from humans. We apply the same architecture to account-level bot detection, achieving nearly perfect classification accuracy (AUC > 99%). Our system outperforms previous state of the art while leveraging a small and interpretable set of features, yet requiring minimal training data.

A Comparison between Theoretical and Experimental Measures of Consciousness as Integrated Information in an Anatomically Based Network of Coupled Oscillators

The rise of mathematical developments in the theories of consciousness has led to new measures to detect consciousness in a system. The Integrated Information Theory (IIT) is one of the best mathematical rooted attempts to quantify the level of consciousness in a system with Φ as the effective information generated in a system above its parts. Recently, the IIT has inspired the Perturbational Complexity Index (PCI) to detect conscious states in patients with disorders of consciousness, and it has shown to have almost perfect classification accuracy. In this study, we explore the statistical correspondence between the theoretical Φ and the experimental PCI through a neurocomputational model of coupled oscillators that can be artificially perturbed, which mainly focuses on the dynamics of collective synchronization between subsets of brain areas. Our results reveal that both measures are statistically related but, in principle, this relationship is far to be perfect. These results are discussed in the context of the model of coupled oscillators, which mainly focuses on the dynamics of collective synchronization between subsets of brain areas.

Steel surface defect classification using multiple hyper-spheres support vector machine with additional information

A novel multiple hyper-spheres support vector machine with additional information (MHSVM+) is proposed for multi-class steel surface defects classification. Originated from binary twin hyper-spheres support vector machine, MHSVM+ uses hyper-sphere to solve classification decision problem. Differently, MHSVM+ is a multi-class classifier, where it builds a corresponding hyper-sphere for each type of defect dataset. Moreover, MHSVM+ introduces learning paradigm using additional information, which means it can learn additional information hidden in defect dataset. Two types of additional information are provided: local neighbor information and local density information. Local neighbor information contains local classification results for defect samples. And local density information is used to capture label noise, isolated samples and important samples in defect dataset. The above two types of additional information are introduced into MHSVM+ model. Finally, MHSVM+ classifier is used to classify six types of steel surface defects. Experimental results show that the novel multi-class classifier has perfect classification accuracy for defect dataset, especially corrupted defect dataset.

Prediction of friction stir weld quality without and with signal features

Building a reliable prediction model can mitigate the need for actual experiments, hence saving time and cost. To this end, this study presents a methodology to predict weld quality for a particular friction stir weld configuration using machine learning and metaheuristic algorithms including K-nearest neighbor (KNN), fuzzy KNN (FKNN), and the artificial bee colony (ABC). The ABC algorithm was utilized to determine the best (F)KNN model with optimal K value and feature subset. First, models were built based on only experimental conditions including spindle rotational speed, plunge force, and feed rate, as well as derived values including a speed ratio and an empirical force index (EFI). The best model was identified to be 1-NN comprised of three features, i.e., rotational speed, feed rate, and EFI, with 93.16% classification accuracy based on leave-one-out cross-validation. The majority of data points leading to error were found to lie mostly on the boundaries between classes. It was shown that classification error could be reduced by removing those points, which is cheating and not recommended. Instead, it is recommended to improve classification accuracy without omitting dissenting data by introducing additional information to better distinguish misclassified data points. To this end, wavelet energy features extracted from weld signals of X-Force, Y-Force, spindle rotational speed, feed rate, and plunge force were added to the original feature pool. In order to determine the impact of each weld signal feature set, each signal feature set was individually tested. After applying ABC to the expanded feature pool to build the best model, perfect classification accuracy was achieved in several cases. The results suggest that adding signal features can greatly improve the effectiveness of model predictability of friction stir weld quality.

IMPLEMENTATION OF A SMARTPHONE WIRELESS GYROSCOPE PLATFORM WITH MACHINE LEARNING FOR CLASSIFYING DISPARITY OF A HEMIPLEGIC PATELLAR TENDON REFLEX PAIR

The patellar tendon reflex response provides fundamental means of assessing a subject’s neurological health. Dysfunction regarding the characteristics of the reflex response may warrant the escalation to more advanced diagnostic techniques. Current strategies involve the manual elicitation of the patellar tendon reflex by a highly skilled clinician with subsequent interpretation according to an ordinal scale. The reliability of the ordinal scale approach is a topic of contention. Highly skilled clinicians have been in disagreement regarding even the observation of asymmetric reflex pairs. An alternative strategy incorporated the ubiquitous smartphone with a software application to function as a wireless gyroscope platform for quantifying the reflex response. Each gyroscope signal recording of the reflex response can be conveyed wirelessly through Internet connectivity as an email attachment. The reflex response is evoked through a potential energy impact pendulum that enables prescribed targeting and potential energy level. The smartphone functioning as a wireless gyroscope platform reveals an observationally representative gyroscope signal of the reflex response. Three notably distinguishable attributes of the reflex response are incorporated into a feature set for machine learning: maximum angular rate of rotation, minimum angular rate of rotation, and time disparity between maximum and minimum angular rate of rotation. Four machine learning platforms such as the J48 decision tree, K-nearest neighbors, logistic regression, and support vector machine, were applied to the patellar tendon reflex response feature set incorporating a hemiplegic patellar tendon reflex pair. The J48 decision tree attained 98% classification accuracy, and the K-nearest neighbors, logistic regression, and support vector machine achieved perfect classification accuracy for distinguishing between a hemiplegic affected leg and unaffected leg patellar tendon reflex pair. The research findings reveal the potential of machine learning for enabling advanced diagnostic acuity respective of the gyroscope signal of the patellar tendon reflex response.

Prediction of lithium response in first-episode mania using the LITHium Intelligent Agent (LITHIA): Pilot data and proof-of-concept.

Individualized treatment for bipolar disorder based on neuroimaging treatment targets remains elusive. To address this shortcoming, we developed a linguistic machine learning system based on a cascading genetic fuzzy tree (GFT) design called the LITHium Intelligent Agent (LITHIA). Using multiple objectively defined functional magnetic resonance imaging (fMRI) and proton magnetic resonance spectroscopy (1 H-MRS) inputs, we tested whether LITHIA could accurately predict the lithium response in participants with first-episode bipolar mania. We identified 20 subjects with first-episode bipolar mania who received an adequate trial of lithium over 8weeks and both fMRI and 1 H-MRS scans at baseline pre-treatment. We trained LITHIA using 18 1 H-MRS and 90 fMRI inputs over four training runs to classify treatment response and predict symptom reductions. Each training run contained a randomly selected 80% of the total sample and was followed by a 20% validation run. Over a different randomly selected distribution of the sample, we then compared LITHIA to eight common classification methods. LITHIA demonstrated nearly perfect classification accuracy and was able to predict post-treatment symptom reductions at 8weeks with at least 88% accuracy in training and 80% accuracy in validation. Moreover, LITHIA exceeded the predictive capacity of the eight comparator methods and showed little tendency towards overfitting. The results provided proof-of-concept that a novel GFT is capable of providing control to a multidimensional bioinformatics problem-namely, prediction of the lithium response-in a pilot data set. Future work on this, and similar machine learning systems, could help assign psychiatric treatments more efficiently, thereby optimizing outcomes and limiting unnecessary treatment.

Smartphone-based colorimetric detection via machine learning.

We report the application of machine learning to smartphone-based colorimetric detection of pH values. The strip images were used as the training set for Least Squares-Support Vector Machine (LS-SVM) classifier algorithms that were able to successfully classify the distinct pH values. The difference in the obtained image formats was found not to significantly affect the performance of the proposed machine learning approach. Moreover, the influence of the illumination conditions on the perceived color of pH strips was investigated and further experiments were conducted to study the effect of color change on the learning model. Non-integer pH levels are identified as their nearest integer pH values, whereas the test results for integer pH levels using JPEG, RAW and RAW-corrected image formats captured under different lighting conditions lead to perfect classification accuracy, sensitivity and specificity, which proves that colorimetric detection using machine learning based systems is able to adapt to various experimental conditions and is a great candidate for smartphone-based sensing in paper-based colorimetric assays.

Multistage feature selection approach for high-dimensional cancer data

Cancer is a serious disease that causes death worldwide. DNA methylation (DNAm) is an epigenetic mechanism, which controls the regulation of gene expression and is useful in early detection of cancer. The challenge with DNA methylation microarray datasets is the huge number of CpG sites compared to the number of samples. Recent research efforts attempted to reduce this high dimensionality by different feature selection techniques. This article proposes a multistage feature selection approach to select the optimal CpG sites from three different DNAm cancer datasets (breast, colon and lung). The proposed approach combines three different filter feature selection methods including Fisher Criterion, t-test and Area Under ROC Curve. In addition, as a wrapper feature selection, we apply genetic algorithms with Support Vector Machine Recursive Feature Elimination (SVM-RFE) as its fitness function, and SVM as its evaluator. Using the Incremental Feature Selection (IFS) strategy, subsets of 24, 13 and 27 optimal CpG sites are selected for the breast, colon and lung cancer datasets, respectively. By applying fivefold cross-validation on the training datasets, these subsets of optimal CpG sites showed perfect classification accuracies of 100, 100 and 97.67%, respectively. Moreover, the testing of the three independent cancer datasets by these final subsets resulted in accuracies 96.02, 98.81 and 94.51%, respectively. The experimental results demonstrated high classification performance and small optimal feature subsets. Consequently, the biological significance of the genes corresponding to these feature subsets is validated using enrichment analysis.

Overcoming data deficiency in reptiles

We have no information on the risk of extinction of 21% of reptiles listed as Data Deficient on the Sampled Red List Index (SRLI), an indicator developed to track global change in species status. Data Deficient species are of high research priority, because they contribute to uncertainty in estimates of extinction risk and are neglected by conservation programmes. We review the causes of data deficiency in reptiles; the likely status of Data Deficient reptiles; and possible solutions for their re-assessment. We find that 52% of Data Deficient reptiles lack information on population status and trends, and that few species are only known from type specimens and old records. We build a random forest model for SRLI species of known extinction risk, based on life-history, environmental and threat information. The final model shows perfect classification accuracy (100%) in ten-fold cross validation. We use the model to predict that 56 of 292 Data Deficient reptiles (19%) are at risk of extinction, so the overall proportion of threatened reptiles in the SRLI (19%) remains unchanged. Regions predicted to contain large numbers of threatened Data Deficient reptiles overlap with known centres of threatened species richness. However, the model shows lower accuracy (79%) on 29 species recently re-assessed in the Global Reptile Assessment. Predictive models could be used to prioritize Data Deficient species and reptiles not included in the SRLI, and new reptile assessments could be used to improve model predictions through adaptive learning.

Quaternion-Based Gesture Recognition Using Wireless Wearable Motion Capture Sensors.

This work presents the development and implementation of a unified multi-sensor human motion capture and gesture recognition system that can distinguish between and classify six different gestures. Data was collected from eleven participants using a subset of five wireless motion sensors (inertial measurement units) attached to their arms and upper body from a complete motion capture system. We compare Support Vector Machines and Artificial Neural Networks on the same dataset under two different scenarios and evaluate the results. Our study indicates that near perfect classification accuracies are achievable for small gestures and that the speed of classification is sufficient to allow interactivity. However, such accuracies are more difficult to obtain when a participant does not participate in training, indicating that more work needs to be done in this area to create a system that can be used by the general population.

Abstract T P176: Classification of DWI Lesion Patterns in Acute Ischemic Stroke Using Shape Context

Introduction: Neuroimaging studies in ischemic stroke hold abundant features about location, type, and extent of brain ischemia. Identification of imaging characteristics that predict recovery of brain tissue at risk remains an ongoing effort and such DWI patterns hold promise. However, systematic analysis of large multi-modal datasets face many challenges. An automated algorithm for classification of DWI lesion patterns in real-time would be ideal. We developed and evaluated a computer vision model that uses 3D shape description of the lesion for classification of DWI lesion patterns into an existing nomenclature. Methods: Inclusion criteria were acute ischemic stroke and DWI performed within 24 hours of symptom onset. Acute lesions were manually segmented on DWI and categorized into 6 types (territorial, other cortical, small superficial, internal border zone, small deep, and other deep infarcts). The computer vision model characterized stroke lesion using a shape context descriptor that accumulates surface points of the lesion into a 3D log-linear spherical histogram. The experiments estimate the accuracy of the model in classifying the lesion patterns using a leave-one-out cross-validation. Results: A total of 344 patients satisfied inclusion criteria. Mean age was 66.1 (range 13-97). Median NIHSS at baseline was 14 (range 0-38). Average lesion volume was 75.6 cc (range 5-256). The number of observations per category was [territorial:81, other cortical:51, small superficial:78, internal border zone:12, small deep:34, other deep infarcts:88]. When combining lesion volume, average location relative to the center of the brain, and shape descriptor as input, the classification models yielded perfect classification accuracy over the entire dataset. Conclusions: An advanced computer vision framework that automatically identifies discriminant features between lesion categories was introduced and successfully evaluated on 344 patients with acute stroke. Our study enables even larger scale retrospective analyses. Computer vision and pattern recognition methods can play a central role for the systematic analysis of big data in stroke.

A Feature-Space Indicator Kriging Approach for Remote Sensing Image Classification

An indicator kriging (IK) approach for remote sensing image classification is proposed. By introducing indicator variables for categorical data, the work of image classification is transformed into estimation of class-dependent probabilities in feature space using ordinary kriging. Individual pixels are then assigned to the class with maximum class probability. The approach is distribution free and yields perfect classification accuracies for training data provided that collocated data in feature space do not exist. Technical considerations regarding implementation of IK such as indicator semivariogram modeling and handling of collocated data in feature space are also described. The IK, Gaussian-based maximum likelihood, nearest neighbor, and support vector machine (SVM) classifiers were applied to study areas within the Shimen reservoir watershed (case A: FORMOSAT-2) and Taipei city (case B: SPOT 4). The results show that the overall accuracies of the proposed IK classifier and SVM can achieve higher than 97% for training data and 81% for testing data. (The overall accuracies of IK are a little higher than those of SVM.) IK and SVM are found to be superior to the other two classifiers in terms of overall accuracies for both training and testing data. The proposed IK classifier has the following advantages: 1) It can deal with anisotropic problem in feature space; 2) it is a nonparametric method and needs not to know the type of probability distribution; and 3) it yields 100% classification accuracy for the training data provided that collocated data in feature space do not exist.

A sparse representation based approach for recognition of power system transients

The present work shows how a new genre of problem solution techniques called compressive sensing (CI) can be utilized for solving power quality problems. Specifically speaking, this work undertakes the power system transient recognition problem and solves it using sparse representation based classification (SRC) technique, a popular, recently emerging technique in CI. The method employs SRC based classification in combination with a feature extraction procedure, carried out using dual tree complex wavelet transform (DTCWT). The proposed method uses statistical characteristics to extract features from the DTCWT coefficients obtained from each signal and then these extracted features are used as input arrays for the SRC based classifier. The PQ disturbance events considered in this work include four different transient conditions, namely, transients due to capacitor switching, transformer inrush currents, transients due to motor switching and transients due to short circuit faults. The proposed algorithm could achieve perfect classification accuracy of the order of 100% and could comfortably outperform several similar contemporary methods known for identical power system transient classification problems.

Assuring the authenticity of northwest Spain white wine varieties using machine learning techniques

Classification of wine represents a multi-criteria decision-making problem characterized by great complexity, non-linearity and lack of objective information regarding the quality of the desired final product. Volatile compounds of wines elaborated from four Galician (NW Spain) autochthonous white Vitis vinifera from four consecutive vintages were analysed by gas chromatography (FID, FPD and MS detectors), and several aroma compounds were used for correctly classifying autochthonous white grape varieties (Albariño, Treixadura, Loureira and Dona Branca). The objective of the work is twofold: to find a classification model able to precisely differentiate between existing grape varieties (i.e. assuring the authenticity), and to assess the discriminatory power of different family compounds over well-known classifiers (i.e. guaranteeing the typicality). From the experiments carried out, and given the fact that Principal Component Analysis (PCA) was not able to accurately separate all the wine varieties, this work investigates the suitability of applying different machine learning (ML) techniques (i.e.: Support Vector Machines, Random Forests, MultiLayer Perceptron, k-Nearest Neighbour and Naïve Bayes) for classification purposes. Perfect classification accuracy is obtained by the Random Forest algorithm, whilst the other alternatives achieved promising results using only part of the available information.

Diagnosis of bladder cancers with small sample size via feature selection

This paper proposes feature selection as an approach to deal with a bladder cancer data set with small sample size. Three feature selection methods and four classifiers were used to determine the best feature subsets that produce perfect classification accuracy. The smallest best feature subsets are used to build neural models with the small data set to achieve 100% training and testing accuracies. Therefore, the mega-trend-diffusion technique proposed by Li et al. to produce artificial samples is actually unnecessary. The similarity classifier proposed by Luukka was also applied to the small data set with the smallest best feature subsets to achieve 100% accuracy using only 4 samples (two with bladder cancer and two normal) for two selected p and m values. Given the same accuracy, using the best feature subsets selected is better than using all 13 features as done by Luukka. Furthermore, several indexes/methods commonly used in filtering feature selection methods were tested for their ability to find the best feature subsets for this particular small bladder cancer data set.

A Comparative Study of Local Matching Approach for Face Recognition

In contrast to holistic methods, local matching methods extract facial features from different levels of locality and quantify them precisely. To determine how they can be best used for face recognition, we conducted a comprehensive comparative study at each step of the local matching process. The conclusions from our experiments include: (1) additional evidence that Gabor features are effective local feature representations and are robust to illumination changes; (2) discrimination based only on a small portion of the face area is surprisingly good; (3) the configuration of facial components does contain rich discriminating information and comparing corresponding local regions utilizes shape features more effectively than comparing corresponding facial components; (4) spatial multiresolution analysis leads to better classification performance; (5) combining local regions with Borda count classifier combination method alleviates the curse of dimensionality. We implemented a complete face recognition system by integrating the best option of each step. Without training, illumination compensation and without any parameter tuning, it achieves superior performance on every category of the FERET test: near perfect classification accuracy (99.5%) on pictures taken on the same day regardless of indoor illumination variations, and significantly better than any other reported performance on pictures taken several days to more than a year apart. The most significant experiments were repeated on the AR database, with similar results.