A medical diagnostic tool based on radial basis function classifiers and evolutionary simulated annealing

ObjectiveThe profusion of data accumulating in the form of medical records could be of great help for developing medical decision support systems. The objective of this paper is to present a methodology for designing data-driven medical diagnostic tools, based on neural network classifiers.MethodsThe proposed approach adopts the radial basis function (RBF) neural network architecture and the non-symmetric fuzzy means (NSFM) training algorithm, which presents certain advantages including better approximation capabilities and shorter computational times. The novelty in this work consists of adapting the NSFM algorithm to train RBF classifiers, and suitably tailoring the evolutionary simulated annealing (ESA) technique to optimize the produced RBF models. The integration of ESA is critical as it helps the optimization procedure to escape from local minima, which could arise from the application of the traditional simulated annealing algorithm, and thus discover improved solutions. The resulting method is evaluated in nine different medical benchmark datasets, where the common objective is to train a suitable classifier. The evaluation includes a comparison with two different schemes for training classifiers, including a standard RBF training technique and support vector machines (SVMs). Accuracy% and the Matthews Correlation Coefficient (MCC) are used for comparing the performance of the three classifiers.ResultsResults show that the use of ESA helps to greatly improve the performance of the NSFM algorithm and provide satisfactory classification accuracy. In almost all benchmark datasets, the best solution found by the ESA-NSFM algorithm outperforms the results produced by the SFM algorithm and SVMs, considering either the accuracy% or the MCC criterion. Furthermore, in the majority of datasets, the average solution of the ESA-NSFM population is statistically significantly higher in terms of accuracy% and MCC at the 95% confidence level, compared to the global optimum solution that its rivals could achieve. As far as computational times are concerned, the proposed approach was found to be faster compared to SVMs.ConclusionsThe results of this study suggest that the ESA-NSFM algorithm can form the basis of a generic method for knowledge extraction from data originating from different kinds of medical records. Testing the proposed approach on a number of benchmark datasets, indicates that it provides increased diagnostic accuracy in comparison with two different classifier training methods.     

基于PCA和LDA方法的肿瘤基因表达谱数据分类

目的基因芯片技术对医学临床诊断、治疗、药物开发和筛选等技术的发展具有革命性的影响。针对高维医学数据降维困难及基因表达谱样本数据少、维度高、噪声大的特点,维数约减十分必要。基于主成分分析（principalcomponentanalysis,PCA）和线性判别分析（1ineardiscriminantanalysis,LDA）方法,有效解决了基因表达谱数据分类问题,并提高了识别率。方法分别引人PCA和LDA方法对基因表达谱数据进行降维,然后用K近邻（K—nearestneighbor,KNN）作为分类器对数据进行分类,并分别在乳腺癌和卵巢癌质谱数据上。结果在两类癌症质谱数据上应用PCA和LDA方法能够有效提取分类特征信息,并在保持较高分类正确率的前提下大幅度降低医学数据的维数。结论利用维数约减的方法对癌症基因表达谱数据进行分类,可辅助临床医生发现新的疾病特征,提高疾病诊断的正确率。     

Vicinal support vector classifier using supervised kernel-based clustering

ObjectiveSupport vector machines (SVMs) have drawn considerable attention due to their high generalisation ability and superior classification performance compared to other pattern recognition algorithms. However, the assumption that the learning data is identically generated from unknown probability distributions may limit the application of SVMs for real problems. In this paper, we propose a vicinal support vector classifier (VSVC) which is shown to be able to effectively handle practical applications where the learning data may originate from different probability distributions.MethodsThe proposed VSVC method utilises a set of new vicinal kernel functions which are constructed based on supervised clustering in the kernel-induced feature space. Our proposed approach comprises two steps. In the clustering step, a supervised kernel-based deterministic annealing (SKDA) clustering algorithm is employed to partition the training data into different soft vicinal areas of the feature space in order to construct the vicinal kernel functions. In the training step, the SVM technique is used to minimise the vicinal risk function under the constraints of the vicinal areas defined in the SKDA clustering step.ResultsExperimental results on both artificial and real medical datasets show our proposed VSVC achieves better classification accuracy and lower computational time compared to a standard SVM. For an artificial dataset constructed from non-separated data, the classification accuracy of VSVC is between 95.5% and 96.25% (using different cluster numbers) which compares favourably to the 94.5% achieved by SVM. The VSVC training time is between 8.75 s and 17.83 s (for 2–8 clusters), considerable less than the 65.0 s required by SVM. On a real mammography dataset, the best classification accuracy of VSVC is 85.7% and thus clearly outperforms a standard SVM which obtains an accuracy of only 82.1%. A similar performance improvement is confirmed on two further real datasets, a breast cancer dataset (74.01% vs. 72.52%) and a heart dataset (84.77% vs. 83.81%), coupled with a reduction in terms of learning time (32.07 s vs. 92.08 s and 25.00 s vs. 53.31 s, respectively). Furthermore, the VSVC results in the number of support vectors being equal to the specified cluster number, and hence in a much sparser solution compared to a standard SVM.ConclusionIncorporating a supervised clustering algorithm into the SVM technique leads to a sparse but effective solution, while making the proposed VSVC adaptive to different probability distributions of the training data.     

A Q-backpropagated time delay neural network for diagnosing severity of gait disturbances in Parkinson’s disease

Parkinson’s disease (PD) is a movement disorder that affects the patient’s nervous system and health-care applications mostly uses wearable sensors to collect these data. Since these sensors generate time stamped data, analyzing gait disturbances in PD becomes challenging task. The objective of this paper is to develop an effective clinical decision-making system (CDMS) that aids the physician in diagnosing the severity of gait disturbances in PD affected patients. This paper presents a Q-backpropagated time delay neural network (Q-BTDNN) classifier that builds a temporal classification model, which performs the task of classification and prediction in CDMS. The proposed Q-learning induced backpropagation (Q-BP) training algorithm trains the Q-BTDNN by generating a reinforced error signal. The network’s weights are adjusted through backpropagating the generated error signal. For experimentation, the proposed work uses a PD gait database, which contains gait measures collected through wearable sensors from three different PD research studies. The experimental result proves the efficiency of Q-BP in terms of its improved classification accuracy of 91.49%, 92.19% and 90.91% with three datasets accordingly compared to other neural network training algorithms.     

CARSVM: a class association rule-based classification framework and its application to gene expression data

OBJECTIVE: In this study, we aim at building a classification framework, namely the CARSVM model, which integrates association rule mining and support vector machine (SVM). The goal is to benefit from advantages of both, the discriminative knowledge represented by class association rules and the classification power of the SVM algorithm, to construct an efficient and accurate classifier model that improves the interpretability problem of SVM as a traditional machine learning technique and overcomes the efficiency issues of associative classification algorithms. METHOD: In our proposed framework: instead of using the original training set, a set of rule-based feature vectors, which are generated based on the discriminative ability of class association rules over the training samples, are presented to the learning component of the SVM algorithm. We show that rule-based feature vectors present a high-qualified source of discrimination knowledge that can impact substantially the prediction power of SVM and associative classification techniques. They provide users with more conveniences in terms of understandability and interpretability as well. RESULTS: We have used four datasets from UCI ML repository to evaluate the performance of the developed system in comparison with five well-known existing classification methods. Because of the importance and popularity of gene expression analysis as real world application of the classification model, we present an extension of CARSVM combined with feature selection to be applied to gene expression data. Then, we describe how this combination will provide biologists with an efficient and understandable classifier model. The reported test results and their biological interpretation demonstrate the applicability, efficiency and effectiveness of the proposed model. CONCLUSION: From the results, it can be concluded that a considerable increase in classification accuracy can be obtained when the rule-based feature vectors are integrated in the learning process of the SVM algorithm. In the context of applicability, according to the results obtained from gene expression analysis, we can conclude that the CARSVM system can be utilized in a variety of real world applications with some adjustments.     

Biomedical named entity recognition using two-phase model based on SVMs

Named entity (NE) recognition has become one of the most fundamental tasks in biomedical knowledge acquisition. In this paper, we present a two-phase named entity recognizer based on SVMs, which consists of a boundary identification phase and a semantic classification phase of named entities. When adapting SVMs to named entity recognition, the multi-class problem and the unbalanced class distribution problem become very serious in terms of training cost and performance. We try to solve these problems by separating the NE recognition task into two subtasks, where we use appropriate SVM classifiers and relevant features for each subtask. In addition, by employing a hierarchical classification method based on ontology, we effectively solve the multi-class problem concerning semantic classification. The experimental results on the GENIA corpus show that the proposed method is effective not only in reducing computational cost but also in improving performance. The F-score (beta=1) for the boundary identification is 74.8 and the F-score for the semantic classification is 66.7.     

A reliable method for cell phenotype image classification

OBJECTIVE: Image-based approaches have proven to be of great utility in the automated cell phenotype classification, it is very important to develop a method that efficiently quantifies, distinguishes and classifies sub-cellular images. METHODS AND MATERIALS: In this work, the invariant locally binary patterns (LBP) are applied, for the first time, to the classification of protein sub-cellular localization images. They are tested on three image datasets (available for download), in conjunction with support vector machines (SVMs) and random subspace ensembles of neural networks. Our method based on invariant LBP provides higher accuracy than other well-known methods for feature extraction; moreover, our method does not require to (direct) crop the cells for the classification. RESULTS AND CONCLUSION: The experimental results show that the random subspace ensemble of neural networks outperforms the SVM in this problem. The proposed approach based on the solely LBP features gives accuracies of 85%, 93.9% and 88.4% on the 2D HeLa dataset, LOCATE endogenous and transfected datasets, respectively, and in combination with other state-of-the-art methods for the cell phenotype image classification we obtain a classification accuracy of 94.2%, 98.4% and 96.5%.     

Ensemble adaptive network-based fuzzy inference system with weighted arithmetical mean and application to diagnosis of optic nerve disease from visual-evoked potential signals

OBJECTIVE: This paper presents a new method based on combining principal component analysis (PCA) and adaptive network-based fuzzy inference system (ANFIS) to diagnose the optic nerve disease from visual-evoked potential (VEP) signals. The aim of this study is to improve the classification accuracy of ANFIS classifier on diagnosis of optic nerve disease from VEP signals. With this aim, a new classifier ensemble based on ANFIS and PCA is proposed. METHODS AND MATERIAL: The VEP signals dataset include 61 healthy subjects and 68 patients suffered from optic nerve disease. First of all, the dimension of VEP signals dataset with 63 features has been reduced to 4 features using PCA. After applying PCA, ANFIS trained using three different training-testing datasets randomly with 50-50% training-testing partition. RESULTS: The obtained classification results from ANFIS trained separately with three different training-testing datasets are 96.87%, 98.43%, and 98.43%, respectively. And then the results of ANFIS trained with three different training-testing datasets randomly with 50-50% training-testing partition have been combined with three different ways including weighted arithmetical mean that proposed firstly by us, arithmetical mean, and geometrical mean. The classification results of ANFIS combined with three different ways are 98.43%, 100%, and 100%, respectively. Also, ensemble ANFIS has been compared with ANN ensemble. ANN ensemble obtained 98.43%, 100%, and 100% prediction accuracy with three different ways including arithmetical mean, geometrical mean and weighted arithmetical mean. CONCLUSION: These results have shown that the proposed classifier ensemble approach based on ANFIS trained with different train-test datasets and PCA has produced very promising results in the diagnosis of optic nerve disease from VEP signals.     

The problem of bias in training data in regression problems in medical decision support.

This paper describes a bias problem encountered in a machine learning approach to outcome prediction in anticoagulant drug therapy. The outcome to be predicted is a measure of the clotting time for the patient; this measure is continuous and so the prediction task is a regression problem. Artificial neural networks (ANNs) are a powerful mechanism for learning to predict such outcomes from training data. However, experiments have shown that an ANN is biased towards values more commonly occurring in the training data and is thus, less likely to be correct in predicting extreme values. This issue of bias in training data in regression problems is similar to the associated problem with minority classes in classification. However, this bias issue in classification is well documented and is an on-going area of research. In this paper, we consider stratified sampling and boosting as solutions to this bias problem and evaluate them on this outcome prediction problem and on two other datasets. Both approaches produce some improvements with boosting showing the most promise.     

Novel network architecture and learning algorithm for the classification of mass abnormalities in digitized mammograms

OBJECTIVE: The main objective of this paper is to present a novel learning algorithm for the classification of mass abnormalities in digitized mammograms. METHODS AND MATERIAL: The proposed approach consists of new network architecture and a new learning algorithm. The original idea is based on the introduction of an additional neuron in the hidden layer for each output class. The additional neurons for benign and malignant classes help in improving memorization ability without destroying the generalization ability of the network. The training is conducted by combining minimal distance-based similarity/random weights and direct calculation of output weights. RESULTS: The proposed approach can memorize training patterns with 100% retrieval accuracy as well as achieve high generalization accuracy for patterns which it has never seen before. The grey-level and breast imaging reporting and data system-based features from digitized mammograms are extracted and used to train the network with the proposed architecture and learning algorithm. The best results achieved by using the proposed approach are 100% on training set and 94% on test set. CONCLUSION: The proposed approach produced very promising results. It has outperformed existing classification approaches in terms of classification accuracy, generalization and memorization abilities, number of iterations, and guaranteed training on a benchmark database.     

Predictive modeling of schizophrenia from genomic data: Comparison of polygenic risk score with kernel support vector machines approach

A major controversy in psychiatric genetics is whether nonadditive genetic interaction effects contribute to the risk of highly polygenic disorders. We applied a support vector machines (SVMs) approach, which is capable of building linear and nonlinear models using kernel methods, to classify cases from controls in a large schizophrenia case–control sample of 11,853 subjects (5,554 cases and 6,299 controls) and compared its prediction accuracy with the polygenic risk score (PRS) approach. We also investigated whether SVMs are a suitable approach to detecting nonlinear genetic effects, that is, interactions. We found that PRS provided more accurate case/control classification than either linear or nonlinear SVMs, and give a tentative explanation why PRS outperforms both multivariate regression and linear kernel SVMs. In addition, we observe that nonlinear kernel SVMs showed higher classification accuracy than linear SVMs when a large number of SNPs are entered into the model. We conclude that SVMs are a potential tool for assessing the presence of interactions, prior to searching for them explicitly.     

Mapping high-dimensional data onto a relative distance plane--an exact method for visualizing and characterizing high-dimensional patterns

We introduce a distance (similarity)-based mapping for the visualization of high-dimensional patterns and their relative relationships. The mapping preserves exactly the original distances between points with respect to any two reference patterns in a special two-dimensional coordinate system, the relative distance plane (RDP). As only a single calculation of a distance matrix is required, this method is computationally efficient, an essential requirement for any exploratory data analysis. The data visualization afforded by this representation permits a rapid assessment of class pattern distributions. In particular, we can determine with a simple statistical test whether both training and validation sets of a 2-class, high-dimensional dataset derive from the same class distributions. We can explore any dataset in detail by identifying the subset of reference pairs whose members belong to different classes, cycling through this subset, and for each pair, mapping the remaining patterns. These multiple viewpoints facilitate the identification and confirmation of outliers. We demonstrate the effectiveness of this method on several complex biomedical datasets. Because of its efficiency, effectiveness, and versatility, one may use the RDP representation as an initial, data mining exploration that precedes classification by some classifier. Once final enhancements to the RDP mapping software are completed, we plan to make it freely available to researchers.     

Recognizing the intensity of strength training exercises with wearable sensors

In this paper we propose a system based on a network of wearable accelerometers and an off-the-shelf smartphone to recognize the intensity of stationary activities, such as strength training exercises. The system uses a hierarchical algorithm, consisting of two layers of Support Vector Machines (SVMs), to first recognize the type of exercise being performed, followed by recognition of exercise intensity. The first layer uses a single SVM to recognize the type of the performed exercise. Based on the recognized type a corresponding intensity prediction SVM is selected on the second layer, specializing in intensity prediction for the recognized type of exercise. We evaluate the system for a set of upper-body exercises using different weight loads. Additionally, we compare the most important features for exercise and intensity recognition tasks and investigate how different sliding window combinations, sensor configurations and number of training subjects impact the algorithm performance. We perform all of the experiments for two different types of features to evaluate the feasibility of implementation on resource constrained hardware. The results show the algorithm is able to recognize exercise types with approximately 85% accuracy and 6% intensity prediction error. Furthermore, due to similar performance using different types of features, the algorithm offers potential for implementation on resource constrained hardware.     

Computer-assisted lip diagnosis on traditional Chinese medicine using multi-class support vector machines

ABSTRACT: BACKGROUND: In Traditional Chinese Medicine (TCM), the lip diagnosis is an important diagnostic method which has a long history and is applied widely. The lip color of a person is considered as a symptom to reflect the physical conditions of organs in the body. However, the traditional diagnostic approach is mainly based on observation by doctor's nude eyes, which is non-quantitative and subjective. The non-quantitative approach largely depends on the doctor's experience and influences accurate the diagnosis and treatment in TCM. Developing new quantification methods to identify the exact syndrome based on the lip diagnosis of TCM becomes urgent and important. In this paper, we design a computer-assisted classification model to provide an automatic and quantitative approach for the diagnosis of TCM based on the lip images. METHODS: A computer-assisted classification method is designed and applied for syndrome diagnosis based on the lip images. Our purpose is to classify the lip images into four groups: deep-red, red, purple and pale. The proposed scheme consists of four steps including the lip image preprocessing, image feature extraction, feature selection and classification. The extracted 84 features contain the lip color space component, texture and moment features. Feature subset selection is performed by using SVM-RFE (Support Vector Machine with recursive feature elimination), mRMR (minimum Redundancy Maximum Relevance) and IG (information gain). Classification model is constructed based on the collected lip image features using multi-class SVM and Weighted multi-class SVM (WSVM). In addition, we compare SVM with k-nearest neighbor (kNN) algorithm, Multiple Asymmetric Partial Least Squares Classifier (MAPLSC) and Naive Bayes for the diagnosis performance comparison. All displayed faces image have obtained consent from the participants. RESULTS: A total of 257 lip images are collected for the modeling of lip diagnosis in TCM. The feature selection method SVM-RFE selects 9 important features which are composed of 5 color component features, 3 texture features and 1 moment feature. SVM, MAPLSC, Naive Bayes, kNN showed better classification results based on the 9 selected features than the results obtained from all the 84 features. The total classification accuracy of the five methods is 84%, 81%, 79% and 81%, 77%, respectively. So SVM achieves the best classification accuracy. The classification accuracy of SVM is 81%, 71%, 89% and 86% on Deep-red, Pale Purple, Red and lip image models, respectively. While with the feature selection algorithm mRMR and IG, the total classification accuracy of WSVM achieves the best classification accuracy. Therefore, the results show that the system can achieve best classification accuracy combined with SVM classifiers and SVM-REF feature selection algorithm. CONCLUSIONS: A diagnostic system is proposed, which firstly segments the lip from the original facial image based on the Chan-Vese level set model and Otsu method, then extracts three kinds of features (color space features, Haralick co-occurrence features and Zernike moment features) on the lip image. Meanwhile, SVM-REF is adopted to select the optimal features. Finally, SVM is applied to classify the four classes. Besides, we also compare different feature selection algorithms and classifiers to verify our system. So the developed automatic and quantitative diagnosis system of TCM is effective to distinguish four lip image classes: Deep-red, Purple, Red and Pale. This study puts forward a new method and idea for the quantitative examination on lip diagnosis of TCM, as well as provides a template for objective diagnosis in TCM.     

A Content-Boosted Collaborative Filtering Algorithm for Personalized Training in Interpretation of Radiological Imaging

Devising a method that can select cases based on the performance levels of trainees and the characteristics of cases is essential for developing a personalized training program in radiology education. In this paper, we propose a novel hybrid prediction algorithm called content-boosted collaborative filtering (CBCF) to predict the difficulty level of each case for each trainee. The CBCF utilizes a content-based filtering (CBF) method to enhance existing trainee-case ratings data and then provides final predictions through a collaborative filtering (CF) algorithm. The CBCF algorithm incorporates the advantages of both CBF and CF, while not inheriting the disadvantages of either. The CBCF method is compared with the pure CBF and pure CF approaches using three datasets. The experimental data are then evaluated in terms of the MAE metric. Our experimental results show that the CBCF outperforms the pure CBF and CF methods by 13.33 and 12.17 %, respectively, in terms of prediction precision. This also suggests that the CBCF can be used in the development of personalized training systems in radiology education.     

Supervised pattern recognition for the prediction of contrast-enhancement appearance in brain tumors from multivariate magnetic resonance imaging and spectroscopy

OBJECTIVE: The purpose of this study was to develop a pattern classification algorithm for use in predicting the location of new contrast-enhancement in brain tumor patients using data obtained via multivariate magnetic resonance (MR) imaging from a prior scan. We also explore the use of feature selection or weighting in improving the accuracy of the pattern classifier. METHODS AND MATERIALS: Contrast-enhanced MR images, perfusion images, diffusion images, and proton spectroscopic imaging data were obtained from 26 patients with glioblastoma multiforme brain tumors, divided into a design set and an unseen test set for verification of results. A k-NN algorithm was implemented to classify unknown data based on a set of training data with ground truth derived from post-treatment contrast-enhanced images; the quality of the k-NN results was evaluated using a leave-one-out cross-validation method. A genetic algorithm was implemented to select optimal features and feature weights for the k-NN algorithm. The binary representation of the weights was varied from 1 to 4 bits. Each individual parameter was thresholded as a simple classification technique, and the results compared with the k-NN. RESULTS: The feature selection k-NN was able to achieve a sensitivity of 0.78+/-0.18 and specificity of 0.79+/-0.06 on the holdout test data using only 7 of the 38 original features. Similar results were obtained with non-binary weights, but using a larger number of features. Overfitting was also observed in the higher bit representations. The best single-variable classifier, based on a choline-to-NAA abnormality index computed from spectroscopic data, achieved a sensitivity of 0.79+/-0.20 and specificity of 0.71+/-0.11. The k-NN results had lower variation across patients than the single-variable classifiers. CONCLUSIONS: We have demonstrated that the optimized k-NN rule could be used for quantitative analysis of multivariate images, and be applied to a specific clinical research question. Selecting features was found to be useful in improving the accuracy of feature weighting algorithms and improving the comprehensibility of the results. We believe that in addition to lending insight into parameter relevance, such algorithms may be useful in aiding radiological interpretation of complex multimodality datasets.     

An improved electromagnetism-like mechanism algorithm and its application to the prediction of diabetes mellitus

Recently, the use of artificial intelligence based data mining techniques for massive medical data classification and diagnosis has gained its popularity, whereas the effectiveness and efficiency by feature selection is worthy to further investigate. In this paper, we presents a novel method for feature selection with the use of opposite sign test (OST) as a local search for the electromagnetism-like mechanism (EM) algorithm, denoted as improved electromagnetism-like mechanism (IEM) algorithm. Nearest neighbor algorithm is served as a classifier for the wrapper method. The proposed IEM algorithm is compared with nine popular feature selection and classification methods. Forty-six datasets from the UCI repository and eight gene expression microarray datasets are collected for comprehensive evaluation. Non-parametric statistical tests are conducted to justify the performance of the methods in terms of classification accuracy and Kappa index. The results confirm that the proposed IEM method is superior to the common state-of-art methods. Furthermore, we apply IEM to predict the occurrence of Type 2 diabetes mellitus (DM) after a gestational DM. Our research helps identify the risk factors for this disease; accordingly accurate diagnosis and prognosis can be achieved to reduce the morbidity and mortality rate caused by DM.     

Automatic detection of unstained viable cells in bright field images using a support vector machine with an improved training procedure

Detection of unstained viable cells in bright field images is an inherently difficult task due to the immense variability of cell appearance. Traditionally, it has required human observers. However, in high-throughput robotic systems, an automatic procedure is essential. In this paper, we formulate viable cell detection as a supervised, binary pattern recognition problem and show that a support vector machine (SVM) with an improved training algorithm provides highly effective cell identification. In the case of cell detection, the binary classification problem generates two classes, one of which is much larger than the other. In addition, the total number of samples is extremely large. This combination represents a difficult problem for SVMs. We solved this problem with an iterative training procedure ("Compensatory Iterative Sample Selection", CISS). This procedure, which was systematically studied under various class size ratios and overlap conditions, was found to outperform several commonly used methods, primarily owing to its ability to choose the most representative samples for the decision boundary. Its speed and accuracy are sufficient for use in a practical system.     

Support vector machine for optical diagnosis of cancer

We report the application of a support vector machine (SVM) for the development of diagnostic algorithms for optical diagnosis of cancer. Both linear and nonlinear SVMs have been investigated for this purpose. We develop a methodology that makes use of SVM for both feature extraction and classification jointly by integrating the newly developed recursive feature elimination (RFE) in the framework of SVM. This leads to significantly improved classification results compared to those obtained when an independent feature extractor such as principal component analysis (PCA) is used. The integrated SVM-RFE approach is also found to outperform the classification results yielded by traditional Fisher's linear discriminant (FLD)-based algorithms. All the algorithms are developed using spectral data acquired in a clinical in vivo laser-induced fluorescence (LIF) spectroscopic study conducted on patients being screened for cancer of the oral cavity and normal volunteers. The best sensitivity and specificity values provided by the nonlinear SVM-RFE algorithm over the data sets investigated are 95 and 96% toward cancer for the training set data based on leave-one-out cross validation and 93 and 97% toward cancer for the independent validation set data. When tested on the spectral data of the uninvolved oral cavity sites from the patients it yielded a specificity of 85%.     

基于迁移学习的小样本重症疾病预后模型

针对临床上重症疾病样本数量少容易导致预后模型过拟合、预测误差大、不稳定的问题,本文提出迁移长短时程记忆算法(transLSTM)。该算法基于迁移学习思想,利用疾病间的相关性实现不同疾病预后模型的信息迁移,借助相关疾病的大数据辅助构建小样本目标病种有效模型,提升模型预测性能,降低对目标训练样本量的要求。transLSTM算法先利用相关疾病数据预训练部分模型参数,再用目标训练样本进一步调整整个网络。基于MIMIC-Ⅲ数据库的测试结果显示,相比传统的LSTM分类算法,transLSTM算法的AUROC指标高出0.02~0.07,AUPRC指标超过0.05~0.14,训练迭代次数仅为传统算法的39%~64%。应用于脓毒症疾病的结果显示,仅100个训练样本的transLSTM模型死亡率预测性能与250个训练样本的传统模型相当。在小样本情况下,transLSTM算法预测精度更高、训练速度更快,具有显著优势。它实现了迁移学习在小样本重症疾病预后模型中的应用。