Classification Algorithm of Parkinson’s Disease Based on Convolutional Sparse Transfer Learning and Sample/Feature Parallel Selection
-
摘要: 基于语音数据分析的帕金森病(PD)诊断存在样本量小、训练与测试数据分布差异明显的问题。为了解决这些问题,需要从降维和样本扩充两个方面同时进行。因此,该文提出结合加噪加权卷积稀疏迁移学习和样本特征并行优选的PD分类算法。该算法可从源域的公共语音库中学习有利于表达PD语音特征的有效结构信息,同时完成降维和样本间接扩充。样本特征并行优选考虑到了样本和语音特征间的关系,从而有助于获取高质量的特征。首先,对公共语音库进行特征提取构造公共特征库;然后,以公共特征库对PD目标域的训练数据集及测试数据集进行稀疏编码,这里分别采用传统稀疏编码(SC)与卷积稀疏编码(CSC)两种稀疏编码方法;接着,对编码后的语音样本段和特征数据进行同时优选;最后,采用支撑向量机(SVM)进行分类。实验结果表明,该算法针对受试者的分类准确率最高值达到了95.0%,均值达到了86.0%,较相关被比较算法有较大提高。此外,研究还发现,相较于传统稀疏编码方法,卷积稀疏编码更有利于提取PD语音数据的高层特征;同样,迁移学习也有利于提高该算法性能。
-
关键词:
- 迁移学习 /
- 帕金森病 /
- 稀疏编码 /
- 卷积稀疏编码 /
- 语音样本特征并行优选
Abstract: To solve the problems that there are few labeled data in speech data for diagnosis of Parkinson’s Disease (PD), and the distributed condition of the training and the test data is different, the two aspects of dimension reduction and sample augment are considered. A novel transfer learning algorithm is proposed based on noise weighting sparse coding combined with speech sample / feature parallel selection. The algorithm can learn the structural information from the source domain and express the effective PD features, and achieves dimension reduction and sample augment simultaneously. Considering the relationship between the samples and features, the higher quality features can be extracted. Firstly, the features are extracted from the public data set and the feature data set is constructed as source domain. Then the training data and test data of the target domain are sparsely represented based on source domain. Spares representing includs traditional Sparse Coding(SC) and Convolutional Sparse Coding(CSC); Next, the sparse representing data are screened according to sample feature selection simultaneously, so as to improve the accuracy of the PD classification; Finally, the Support Vector Machine(SVM) classifier is adopted. Experiments show that it achieves the highest classification accuracy of 95.0% and the average classification accuracy of 86.0%, and obtains obvious improvement according to the subjects, compared with the relevant algorithms. Besides, compared with sparse coding, convolutional sparse coding can be beneficial to extracting high level features from PD data set; moreover, it is proved that transfer learning is effective. -
表 1 基于语音卷积稀疏迁移学习和并行优选的PD分类算法
输入:公共数据集${\text{S}}$,目标领域数据集${\text{A}}$,样本总数H,每个样
本的特征数N,受试者数M输出:分类准确率,灵敏度,特异度 步骤: (1)对公共数据集${\text{S}}$叠加不同类型不同信噪比的噪声扩展其为
集合${\text{S}}'$;(2)根据式(1),对集合${\text{S}}'$的语音样本提取特征构造特征库即
源领域数据集${\text{Y}}$;(3)根据2.2.2节及2.2.3节,提取特征库${\text{Y}}$的字典(稀疏编
码),卷积核(卷积稀疏编码);(4)根据1.2.4节,计算目标领域数据集${\text{A}}$的稀疏表达系数矩
阵${\text{E}}$(稀疏编码)或特征图矩阵${\text{E}}$(卷积稀疏编码);(5)根据语音样本特征并行优选算法,将特征矩阵${\text{E}}$进行特征
扩展为${\text{G}}$并归一化为${\text{G}}'$,并进行样本特征同时优选得矩阵${\text{P}}$;(6)基于SVM分类器进行受试者留一法(LOSO)分类计算。 表 2 各种算法分类结果对比(%)
分类算法 基于受试者的留一法 准确率 灵敏度 特异度 SVM(线性核函数) 平均 65.0 65.0 65.0 最好 65.0 65.0 65.0 SVM(径向基核函数) 平均 67.5 80.0 55.0 最好 67.5 80.0 55.0 文献[6]算法 平均 52.0 55.0 49.0 最好 85.0 85.0 90.0 DBN算法 平均 54.6 52.4 56.8 最好 57.0 56.0 58.0 CNN算法 平均 60.0 63.0 57.0 最好 65.0 61.0 69.0 autoencoder+SVM(TL) 平均 72.5 75.0 70.0 最好 72.5 75.0 70.0 autoencoder+SVM 平均 67.5 65.0 70.0 最好 67.5 65.0 70.0 DBN+SVM(TL) 平均 55.5 60.0 51.0 最好 60.0 65.0 55.0 DBN+SVM 平均 50.5 53.0 48.0 最好 57.5 65.0 50.0 PD_SC&S2 平均 68.5 69.5 67.5 最好 90.0 85.0 95.0 PD_SC&S2_TL 平均 81.0 79.5 82.5 平均 92.5 95.0 90.0 PD_CSC&S2 平均 70.0 73.0 67.0 最好 75.0 74.0 76.0 PD_CSC&S2_TL 平均 86.0 91.0 81.0 最好 95.0 100.0 90.0 -
BENGE J F, ROBERTS R L, KEKECS Z, et al. Brief report: knowledge of, interest in, and willingness to try behavioral interventions in individuals with Parkinson's disease[J]. Advances in Mind-Body Medicine, 2018, 32(1): 8–12. PLOUVIER A O A, OLDE HARTMAN T C, VAN LITSENBURG A, et al. Being in control of Parkinson's disease: a qualitative study of community-dwelling patients' coping with changes in care[J]. European Journal of General Practice, 2018, 24(1): 138–145. doi: 10.1080/13814788.2018.1447561 CHIU Y F and FORREST K. The impact of lexical characteristics and noise on intelligibility of Parkinsonian speech[J]. Journal of Speech, Language, and Hearing Research, 2018, 61(4): 837–846. doi: 10.1044/2017_JSLHR-S-17-0205 LITTLE M A, MCSHARRY P E, HUNTER E J, et al. Suitability of dysphonia measurements for telemonitoring of Parkinson's disease[J]. IEEE Transactions on Biomedical Engineering, 2009, 56(4): 1015–1022. doi: 10.1109/TBME.2008.2005954 TSANAS A, LITTLE M A, MCSHARRY P E, et al. Novel speech signal processing algorithms for high-accuracy classification of Parkinson's disease[J]. IEEE Transactions on Biomedical Engineering, 2012, 59(5): 1264–1271. doi: 10.1109/TBME.2012.2183367 SAKAR B E, ISENKUL M E, SAKAR C O, et al. Collection and analysis of a Parkinson speech dataset with multiple types of sound recordings[J]. IEEE Journal of Biomedical and Health Informatics, 2013, 17(4): 828–834. doi: 10.1109/JBHI.2013.2245674 GODINO-LLORENTE J I, SHATTUCK-HUFNAGEL S, CHOI J Y, et al. Towards the identification of idiopathic Parkinson's disease from the speech. New articulatory kinetic biomarkers[J]. PLOS One, 2017, 12(12): e0189583. doi: 10.1371/journal.pone.0189583 KAYA M and BILGE H Ş. Classification of Parkinson speech data by metric learning[C]. Proceedings of the 2017 International Artificial Intelligence and Data Processing Symposium, Malatya, Turkey, 2017: 1–5. JI Wei and LI Yun. Stable dysphonia measures selection for Parkinson speech rehabilitation via diversity regularized ensemble[C]. International Conference on Acoustics, Speech and Signal Processing, Shanghai, China, 2016: 2264–2268. doi: 10.1109/ICASSP.2016.7472080. KIM J, NASIR M, GUPTA R, et al. Automatic estimation of Parkinson's disease severity from diverse speech tasks[C]. Proceedings of the 16th Annual Conference of the International Speech Communication Association, Dresden, Germany, 2015: 914–918. SHAHBAKHTI M, TAHERIFAR D, and SOROURI A. Linear and non-linear speech features for detection of Parkinson's disease[C]. Proceedings of the 6th 2013 Biomedical Engineering International Conference, Amphur Muang, Thailand, 2013: 1–3. doi: 10.1109/BMEiCon.2013.6687667. DUBEY H, GOLDBERG J C, ABTAHI M, et al. EchoWear: Smartwatch technology for voice and speech treatments of patients with Parkinson's disease[C]. Proceedings of the Conference on Wireless Health, Bethesda, USA, 2015: Article No.15. doi: 10.1145/2811780.2811957. ARIAS-VERGARA T, VASQUEZ-CORREA J C, OROZCO-ARROYAVE J R, et al. Parkinson's disease progression assessment from speech using GMM-UBM[C]. Conference of the International Speech Communication Association, San Francisco, USA, 2016: 1933–1937. doi: 10.21437/Interspeech.2016-1122. GEMAN O. Data processing for Parkinson's disease: Tremor, speech and gait signal analysis[C]. Proceedings of the 2011 E-Health and Bioengineering Conference, Iasi, Romania, 2011: 1–4. 李勇明, 杨刘洋, 刘玉川, 等. 基于语音样本重复剪辑和随机森林的帕金森病诊断算法研究[J]. 生物医学工程学杂志, 2016, 33(6): 1053–1059.LI Yongming, YANG Liuyang, LIU Yuchuan, et al. Research on diagnosis algorithm of Parkinson's disease based on speech sample multi-edit and random forest[J]. Journal of Biomedical Engineering, 2016, 33(6): 1053–1059. 张小恒, 王力锐, 曹垚, 等. 混合语音段特征双边式优选算法用于帕金森病分类研究[J]. 生物医学工程学杂志, 2017, 34(6): 942–948. doi: 10.7507/1001-5515.201704061ZHANG Xiaoheng, WANG Lirui, CAO Yao, et al. Combining speech sample and feature bilateral selection algorithm for classification of Parkinson's disease[J]. Journal of Biomedical Engineering, 2017, 34(6): 942–948. doi: 10.7507/1001-5515.201704061 SHARMA P, ABROL V, DILEEP A D, et al. Sparse coding based features for speech units classification[J]. Computer Speech & Language, 2018, 47: 333–350. doi: 10.1016/j.csl.2017.08.004 ZHOU Haotian, ZHUANG Yin, CHEN Liang, et al. Ship Detection in Optical Satellite Images Based on Sparse Representation[M]. Singapore: Springer, 2018: 164–171. doi: 10.1007/978-981-10-7521-6_20. LI Jinming. Sparse representation based single image super-resolution with low-rank constraint and nonlocal self-similarity[J]. Multimedia Tools and Applications, 2018, 77(2): 1693–1714. doi: 10.1007/s11042-017-4399-1 LEE H, BATTLE A, RAINA R, et al. Efficient sparse coding algorithms[C]. Proceedings of the 19th International Conference on Neural Information Processing Systems, Canada, 2006: 801–808. CHANG Hang, HAN Ju, ZHONG Cheng, et al. Unsupervised transfer learning via multi-scale convolutional sparse coding for biomedical applications[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(5): 1182–1194. doi: 10.1109/TPAMI.2017.2656884 ZHANG He and PATEL V M. Convolutional sparse and low-rank coding-based image decomposition[J]. IEEE Transactions on Image Processing, 2018, 27(5): 2121–2133. doi: 10.1109/TIP.2017.2786469 HU Xuemei, HEIDE F, DAI Qionghai, et al. Convolutional sparse coding for RGB+NIR imaging[J]. IEEE Transactions on Image Processing, 2018, 27(4): 1611–1625. doi: 10.1109/TIP.2017.2781303 WOHLBERG B. Efficient algorithms for convolutional sparse representations[J]. IEEE Transactions on Image Processing, 2016, 25(1): 301–315. doi: 10.1109/TIP.2015.2495260 PERKINS S and THEILER J. Online feature selection using grafting[C]. Proceedings of the Twentieth International Conference on Machine Learning, Washington DC, USA, 2003: 592–599. BOYD S, PARIKH N, CHU E, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers[J]. Foundations and Trends ® in Machine Learning, 2011, 3(1): 1–122. doi: 10.1561/2200000016 EMADI M, MIANDJI E, and UNGER J. OMP-based DOA estimation performance analysis[J]. Digital Signal Processing, 2018, 79: 57–65. doi: 10.1016/j.dsp.2018.04.006 ŠOREL M and ŠROUBEK F. Fast convolutional sparse coding using matrix inversion lemma[J]. Digital Signal Processing, 2016, 55: 44–51. doi: 10.1016/j.dsp.2016.04.012