Machine learning aided classification of tremor in multiple sclerosis.

BACKGROUND: Tremors are frequent and disabling in people with multiple sclerosis (MS). Characteristic tremor frequencies in MS have a broad distribution range (1-10 Hz), which confounds the diagnostic from other forms of tremors. In this study, we propose a classification method for distinguishing MS tremors from other forms of cerebellar tremors. METHODS: Electromyogram (EMG), accelerometer and clinical data were obtained from a total of 120 [40 MS, 41 essential tremor (ET) and 39 Parkinson's disease (PD)] subjects. The proposed method - Soft Decision Wavelet Decomposition (SDWD) - was used to compute power spectral densities and receiver operating characteristic (ROC) analysis was performed for the automatic classification of the tremors. Association between the spectral features and clinical features (FTM - Fahn-Tolosa-Marin scale, UPDRS - Unified Parkinson's Disease Rating Scale), was assessed using a support vector regression (SVR) model. FINDINGS: Our developed analytical framework achieved an accuracy of up to 91.67% using accelerometer data and up to 91.60% using EMG signals for the differentiation of MS tremors and the tremors from ET and PD. In addition, SVR further revealed strong significant correlations between the selected discriminators and the clinical scores. INTERPRETATION: The proposed method, with high classification accuracy and strong correlations of these features to clinical outcomes, has clearly demonstrated the potential to complement the existing tremor-diagnostic approach in MS patients. FUNDING: This work was supported by the German Research Foundation (DFG): SFB-TR-128 (to SG, MM), MU 4354/1-1(to MM) and the Boehringer Ingelheim Fonds BIF-03 (to SG, MM).

Discrimination between parkinsonian tremor and essential tremor using artificial neural network with hybrid features.

BACKGROUND: Essential tremor (ET) and the tremor in Parkinson's disease (PD) are the two most common pathological tremors with a certain overlap in the clinical presentation. OBJECTIVE: The main purpose of this work is to use an artificial neural network to select the best features and to discriminate between the two types of tremors. The features used are of hybrid type obtained from two different algorithms: the statistical signal characterization (SSC) of the signal describing its morphology, and the soft-decision wavelet-decomposition (SDWD) features extracted from the accelerometer and surface EMG signals. METHODS: The SSC method is used to obtain morphology-based features of the spectrum of the accelerometer and two surface EMG signals. The SDWD technique is used in this work to obtain the approximate spectral representation of both accelerometer and the two surface EMG signals. Two sets of data (training and test) are used in this paper. The training set consists of 21 ET subjects and 19 PD subjects, while the test set consists of 20 ET and 20 PD subjects. A neural network of the type feed forward back propagation has been used to combine best SSC features and best SDWD features of the accelerometer and EMG signals. RESULTS: Efficiency result of 92.5% was obtained using best hybrid features. CONCLUSIONS: The artificial neural network has been used successfully to combine two types of features in an automatic discrimination system between PD and ET.

Investigation of heart rate variability of patients undergoing coronary artery bypass grafting using the statistical signal characterization method.

BACKGROUND: Autonomic function can be estimated non-invasively using heart rate variability (HRV). HRV of patients undergoing coronary artery bypass grafting (CABG) is investigated in time-domain and frequency-domain before and after CABG to study the effect of operation on the status of patients. OBJECTIVE: The main purpose of this work is to evaluate the effect of CABG surgery on patients with ischemic heart disease (IHD) before operation, and to monitor the status of patients on day 6 and day 30 after the CABG operation. METHODS: The statistical signal characterization (SSC) technique is used in this work in order to derive different morphology-based parameters to indirectly describe time-domain and frequency-domain HRV parameters in 24 patients undergoing CABG operation, before the operation (Group 1: G1), 6 days after operation (Group 2: G2) and 30 days after operation (Group 3: G3). The data is obtained from the Sultan Qaboos University Hospital in Oman. RESULTS: The SSC parameters Mean(mt) and Mean(dt) are reduced in all 24 patients and in 23 out of 24 patients in G2 compared to G1 (6-days after operation compared with before operation), respectively. Comparing G3 to G1 the reduction in Mean(mt) and Mean(dt) is noted in 18 of the 24 patients. CONCLUSIONS: The parameters Mean(mt) and Mean(dt) are successful parameters to follow the HRV for patients undergoing CABG surgery. A relation between those SSC parameters and the HRV time-domain and frequency-domain parameters is investigated in this paper to understand the physiological behavior of the patients.

A new simple efficient classification technique for severity of sleep apnea with mathematical model and interpretation.

BACKGROUND: Obstructive Sleep Apnea (OSA) is the cessation of breathing during sleep due to the collapse of the upper airway. Polysomnographic recording is a conventional method for detection of OSA. Although it provides reliable results, it is expensive and cumbersome. Thus, an advanced non-invasive heart rate variability (HRV) signal processing technique other than the standard spectral analysis, which also has efficiency limitations, is needed for identification of OSA and classification of apnea levels. OBJECTIVE: The main purpose of this work was to predict the severity of sleep apnea using an efficient method based on the combination of time-domain and frequency-domain analysis of the HRV to classify sleep apnea into three different levels (mild, moderate, and severe) according to its severity and to distinguish them from normal subjects. METHODS: The statistical signal characterization of the FFT-based spectrum of the RRI data is used in this work in order to rank patients to full polysomnography. Data of 20 normal subjects, 20 patients with mild apnea, 20 patients with moderate apnea and 20 patients with severe apnea were used in this study. RESULTS: Accuracy result of 100% was obtained between severe and normal subjects, 100% between mild and normal subjects, and 100% between apnea (mild, moderate, severe) and normal subjects. This perfect accuracy is obtained using the parameter mean (mt). The physiological interpretation of the SSC parameters has been derived using a mathematical model system. CONCLUSIONS: An efficient method for screening of sleep apnea with 100% efficiency in classification of sleep apnea levels, is investigated in this work.

Cardio-autonomic functions and sleep indices before and after coronary artery bypass surgery.

BACKGROUND: Earlier studies showed a short-term impairment of cardio-autonomic functions following coronary artery bypass grafting (CABG). There is a lack of consistency in the time of recovery from this impairment. Studies have attributed the post-CABG atrial fibrillation to preexisting obstructive sleep apnea (OSA) without an objective sleep assessment. The aim of this study was to evaluate the effect of CABG on cardio-autonomic and hemodynamic functions and on OSA indices in patients with ischemic heart disease (IHD). METHODS: Cardio-autonomic function using heart rate variability indices, hemodynamic parameters, and sleep studies were performed in 26 patients with stable IHD before, on day-6, and day-30 post-CABG surgery. RESULTS: The high-frequency powers of normalized R-R intervals significantly (P = 0.002) increased from the preoperative value of 46.09 to 66.52 on day-6 and remained unchanged on day-30 postsurgery. In contrary, the low-frequency powers of normalized R-R interval decreased from 53.91 to 33.48 during the same period (P = 0.002) and remained unchanged on day 30 postsurgery. Baroreceptor sensitivity, obstructive and central apnea indices, desaturation index, and lowest O2 saturation were not significantly different between preoperative, day-6, and day-30 postsurgery. CONCLUSION: Our study revealed that recovery of autonomic functions following CABG occurs as early as 30 days of postsurgery. CABG does not seem to have short-term effects on sleep study indices. However, long-term effects need further evaluation.

A neural network approach for feature extraction and discrimination between Parkinsonian tremor and essential tremor.

BACKGROUND: Essential tremor (ET) and the tremor in Parkinson's disease (PD) are the two most common pathological tremor with a certain overlap in the clinical presentation. OBJECTIVE: The main purpose of this work is to use an artificial neural network to select the best features and to discriminate between the two types of tremors using spectral analysis of tremor time-series recorded by accelerometry and surface EMG signals. METHODS: The Soft-Decision wavelet-based technique is to be used in this work in order to obtain a 16 bands approximate spectral representation of both accelerometer and two EMG signals of two sets of data (training and test). The training set consists of 21 ET subjects and 19 PD subjects while the test set consists of 20 ET and 20 PD subjects. The data has been recorded for diagnostic purposes in the Department of Neurology of the University of Kiel, Germany. A neural network of the type feed forward back propagation has been used to find the frequency bands associated with the different signals that yield better discrimination efficiency on training data. The same designed network is used to discriminate the test set. RESULTS: Efficiency result of 87.5% was obtained using two different bands from each of the three signals under test. CONCLUSIONS: The artificial neural network has been used successfully in both feature extraction and in pattern matching tasks in a complete classification system.

A new diagnostic test to distinguish tremulous Parkinson's disease from advanced essential tremor.

BACKGROUND: Clinical distinction between advanced essential tremor and tremulous Parkinson's disease can be difficult. METHODS: In selected power spectra of accelerometric postural tremor recordings on the more affected side of 41 patients with essential tremor and 39 patients with tremulous Parkinson's disease being indistinguishable by tremor frequency, peak power or number of harmonic peaks, waveform asymmetry (autocorrelation decay), and mean peak power of all harmonic peaks were computed. Cutoff for essential tremor-Parkinson's disease distinction was determined by receiver operating characteristics. Diagnostic yield was tested in 12 clinically unclear patients with monosymptomatic tremor, subsequently definitively diagnosed with essential tremor (n = 2) or Parkinson's disease (n = 10) by 123-I FP-CIT-single-photon emission computed tomography, fluorodopa-positron emission tomography, or clinical course. RESULTS: By autocorrelation decay 64%, by mean harmonic peak power 94% (Parkinson's disease > essential tremor) of patients with a definite clinical diagnosis, and 11 of 12 clinically unclear patients were classified correctly. CONCLUSIONS: Mean harmonic power is a useful measure to separate clinically difficult cases of advanced essential tremor from tremulous Parkinson's disease.

Identification of patients with congestive heart failure using different neural networks approaches.

A new technique for identification of patients with congestive heart failure (CHF) from normal controls is investigated in this paper using spectral analysis and neural networks. The identification system consists of two parts: feature extraction part and classification part. The feature extraction part uses the method of approximate spectral density estimation of R-R-Intervals (RRI) data by implementing the soft decision sub-band decomposition technique. In the classification part, two different methods of machine learning approaches with neural networks are implemented and compared in their performances. Those approaches are: supervised neural network (back-propagation) and unsupervised neural network (Kohonen self organizing maps). The data used in this work is obtained from Massachusetts Institute of Technology (MIT) databases. A data set of 17 CHF and 53 normal subjects is used as original training data set, while another set of 12 CHF and 12 normal subjects is used as original test data set. The classification features are the spectral density of 6 different regions covering the whole spectrum of the RRI data obtained by 32-bands soft decision algorithm. A larger training data set, which is obtained by simulating 1000 CHF and 1000 normal subjects according to the spectral features obtained from the original training data, is used to train the neural network. The neural network is used then to test another simulated data set of the same size of the training date set (simulated according to the spectral features obtained from the original test data set). The accuracy of the classification is found to be about 83.65% and 91.43% with supervised neural networks and unsupervised neural networks respectively.

Statistical signal characterization for congestive heart failure patient's classification.

This paper aims at investigating a new technique of time-domain analysis of heart variability (R-R interval (RRI)) for the screening of patients with Congestive Heart Failure (CHF). This method is based on the Statistical Signal Characterization (SSC) of the analytical signal that is generated using Hilbert transformation of the RRI data. The four SSC parameters are: amplitude mean, period mean, amplitude deviation and period deviation. These parameters and their maximum and minimum values are determined over sliding segments of 300-samples, 32-samples and 16-samples for both the instantaneous amplitudes and the instantaneous frequencies derived from the analytical signal of the RRI data. Data used in this work are drawn from MIT database. Threshold values used in the identification of CHF patients from normal records are selected using the Receiver Operating Characteristics (ROC) curves on trial data. This new technique correctly classifies 31/33 of trial data and 65/70 of test data.

A soft decision algorithm for obstructive sleep apnea patient classification based on fast estimation of wavelet entropy of RRI data.

A soft decision algorithm for Obstructive Sleep Apnea (OSA) patient classification using R-R interval (RRI) data is investigated. This algorithm is based on fast and approximate estimation of the entropy of the wavelet-decomposed bands of the RRI data. The classification is done on the whole record as OSA patient or non-patient (normal). The ratio of the estimated entropy of the low-frequency (LF) band to that of the very-low frequency (VLF) band is used as a classification factor. RRI data used in this work are drawn from MIT database. The MIT trial records are used to set the threshold value of the classification factor using the Receiver Operating Characteristics (ROC). This threshold value is used then to classify the MIT challenge (test) records to obtain the efficiency of classification. The new algorithm classifies correctly 30/30 of MIT-test data using different wavelet filters. Comparison of the results of different wavelet filters is done in terms of complexity and distance parameters. The method is also compared with other two techniques using wavelets in their analysis. The consistency of the results is examined using the leave-one-out technique.

Screening of obstructive sleep apnea based on statistical signal characterization of Hilbert transform of RRI data.

A new technique of time-domain analysis for screening of Obstructive Sleep Apnea (OSA) using R-R interval (RRI) data is investigated. This method is based on the Statistical Signal Characterization (SSC) of the analytical signal that is generated using Hilbert transformation of the RRI data. The four SSC parameters: amplitude mean, period mean, amplitude deviation and period deviation, and their maximum and minimum values are found over a 5-minutes sliding window for both the instantaneous amplitudes and the instantaneous frequencies derived from the analytical signal of the RRI data. Data used in this work are drawn from both MIT database as well as from the Sleep Laboratory at Sultan Qaboos University (SQU) hospital. Threshold values used in the identification of OSA from normal subjects are selected using the Receiver Operating Characteristics (ROC) curves. The new technique classifies correctly 29/30 of MIT Trial data, 27/30 of MIT challenge data, and 30/30 of SQU data.