C. Marchesi

Artificial neural networks for automatic ECG analysis

The analysis of ECGs can benefit from the wide availability of computing technology. This paper presents some results achieved by carrying out the classification tasks of equipment integrating the most common features of the ECG analysis: arrhythmia, myocardial ischemia, chronic alterations. Several ANN architectures are implemented, tested, and compared with competing alternatives. The approach, structure, and learning algorithm of ANNs are designed according to the features of each particular classification task. The trade-off between the time consuming training of ANNs and their performance is also explored. Data pre- and post-processing efforts for system performance are critically tested. The crucial role of these efforts for the reduction of input space dimensions, for a more significant description of the input features, and for improving new or ambiguous event processing is also documented. Finally, algorithm assessment is done on data coming from available ECG databases.

Identification of a predominant COPD phenotype in clinical practice

BACKGROUND Chronic obstructive pulmonary disease (COPD) is characterized by airflow limitation caused by small airways increased resistance and/or terminal airspaces emphysematous destruction. Spirometric detection of not fully reversible airflow limitation unifies under the acronym COPD, a spectrum of heterogeneous conditions, whose clinical presentations may be substantially different. In a cross-sectional study we aimed to ascertain whether COPD phenotypes reflecting different mechanisms of airflow limitation could be clinically identified. METHODS Multidimensional scaling was used to visualize as a single point in a two-dimension space the multidimensional variables derived from each of 322 COPD patients (derivation set) by clinical, functional, and chest radiographic evaluation. Cluster analysis assigned then a cluster membership to each patient data point. Finally, using cluster membership as dependent variable and all data acquired as independent variables, we developed multivariate models to prospectively classify another group of 93 COPD patients (validation set) in whom high-resolution computerized tomography (HRCT) density parameters were measured. RESULTS A multivariate model based on nine variables acquired from the derivation set by history (sputum characteristics), physical examination (adventitious sounds, hyperresonance), FEV1/VC, and chest radiography (increased vascular markings, bronchial wall thickening, increased lung volume, reduced lung density) partitioned the validation set into two groups whose clinical, functional, chest radiographic, and HRCT characteristics corresponded to either an airways obstructive or a parenchymal destructive COPD phenotype. CONCLUSION Patients with COPD can be assigned a clinical phenotype reflecting the prevalent mechanism of airflow limitation. The standardized identification of the predominant phenotype may permit to clinically characterize COPD beyond its unifying spirometric definition.

A system for the detection of ischemic episodes in ambulatory ECG

Describes a system for the automated detection of ischemic changes in double channel ambulatory EGG. The high performance QRS detector, previously designed and tested by the authors, was applied. A subset of ECG records of the European ST-T Database was used for development, whereas the full set of 90 records and selected 24-hour ECG recordings were used for evaluation. Noisy ECG segments were rejected by the QRS detector or processed by the use of median filters. ST segment deviations, measured beat by beat on both channels, were used for detection of ischemic episodes. The time series of parameters were smoothed by nonlinear filters. Episode detection is based on the analysis of the 2-D path of the two ST deviations. Results of performance evaluation on the European ST-T Database are reported.

Explorative data analysis techniques and unsupervised clustering methods to support clinical assessment of Chronic Obstructive Pulmonary Disease (COPD) phenotypes

Chronic Obstructive Pulmonary Disease (COPD) is the fourth leading cause of death worldwide and represents one of the major causes of chronic morbidity. Cigarette smoking is the most important risk factor for COPD. In these patients, the airflow limitation is caused by a mixture of small airways disease and parenchyma destruction, the relative contribution of which varies from person to person. The twofold nature of the pathology has been studied in the past and according to some authors each patient should be classified as presenting a predominantly bronchial or emphysematous phenotype. In this study we applied various explorative analysis techniques (PCA, MCA, MDS) and recent unsupervised clustering methods (KHM) to study a large dataset, acquired from 415 COPD patients, to assess the presence of hidden structures in data corresponding to the different COPD phenotypes observed in clinical practice. In order to validate our methods, we compared the results obtained from a training set of 415 patients with lung density data acquired in a test set of 93 patients who underwent HRCT (High Resolution Computerized Tomography).

Development of an electronic medical record for patient care in cardiology

An electronic medical record system has been designed and developed with the aim of supporting patient care in the department of cardiology. A clinical information system was implemented to integrate the different heterogeneous sources of patient information. All the clinically relevant patient data were gathered in a timely way and stored in the clinical database. A viewer/editor of patient medical records was designed, addressing friendliness, flexibility and communication. The basic function for consultation is the time-oriented (weekly or daily) representation of patient parameters, care events and examinations. World Wide Web technology has been applied to implement the medical record. The system is currently under clinical evaluation.

Continuous monitoring and detection of ST-T changes in ischemic patients

The authors developed a complete two channel ST episode detection system for long term ECG records. To improve the system sensitivity, a high performance QRS detector was implemented and some noise criteria were applied, to reject too noisy measure values (sens: 97.51% PPA: 99.96%). A three layer feedforward Artificial Neural Network (ANN), trained by backpropagation algorithm, was introduced. It processed the inputs (ST amplitude and ST slope, both in absolute value) in a nonlinear way so that the ST episodes became more easily recognizable from ANN output and the system sensitivity resulted improved (sens: 85% PPA; 88% with vs. sens: 78% PPA: 90% without ANN). The training set was built with 3 out of the 50 records of the European Society of Cardiology ST-T Database. The remaining records were used for system evaluation.<<ETX>>

The Long-Term ST database: a research resource for algorithm development and physiologic studies of transient myocardial ischemia

Presents the Long Term ST Database, a collection of eighty 24-hour two and three lead ECG records from ambulatory subjects with transient ST segment abnormalities. The database provides a comprehensive standard research resource for quantitatively assessing the performance of automated detectors of transient ischemia, and for supporting basic research into the mechanisms and dynamics of transient ischemia. Records of the database contain annotated significant transient ischemic ST episodes, non-ischemic ST episodes caused by heart rate related changes, non-ischemic ST events due to axis shifts or QRS conduction changes, and individual QRS and rhythm annotations, all made by human experts.

A multichannel template based data compression algorithm

Data compression methods for electrocardiogram (EGG) signals play an important role in computer processing and analysis of ECG. Ambulatory monitoring of ECG (AECG) is the area of application which mostly benefits from data compression techniques. AECG is usually done using the conventional Holter monitor, which consists of a 24 hour, 2-channel, analog, cassette recording of the EGG. Modern Holter monitors with digital memory cards are now expected to improve fidelity of recording and decrease system size. The cost of high capacity integrated memory cards makes it convenient to store ECG samples in compact form. A powerful template matching algorithm (ABS) has been recently introduced. The authors present a method which improves ABS algorithm, in order to achieve a better compression ratio as well as faster processing.<<ETX>>

Supervised and unsupervised learning for diagnostic ECG classification

A hybrid system with RBF pre-processing, a system with supervised learning, is compared with some Kohonen self-organizing maps in a subtle ECG classification task. Based on ECG measures, they are supposed to detect normal condition, presence of infarction and of hypertrophy, and at the same time to sub-classify those pathologies. During the evaluation process the hybrid system produces better results. In terms of average sensitivity and specificity (83% vs. 62% of sensitivity and 84% vs. 92% of specificity), but Kohonen maps allow a detailed description of the similarities among input data. An integration of the two techniques should improve the final results.

Fuzzy pattern classification and the connectionist approach

Several hybrid architectures combining fuzzy pattern classification and the connectionist approach will be developed and tested for the particular problem of diagnostic classification in computerized electrocardiography. The first level of fuzzy description of the input parameters is performed by a layer of Radial Basis Functions, and this step can be seen as a level of data abstraction. A subsequent classical NN processes these fuzzy descriptions. Several experiments have been performed on the components of the resulting architecture in order to point out their influence on the overall performance in the diagnostic classification task. A large validated database has been used for the validation of the proposed hybrid architecture.