Analysis Of Medical Time Series Based On Deep Learning

In recent years,the popular research on deep learning and its rapid developments have brought a huge revolution in medical big data field.Due to the complexity of medical data,applying deep learning to medical field is still full of challenges,so as to further research.As an important part of medical data,time series data can be divided into two categories according to the characteristics of the data.One is non-continuous multivariate clinical time series which is usually sparse.The information and features in time series may be hidden or lost resulting from the equipment failures and manual recording irregularities.The other is continuous waveform time series,which contains enormous information in a short period.However,the presence of interference signals may cause diagnostic errors for doctors.Hence mining useful information from medical data remain to be fully solved.To explore of the sparse nature of multivariate clinical time series and the time-frequency nature of waveform physiological signal data,the following research has been conducted on the optimal application of deep learning models in medical time senes:·A novel missing values handling algorithm is proposed,in which,missing values are interpolated by truncated nearest-neighbor values meanwhile the mask of missing values is also treated as inputs.The result illustrates that the prediction performance of deep learning network is improved when dealing with sparse multivariate clinical time series.The algorithm is implemented on the EHR data extracted from the MIMIC-? clinical data set then the processed data is trained by a deep recurrent neural network.As comparison,a conventional machine learning method,logistic regression is adopted as a benchmark model.With the proposed approach,one can not only omit the step the feature extraction,but also obtain better performance,which is of vital meaning for early warning of critical patients in practice.·By combining empirical mode decomposition(EMD)and deep learning,a new waveform physiological signal analyzing frame is proposed to assist medical diagnosis.It is a highlight that the inputs of neural network are feed by the decomposition result of EMD.Because EMD has the function of automatically extracting noise and baseline drift,the time-frequency information of waveform physiological signals can be fully mined without manual feature extraction.With the proposed frame,the length of ECG signal is trivial;waveform detection,location and segmentation are unnecessary.The effect is verified on the MIT-BIT arrhythmia dataset.It is illustrated that the EMD auxiliary layer has a significant gain for improving the classification performance of simple networks.EMD is still available in complex network.Moreover,compared with existing deep learning techniques for heartbeat classification,it shows that the model is superior in comprehensive evaluation indicators.