| 研究課題/領域番号 |
20K12080
|
|---|
| 研究機関 | 会津大学 |
|---|
研究代表者 |
MARKOV K 会津大学, コンピュータ理工学部, 教授 (80394998)
|
|---|
| 研究分担者 |
松井 知子 統計数理研究所, モデリング研究系, 教授 (10370090)
齋藤 純平 福島県立医科大学, 医学部, 講師 (50332929)
|
|---|
| 研究期間 (年度) |
2020-04-01 – 2023-03-31
|
| キーワード | cough recognition / cough monitoring |
|---|
| 研究実績の概要 |
Persistent cough is a symptom common to a number of respiratory disorders, However, reliable monitoring of cough frequency and cough severity over an extended period of time can be a challenge. In this study, we develop a baseline cough classification detection systems using classical machine learning method such as Hidden Markov Model with Gaussian mixture state probability distribution. We also experimented with some deep learning models such as Feed-forward neural networks (FFNN) and Recurrent neural networks (RNN) and evaluated and compared their performance.
All models in this study were evaluated on data collected from 11 real patient in Fukushima Medical University who suffered various respiratory diseases. In order to record the real cough sound, the patients were encouraged to perform their usual daily routines. Each patient's audio was recorded continually in about 24 hours including daytime sounds and sleeping sounds. In total, there were about 15000 manually and 70000 automatically labeled segments.
All models were trained on data from 8 patients totaling more than 6000 segments and were tested on about 900 segments. As feature vectors we used 39 dimensional MFCC coefficients. The frame shift length is 10 ms, and frame shift is 25 ms. We performed cough classification and cough detection tasks and used the F1 measure as a evaluation metric. For the cough classification we obtained the following results: HMM-GMM 93.4%, HMM-DNN 93.5%, HMN-RNN 94.6% and RNN 97.1%. For the cough detection, the results are: HMM-GMM 75.9% and HMM-DNN 77.6%.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
In cough classification experiments, F-measure was higher in DNN-HMM and RNN-
HMM than in GMM-HMM. The advantage of the hybrid system with HMM using deep
learning is that it can learn time series and learn features that do not follow Gaussian distribution. In cough detection, it can be said that a hybrid system with HMM using deep learning is effective. RNN showed the best F-measure value. This is limited by the ability of the HMM to capture long-term dependencies, and during HMM training, past remote events have much less effect on state variables than recent events. Therefore, it is considered that the results of the model are quite independent of past remote inputs and outputs. On the other hand, RNN thinks that long-term dependencies could be appropriately modeled by LSTM that controls internal memory and bidirectional structure that can handle time series from the future. It is assumed that RNN is most suitable for cough classification task.In cough detection experiments, DNN-HMM had higher F-measure than GMM-HMM. It is considered that DNN can handle more features than GMM, as in the cough discrimination experiment. Also, since the existence range of the likelihood is different between the GMM-HMM and the DNN-HMM, it is necessary to find the optimal recognition parameters using Julius. In the DNN-HMM, it is considered that the detection accuracy may be further improved by obtaining the optimal parameters and optimizing the structure of the DNN.
|
| 今後の研究の推進方策 |
We obtained good system performance (>90%) for the cough classification task. However, the results for cough detection are still rather low (~76%) and the feature work will be focused on improving the cough detection performance. We plan to study several other state-of-the-art deep learning approaches and methods which were found to work for other sound detection tasks. Those methods include Bidirectional LSTM networks, encode-decoder structures such as transformers and their derivatives. In addition, we plan to use another signal representation - spectrogram in combination with Convolutional neural networks (CNN). They have shown great potential in speech recognition and audio analysis tasks where they are used as nonlinear, smart feature extractors. The best approach, in our view will be a combination of CNN based feature extraction with encoder-decoder based cough detection models.
On the other hand, the currently available data from 11 patients may not be big enough to train deep neural network models with many parameters. Thus, we plan to increase the number of data subjects no less than 20. This will raise the problem of data labeling, since all data come without any labels. We plan to use automatic or semi-automatic labeling as we did before, but this time using some confidence measures in order to improve the automatic labeling accuracy.
Finally, we plan to compare our system with some other existing cough detection and monitoring systems such as the Leicestor monitor and show how much better our system performs.
|
| 次年度使用額が生じた理由 |
One of the Co-Investigators was unable to spend the 2020 fund because of the COVID-19 pandemic, but will use it during year 2022.
|
