Time Series Classification Techniques Used in Biomedical Applications: History
View Latest Version
Please note this is an old version of this entry, which may differ significantly from the current revision.
Contributor:
Will Ke Wang
,
Ina Chen
,
Leeor Hershkovich
,
Jiamu Yang
,
Ayush Shetty
,
Geetika Singh
,
Yihang Jiang
,
Aditya Kotla
,
Jason Zisheng Shang
,
Rushil Yerrabelli
,
Ali R. Roghanizad
,
Md Mobashir Hasan Shandhi
,
Jessilyn Dunn

Time series classification (TSC) is very commonly used for modeling digital clinical measures. Time Series Classification (TSC) involves building predictive models that output a target variable or label from inputs of longitudinal or sequential observations across some time period. These inputs could be from a single variable or multiple variables measured across time, where the measurements can be ordinal or numerical (discrete or continuous).

  • time series classification
  • digital clinical measures
  • machine learning

1. Introduction

Time series data are a very common form of data, containing information about the (changing) state of any variable. Some common examples include stock market prices and temperature values across some period of time. Time series modeling tasks include classification, regression, and forecasting. There are unique challenges that come with modeling time series, given that measurements obtained in real-life settings are subject to random noise, and that any measurement at a particular point in time could be related to or influenced by measurements at other points in time [1]. Given this nature of time series data, it is impractical to simply utilize established machine learning algorithms such as logistic regression, support vector machine, or random forest on the raw time series datasets because these data violate the basic assumptions of those models. In recent years, two vastly different camps of time series classification techniques have emerged: deep-learning-based models vs non-deep-learning-based models. While deep learning models are extremely powerful and show great promise in classification performance and generalizability, they also present challenges in the areas of hyperparameter tuning, training, and model complexity decisions. 

2. Time Series Classification Techniques Used in Biomedical Applications

2.1. Preprocessing Methods

The most common preprocessing method is filtering, which is used mainly for artifact removal or noise reduction. Some other common preprocessing methods include re-sampling (downsampling for lower frequency or upsampling for higher frequency), segmentation, and smoothing. Other common methods are the use of discrete wavelet transform to decompose the original signal into different frequency bands [2][3][4], the use of continuous wavelet transform to expand the feature space [5], and the use of Fourier transform for signal decomposition and feature extraction [6][7]. There are also intelligent upsampling techniques, such as the use of synthetic data generation for a larger sample during preprocessing [8].

2.2. Feature Engineering Methods

Feature engineering is the most commonly used method of time series classification. The feature engineering pipeline usually consists of the following steps:
  • Preprocessing: this step takes raw data as the input and performs some manipulation of the data to return cleaner signals. Common steps include artifact removal, filtering, and segmentation.
  • Signal transformation: this step can be used in preprocessing and also as a precursor to feature extraction. Some manipulation is performed on the signal to represent it in a different space. Common choices are Fourier Transform and wavelet transforms.
  • Feature extraction: in this step, features are extracted from the time series data as a new representation of the original time series.
  • Feature selection: this step selects the features that are the most descriptive, or have the most explanation power. Feature selection is also frequently performed in conjunction with model building.
  • Model selection: the best model is found through hyperparameter tuning and/or comparisons between different types of algorithms.
  • Model validation: performance metrics are calculated for all of the final models. This is frequently done in conjunction with model selection and often using some form of cross-validation.
An example feature engineering technique for a time series is shown in Figure 1.
/media/item_content/202211/6368586b3ada6sensors-22-08016-g004.png
Figure 1. Illustration of different types of feature engineering techniques [9].

2.3. Other Methods

Ensemble Methods: Ensemble-based methods are characterized by the connection of multiple algorithmic models that join forces to make the final prediction. These methods may or may not need an additional feature engineering step. Some algorithms that do not necessitate feature engineering in this category are Hierarchical Vote Collective of Transformation-based Ensembles and Bag of Symbolic Fourier Approximation Symbols ensemble algorithms (BOSS) [10].
State-space Models: State-space models are characterized by the construction of a state and transition model where the transitions are modeled by probabilities. Often, state-space models are most intuitively used for sequence-to-sequence or point-wise classification. For example, She et al. [11] introduced an adaptive transfer learning algorithm to classify and segment events from non-stationary, multi-channel temporal data recorded by an Empatica E4 wristband, including 3-axis accelerometry (ACC), blood volume pulse (BVP), skin temperature (TEMP), and electrodermal activity (EDA). Using a multivariate Hidden Markov Model (HMM) and Fisher’s Linear Discriminant Analysis (FLDA), the algorithm adaptively adjusts to shifts in the distribution over time, thereby achieving an accuracy of 0.9981 and F1-score of 0.9987. 
Shape/Pattern-based: These models are characterized by mining or comparing shapes or patterns in a time or sequence vector. For example, Zhou et al. [12] published an algorithm that can take into consideration the interaction among signals collected at spatiotemporally distinct points, where fuzzy temporal patterns are used to characterize and differentiate between different classes of multichannel EEG data. This algorithm achieved an accuracy of 0.9318 and an F1-score of 0.931, thereby classifying positive vs negative emotion states.
Distance-based: These models calculate the distance (or differences) of time series data vectors. For example, Forestier et al. [13] propose an efficient algorithm to find the optimal partial alignment (optimal subsequence matching) and a prediction system for multivariate signals using maximum a posteriori probability estimation and filtering. This scoring function is based on dynamic time warping. They were able to achieve an accuracy of 0.95, an F1-score of 0.926, and a sensitivity of 0.896.
Other: There are other methodologies that are difficult to characterize. One common method is performed by using statistical modeling of some sort. For example, İşcan et al. [14] published a high performance method to classify and discriminate various ECG patterns (to identify and classify QRS complexes). The model is called LLGMN, which is composed of a Log-Linear Model and a Gaussian Mixture Model (GMM), and gives a posterior probability for the training data. This model was able to achieve the highest accuracy, which was 0.9924.
Another common method is designing a composite metric or index based on domain knowledge or data-driven metrics. For example, Zhou et al. [15] proposed a new algorithm to detect gait events on three walking terrains in real-time based on an analysis of acceleration jerk signals with a time–frequency method to obtain gait parameters, as well as detecting the peaks of jerk signals using peak heuristics. The performance of the newly proposed algorithm was evaluated in eight healthy subjects walking on level ground, upstairs, and downstairs. The mean F1-score was above 0.98 for HS (heel-strike) event detection and 0.95 for TO (toe-off) event detection on the three terrains.

2.4. Interpretation Methods

Model interpretability is a significant aspect of model building. In time series classification for biomedical applications, the interpretation of models that have been built and validated could highlight potential insights into the biomedical phenomenon of interest. Some models have a built-in methodology of interpretation, such as statistical modeling (Hidden Markov Models, Bayesian Models, or ARIMA models) and indices that are informed based on domain knowledge. For many more models with great performance, however, interpretability is a challenge.

2.5. Best Performing Algorithms

Overall, the statistical modeling classifiers and feature engineering methods performed the best and most consistently for all input signal types. Wavelet transformation is consistently and widely used and achieving great performances as a preprocessing method, feature extraction method, or as an integral part of index development.

3. Conclusions

In conclusion, non-deep learning time series classification techniques can achieve competitive performances, while also allowing for great interpretability. However, this field still lacks standardization for model testing, validation procedures, and reporting metrics, which should be addressed to allow for better reproducibility and understanding of the  presented algorithms.

This entry is adapted from the peer-reviewed paper 10.3390/s22208016

References

  1. Bock, C.; Moor, M.; Jutzeler, C.R.; Borgwardt, K. Machine Learning for Biomedical Time Series Classification: From Shapelets to Deep Learning. In Artificial Neural Networks; Cartwright, H., Ed.; Springer: New York, NY, USA, 2021; pp. 33–71.
  2. Mole, S.S.S.; Sujatha, K. An efficient Gait Dynamics classification method for Neurodegenerative Diseases using Brain signals. J. Med. Syst. 2019, 43, 245.
  3. Joshi, D.; Khajuria, A.; Joshi, P. An automatic non-invasive method for Parkinson’s disease classification. Comput. Methods Programs Biomed. 2017, 145, 135–145.
  4. Tor, H.T.; Ooi, C.P.; Lim-Ashworth, N.S.; Wei, J.K.E.; Jahmunah, V.; Oh, S.L.; Acharya, U.R.; Fung, D.S.S. Automated detection of conduct disorder and attention deficit hyperactivity disorder using decomposition and nonlinear techniques with EEG signals. Comput. Methods Programs Biomed. 2021, 200, 105941.
  5. Mesbah, S.; Gonnelli, F.; Angeli, C.A.; El-Baz, A.; Harkema, S.J.; Rejc, E. Neurophysiological markers predicting recovery of standing in humans with chronic motor complete spinal cord injury. Sci. Rep. 2019, 9, 14474.
  6. Anh, N.X.; Nataraja, R.; Chauhan, S. Towards near real-time assessment of surgical skills: A comparison of feature extraction techniques. Comput. Methods Progr. Biomed. 2019, 187, 105234.
  7. Durongbhan, P.; Zhao, Y.; Chen, L.; Zis, P.; De Marco, M.; Unwin, Z.C.; Venneri, A.; He, X.; Li, S.; Zhao, Y.; et al. A Dementia Classification Framework Using Frequency and Time-Frequency Features Based on EEG Signals. IEEE Trans. Neural Syst. Rehabil. Eng. Publ. IEEE Eng. Med. Biol. Soc. 2019, 27, 826–835.
  8. Bhattacharya, S.; Mazumder, O.; Roy, D.; Sinha, A.; Ghose, A. Synthetic Data Generation Through Statistical Explosion: Improving Classification Accuracy of Coronary Artery Disease Using PPG. In Proceedings of the ICASSP 2020–2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 4–8 May 2020; pp. 1165–1169.
  9. Walter, S.; Gruss, S.; Limbrecht-Ecklundt, K.; Traue, H.C.; Werner, P.; Al-Hamadi, A.; Diniz, N.; da Silva, G.M.; Andrade, A.O. Automatic pain quantification using autonomic parameters. Psychol. Neurosci. 2014, 7, 363–380.
  10. Ruiz, A.P.; Flynn, M.; Large, J.; Middlehurst, M.; Bagnall, A. The great multivariate time series classification bake off: A review and experimental evaluation of recent algorithmic advances. Data Min. Knowl. Discov. 2020, 35, 401–449.
  11. She, X.; Zhai, Y.; Henao, R.; Woods, C.W.; Chiu, C.; Ginsburg, G.S.; Song, P.X.K.; Hero, A.O. Adaptive Multi-Channel Event Segmentation and Feature Extraction for Monitoring Health Outcomes. IEEE Trans. Biomed. Eng. 2020, 68, 2377–2388.
  12. Zhou, P.-Y.; Chan, K.C.C. Fuzzy Feature Extraction for Multichannel EEG Classification. IEEE Trans. Cogn. Dev. Syst. 2016, 10, 267–279.
  13. Forestier, G.; Petitjean, F.; Riffaud, L.; Jannin, P. Automatic matching of surgeries to predict surgeons’ next actions. Artif. Intell. Med. 2017, 81, 3–11.
  14. Iscan, M.; Yigit, F.; Yilmaz, C. Heartbeat pattern classification algorithm based on Gaussian mixture model. In Proceedings of the 2016 IEEE International Symposium on Medical Measurements and Applications (MeMeA), Benevento, Italy, 15–18 May 2016; pp. 1–6.
  15. Zhou, H.; Ji, N.; Samuel, O.W.; Cao, Y.; Zhao, Z.; Chen, S.; Li, G. Towards Real-Time Detection of Gait Events on Different Terrains Using Time-Frequency Analysis and Peak Heuristics Algorithm. Sensors 2016, 16, 1634.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.
This entry is offline, you can click here to edit this entry!
Video Production Service
Feedback