Transfer learning between deep neural networks using heterogeneous electrical biosignals

Andrea Leão Jesus Menezes dos Santos

Transfer learning between deep neural networks using heterogeneous electrical biosignals

The global health systems are currently unable to adequately meet the high demand for care for people with neurological disorders. This impacts the quality of treatment offered, leading to issues such as the prescription of improper medications, difficulty accessing treatment, late detection of diseases, and more. Neurological disorders include conditions such as dementia, epilepsy, Alzheimer's, Parkinson's, multiple sclerosis, and others. To improve the treatment of these diseases, devices for the acquisition of electrical biosignals have been developed to provide greater accuracy, patient comfort, and, in some cases, lower costs. Recognizing this scenario, we aimed to investigate the possibility of using transfer learning among artificial neural networks to address these problems. Additionally, we attempted to reduce the mathematical complexity of electrical biosignal data by transforming it from time domain to frequency domain, representing it as algebraic functions rather than mathematical sine functions. Based on these ideas, we explored the potential of transfer learning to enhance the predictive accuracy of a neural network model processing diverse electrical biosignals with non-identical feature and label spaces in a frequency domain. To prevent negative transfer learning, we integrated similarity analysis between biosignals into our methodology using the dynamic time warping (DTW) technique. We selected the long short-term memory (LSTM) neural network to develop the proposed architecture, and the public datasets used for the experiment were the TUEG EEG Corpora (electroencephalogram), ECG Heartbeat Categorization (electrocardiogram), and EMG Classify Gestures (electroneuromyography). Using the baseline outcomes as a reference, we selected the ECG as the source domain. Then, we calculated the similarity between the biosignals, trained the model with the features identified as having the lowest distance, and transferred the weights and bias to the EEG and EMG models to process their own dataset, named the target domain. In summary, we present two different scenarios to experiment and explore the potential of an effective transfer learning application with heterogeneous electrical biosignals in the frequency domain, from ECG to EMG and from ECG to EEG, respectively. In the first scenario, we discovered a promising outcome when the source and target datasets were balanced, even with a small target dataset. In the second context, we observed a discreet decrease in performance, also referred to as negative transfer learning, when utilizing a balanced source domain with an imbalanced and robust target dataset. Although we encountered some limitations, such as the high computational cost of calculating the similarity between the biosignals and the preprocessing strategy applied, among others detailed in this work, our experiment demonstrated the potential for transferring learning between neural networks processing heterogeneous electric biosignal dataset.

Luciano Rebouças de Oliveira

Marcos Ennes Barreto

Michele Fúlvia Angelo - UEFS

http://lattes.cnpq.br/6032273849847285

Vinicius Gadis Ribeiro - UFRGS

http://lattes.cnpq.br/2937182050702659

Jefferson Fontinele - UFMA

http://lattes.cnpq.br/4592887904839028

Kalyf Abdalla - IFBA

http://lattes.cnpq.br/4613476711431079

20 Sep, 2024

2:00 PM

The main contributions of our research are as follows: We present an exploratory study on heterogeneous transfer learning using three distinct electrical biosignal datasets (ECG, EMG, and EEG) in the frequency domain. In this context, our findings suggest that it is feasible to apply transfer learning between neural networks processing different types of electrical biosignals, particularly between ECG and EMG. Additionally, our work indicates that processing data in the frequency domain can mitigate the effects of unbalanced data; however, it was not sufficiently effective to avoid slight negative transfer learning in the ECG-EEG scenario. Finally, our study highlights that the similarity analysis did not significantly enhance the accuracy of the target models.