Explainable artificial intelligence for searching frequency characteristics in Parkinson’s disease tremor

Abstract

The distinction between Parkinson’s disease (PD) and essential tremor (ET) tremors is subtle, posing challenges in differentiation. To accurately classify the PD and ET, BiLSTM-based recurrent neural networks are employed to classify between normal patients (N), PD patients, and ET patients using accelerometry data on their lower arm (L), hand (H), and upper arm (U) as inputs. The trained recurrent neural network (RNN) has reached 80% accuracy. The neural network is analyzed using layer-wise relevance propagation (LRP) to understand the internal workings of the neural network. A novel explainable AI method, called LRP-based approximate linear weights (ALW), is introduced to identify the similarities in relevance when assigning the class scores in the neural network. The ALW functions as a 2D kernel that linearly transforms the input data directly into the class scores, which significantly reduces the complexity of analyzing the neural network. This new classification method reconstructs the neural network’s original function, achieving a 73% PD and ET tremor classification accuracy. By analyzing the ALWs, the correlation between each input and the class can also be determined. Then, the differentiating features can be subsequently identified. Since the input is preprocessed using short-time Fourier transform (STFT), the differences between the magnitude of tremor frequencies ranging from 3 to 30 Hz in the mean N, PD, and ET subjects are successfully identified. Aside from matching the current medical knowledge on frequency content in the tremors, the differentiating features also provide insights about frequency contents in the tremors in other frequency bands and body parts.