A feature fusion sequence learning approach for quantitative analysis of tremor symptoms based on digital handwriting
Introduction
Essential tremor (ET) is a neurodegenerative disease that affects approximately 6.38 to 7.63 million people in the United States, close to 2.2% of the population (Louis, 2014). ET typically manifests as postural and intentional tremors involving the upper extremities at a frequency of 8–12 Hz (Louis, 2019), and ET patients have a more than 4-fold increased risk of Parkinson's disease (PD) (Benito-Leon, Louis, & Bermejo-Pareja, 2009). The office history and examination are the sole components required for diagnosis (Shanker, 2019). Neurologists observe patients performing several specific tasks with the naked eye and score each task according to the Clinical Rating Scale for Tremor (CRST) criteria (Fernandez et al., 2019). However, such a protocol depends on the clinical experience and expertise of the physician and a thorough review of the patient, and subjective bias leads to unsatisfactory results in the assessment of movement disorders (Kassubek, 2014). In addition, tremor symptoms can develop and change over time, and manual tracking of disease progression often takes up significant healthcare resources. Further, assessing the disease remains challenging for the elderly, those with poor cognitive abilities or limited mobility (Hu et al., 2011). Therefore, early diagnosis of the disease adapted to the needs of individual patients with movement disorders and accurate assessment of symptom severity is crucial for prognosis and treatment (Ossig & Reichmann, 2015). With this goal in mind, the interest in computer-assisted diagnosis has grown from all walks of life in society over the years (Ali et al., 2019, Parisi et al., 2018). There is an urgent need to provide physicians with objective, multi-modal, quantitative evaluation criteria to aid decision-making.
Neurological disorders that cause brain changes caused by ET and PD, such as neuronal loss, synaptic dysfunction, brain atrophy, etc., may lead to dysfunction of the motor system and its components. According to this view, the unique role of handwriting can be confidently assumed in the context of symptom assessment. Handwriting is a complex activity involving perceptual-motor and cognitive components, and its changes can be considered a promising biomarker for disease assessment (De Stefano et al., 2019, Faundez-Zanuy et al., 2020, Vessio, 2019). A growing body of knowledge has proved that PD and healthy individuals can be automatically distinguished through a simple and easy drawing task (Angelillo et al., 2019, Diaz et al., 2019, Diaz et al., 2021, Drotár et al., 2013, Drotár et al., 2014, Drotár et al., 2014, Drotár et al., 2016, Gupta and Bansal, 2019, Impedovo, 2019, Kamran et al., 2021, Moetesum et al., 2019, Naseer et al., 2020). However, no studies have collected enough handwriting data from ET patients and obtained reliable expert scores to support the construction of an automated quantitative tremor assessment system. Developing a handwriting-based decision support tool is desirable because it could provide a noninvasive, real-time, low-cost solution to optimize standard clinical assessments performed by human experts.
Following this research direction, an online (dynamic) system that uses a digital tablet combined with a stylus with high-precision sensing can be employed (Diaz, Moetesum, Siddiqi, & Vessio, 2021). Such a device can capture temporal-spatial variables during the writing process, as well as information about the pressure exerted by the pen tip on the writing surface, the tilt direction, and even the measurement of the rotation angle. In contrast to the offline (static) morphological characteristics of handwriting, dynamic handwriting analysis deals with the kinetic characteristics during the writing process and is more suitable for disease diagnosis (Angelillo et al., 2019, Diaz et al., 2021, Gupta and Bansal, 2019, Impedovo, 2019). Therefore, selecting the appropriate features is crucial for designing automatic diagnostic systems. Classical statistical feature classifiers focus on the original time series and are susceptible to high dimensionality burden, which leads to overfitting. For this reason, some kinetic features are derived from the raw data and can represent high-level information in handwriting, such as entropy features characterizing sequence complexity or signal-to-noise features (Drotár et al., 2014, Impedovo, 2019, Rosenblum et al., 2013). Although such high-level features can have the ability to aid the model to obtain global information about the handwriting, on the other hand, it also loses detailed local features since arbitrarily long sequences are compressed into single-valued features.
Another approach to obtain tremor information in digital handwriting is to use deep learning models to automatically sense differences in information. Some recent work based on convolutional neural networks (CNN) has exploited morphological features extracted from two-dimensional static images to automatically achieve the recognition of tremor (Diaz et al., 2019, Kamran et al., 2021, Moetesum et al., 2019, Naseer et al., 2020). While this approach can replace manually designed features, it only provides a holistic view of handwriting patterns, and dynamic changes in the temporal direction may also help enhance the learning effect. Kamran et al. (Kamran et al., 2021) constructed a recurrent neural network (RNN) model to learn the dynamic features of electronic handwriting sequences, but it also neglected to explore global information that would likely provide the network with meaningful additional information. The cutting-edge deep-learning model like Transformer, which has natural advantages in processing time series data without losing spatial–temporal information, have not been fully explored. Transformer has significant advantages in almost all domains (including computer vision), with advantages such as its self-attention mechanism for efficient acquisition of high-level features. Rather than compressing the raw data as many researchers have done, we would like to take advantage of the time series nature of the data and fully consider multi-scale channel and time orientation information, with the final decision layer considering fusing depth features and global single-value features to accommodate digital handwriting of variable length.
This paper reports the first attempt to present an automated quantitative assessment system based on Transformer model for the drawing task in movement disorders, as shown in Fig. 1. The system is based on a highly sensitive digital tablet that collects handwriting from a high-dimensional multisensory fusion. The method explores tremor patterns through a cascaded deep network. 1) Constructing multi-scale squeeze-and-excitation-ResNet (MSSE-ResNet) based on raw digital ink data and dynamic features, which serves as an embedding of Transformer model layer to provide meaningful high-level features for the encoder. 2) Further construct encoder with multi-headed self-attentive mechanism, and fuse global kinematic features of handwriting before the classification decision layer to output the severity of tremors. The main contributions of this paper are summarized as follows:
1) Relying on rigorous patient screening experiments (biochemistry, imaging, physical examination, medical history), independent scoring by the neurological expert committee, and a comprehensive consensus score for supervised learning, we established the first essential tremor patient composed of the drawing task tremor inspection task database.
2) We propose the first digital ink Transformer model based on sequence learning combined with a CNN deep feature extraction network for an automated quantification system of movement disorder diseases.
3) We designed complete validation experiments to demonstrate the optimal performance of the proposed method in the evaluation of ET tremor and early diagnosis of PD.
Section snippets
Subjects and protocol
This study was based on the clinical trial “Efficacy and Safety Study of ExAblate Transcranial MRgFUS Thalamic Disruption for Drug-Refractory Idiopathic Tremor” (trial protocol code ET002J) at PLA General Hospital. It was approved by the Ethics Committee of Chinese PLA General Hospital (S2018-021–00/01). The ET002J clinical trial is part of a prospective, single-arm, multi-center clinical trial bid by InSightec (ClinicalTrials.gov Identifier: NCT03253991). Patients with symptoms of movement
Experiments
In this section, we have designed several experiments for the proposed method to evaluate its effectiveness:
1) Evaluate the potential of the proposed approach in quantifying tremor severity and explore the impact of multi-modal information on model performance using multiple individual subsets of digital ink features to confirm the validity of the design features.
2) Evaluate the impact of some architectural choices in our modified DIT model on quantifying tremor performance to verify the
Conclusions
Computer-aided digital handwriting analysis holds promise for a wide range of applications, such as handwriting measures that can capture individuals' physical and cognitive characteristics. Patients with movement disorders often develop symptoms of writing difficulties, so the analysis of electronic handwriting is hopeful to help early clinical diagnosis and even automatically quantify the severity of symptoms. Domain experts can access these easy-to-use, user-friendly tools in their daily
Funding
This work was supported by Natural Science Foundation of China (6217012292), Beijing Municipal Science and Technology (Z181100001918023) and Big Data Research & Development Project of Chinese PLA General Hospital (2018MBD-08, 2018MBD-09).
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (45)
- et al.
Early diagnosis of Parkinson’s disease from multiple voice recordings by simultaneous sample and feature selection
Expert Systems with Applications
(2019) - et al.
Handwriting in patients with Parkinson disease: Effect of L-dopa and stimulation of the sub-thalamic nucleus on motor anticipation
Human Movement Science
(2011) - et al.
Handwriting analysis to support neurodegenerative diseases diagnosis: A review
Pattern Recognition Letters
(2019) - et al.
Dynamically enhanced static handwriting representation for Parkinson’s disease detection
Pattern Recognition Letters
(2019) - et al.
Sequence-based dynamic handwriting analysis for Parkinson’s disease detection with one-dimensional convolutions and BiGRUs
Expert Systems with Applications
(2021) - et al.
Analysis of in-air movement in handwriting: A novel marker for Parkinson's disease
Computer Methods Programs in Biomedicine
(2014) - et al.
Evaluation of handwriting kinematics and pressure for differential diagnosis of Parkinson's disease
Artificial Intelligence in Medicine
(2016) - et al.
Unsupervised home monitoring of Parkinson's disease motor symptoms using body-worn accelerometers
Parkinsonism Related Disorders
(2016) - et al.
Agreement, the F-Measure, and Reliability in Information Retrieval
Journal of the American Medical Informatics Association
(2005) - et al.
How common and what are the determinants of sub-optimal care for Parkinson’s disease patients: The Milton Keynes community study
Parkinsonism Related Disorders
(2011)
Handwriting dynamics assessment using deep neural network for early identification of Parkinson’s disease
Future Generation Computer Systems
Diagnostic procedures during the course of Parkinson's Disease
Basal Ganglia
Machine learning-based classification of simple drawing movements in Parkinson's disease
Biomedical Signal Processing and Control
Dynamic signatures: A review of dynamic feature variation and forensic methodology
Forensic Science International
Assessing visual attributes of handwriting for prediction of neurological disorders—A case study on Parkinson’s disease
Pattern Recognition Letters
Treatment strategies in early and advanced Parkinson disease
Neurologic clinics
Feature-driven machine learning to improve early diagnosis of Parkinson's disease
Expert Systems with Applications
Performance-driven handwriting task selection for Parkinson’s disease classification
Differential diagnosis between Parkinson's disease and essential tremor using the smartphone's accelerometer
PLoS ONE
Risk of incident Parkinson’s disease and parkinsonism in essential tremor: A population based study, Journal of Neurology, Neurosurgery
Psychiatry
Hypometria and bradykinesia during drawing movements in individuals with Parkinson’s disease
Experimental Brain Research
A perspective analysis of handwritten signature technology
The Journal of the ACM Computing Surveys
Cited by (6)
Parkinson's disease diagnosis using deep learning: A bibliometric analysis and literature review
2024, Ageing Research ReviewsTremor detection Transformer: An automatic symptom assessment framework based on refined whole-body pose estimation
2023, Engineering Applications of Artificial IntelligenceAutomatic diagnosis of multi-task in essential tremor: Dynamic handwriting analysis using multi-modal fusion neural network
2023, Future Generation Computer SystemsTwo-stage framework for automatic diagnosis of multi-task in essential tremor via multi-sensory fusion parameters
2022, Journal of King Saud University - Computer and Information SciencesCitation Excerpt :Most of these advanced approaches are based on high-precision digital sensing information, collecting task-specific data from subjects, and designing intelligent algorithms for automated diagnosis. For example, spiral drawing can be performed directly on a digital pad, and the digital ink can record kinetic information such as nib pressure and acceleration for more accurate quantification than paper handwriting (Ma et al., 2022). In addition, ACC can visually measure the object's state of motion, providing information such as adequate tremor amplitude and frequency (Khodakarami et al., 2019; Battista and Romaniello, 2020; van Brummelen et al., 2020; San-Segundo et al., 2020; Sigcha et al., 2021).
Parkinson’s disease diagnosis using recurrent neural network based deep learning model by analyzing online handwriting
2024, Multimedia Tools and Applications