Keystroke-Dynamics for Parkinson's Disease Signs Detection in an At-Home Uncontrolled Population: A New Benchmark and Method

Abstract

Parkinson's disease (PD) is the second most prevalent neurodegenerative disease disorder in the world. A prompt diagnosis would enable clinical trials for disease-modifying neuroprotective therapies. Recent research efforts have unveiled imaging and blood markers that have the potential to be used to identify PD patients promptly, however, the idiopathic nature of PD makes these tests very hard to scale to the general population. To this end, we need an easily deployable tool that would enable screening for PD signs in the general population. In this work, we propose a new set of features based on keystroke dynamics, i.e., the time required to press and release keyboard keys during typing, and used to detect PD in an ecologically valid data acquisition setup at the subject's homes, without requiring any pre-defined task. We compare and contrast existing models presented in the literature and present a new model that combines a new type of keystroke dynamics signal representation using hold time and flight time series as a function of key types and asymmetry in the time series using a convolutional neural network. We show how this model achieves an Area Under the Receiving Operating Characteristic curve ranging from 0.80 to 0.83 on a dataset of subjects who actively interacted with their computers for at least 5 months and positively compares against other state-of-the-art approaches previously tested on keystroke dynamics data acquired with mechanical keyboards.

Fig 1.

Keystroke dynamics features.

Fig 2.

The figure above shows the complete pipeline of the feature extraction process. The HT/FT sequences are used to generate the features maps GEN, ASYM, and COG using the key type of interest (KTI). The feature extraction method is expanded and shows the transformation of HT/FT signals to the binned probability densities (20) in the form of matrices. These matrices are broken down into fixed size matrices of dimension 21 × 20. The fixed size hold-time matrices (htm) are stacked with the fixed size flight-time matrices (ftm) along the depth. The stacking is performed for every time a HT and FT sequence is processed, determining the depth of the final feature matrices.

Fig 3.

The convolutional neural network architecture consists of 4 convolutional layers namely, Conv Layer 1, 2, 3, and 4. Conv Layer 1 and 3 have a filter of height 2 and width 3. Conv Layer 2 and 4 have a filter of height 2 and width 5. The output classes are Parkinson’s and Control.

Fig 4.

The attributions (positive, negative, and zero) for the confirmed diagnosis set. The positive (blue) attributions are obtained after setting a threshold of greater than 0 in the attribution matrix. Similarly, negative attributions (red) are generated after setting a threshold of less than 0 in the attribution matrix. Zero valued attributions are the values in the attribution matrix equal to 0. “All-keys-HT” and “All-keys-FT” do not have any zero-valued attributions since they project the sum of attributions from all the keys present in the set and thus unless all the attributions are 0 the total of the attribution is a non-zero value. Attribution matrix is obtained using integrated gradients [26].

Fig 5.

The attributions (positive, negative, and zero) for the unconfirmed diagnosis set. The positive (blue) attributions are obtained after setting a threshold of greater than 0 in the attribution matrix. Similarly, negative attributions (red) are generated after setting a threshold of less than 0 in the attribution matrix. Zero valued attributions are the values in the attribution matrix equal to 0. “All-keys-HT” and “All-keys-FT” do not have any zero-valued attributions since they project the sum of attributions from all the keys present in the set and thus unless all the attributions are 0 the total of the attribution is a non-zero value. Attribution matrix is obtained using integrated gradients [26].

Fig 6.

The figure shows a stacked bar chart for the positive attributions for each of the valid windows summed across all subjects in both the confirmed and unconfirmed diagnosis arms. The feature representations used to train GEN-ASYM-NET are color-coded and labeled as shown in the figure legend. The y-axis represents the number of days over which the keystroke activity was evaluated. The x-axis represents the values of positive attributions summed across all subjects. The positive attributions are obtained after setting a threshold of greater than 0 in the attribution matrix obtained after applying integrated gradients [26] on the input feature matrix. Positive attributions represent the components of the feature matrix that contribute positively during training.

Fig 7.

Receiver operating characteristic curves (ROC) showing the comparison of the performances of methods GEN-COG_NET, GEN-ASYM-NET, GEN-NET, presented in the paper. The performance of the methods shown in the figure above is evaluated by averaging the AUC scores over 5-folds and 30 repetitions for the unconfirmed and confirmed diagnosis sets. After averaging the AUC scores, we see the GEN-ASYM-NET outperforms the rest of the proposed methods.