Figure 1.

Illustration of (a) feature vector extraction from a given keystroke dynamics variable of a typing session and (b) classification pipeline of each subject based on hold time (HT), normalised flight time (NFT) and pressure (NP) information. (a) Given a keystroke dynamics variable sequence a_n, a ∈ {HT, NFT, NP}: (1) The sequence is split in subsequences $a_n^i$ using 15-seconds non-overlapping time windows; (2) For each subsequence, the first- up to fourth-order statistical moments (mean μ_i, standard deviation σ_i, kurtosis K_i, and skewness S_i) of the elements are computed; (3) The probability density function (PDF) f ⁱ(x) of each subsequence is estimated through kernel density estimation (KDE) and the matrix of sample covariance C(i, j) between the PDFs of all subsequences is calculated. Feature vectors v_a representing each typing session are formed by the mean $\overline{•}$ and standard deviation (std) σ_⋅ of the moments extracted in (2), across time windows (subsequences), and the mean, std and sum of absolute values of the upper triangle C_U(i, j) of the covariance matrix calculated in (3). (b) The proposed two-stage multi-model pipeline for classifying subjects as PD patients or healthy controls: (1st Stage) Feature vector sets {v_a} of a given subject, with each vector representing a typing session, serve as input to three trained models M_a, each one dedicated to a keystroke dynamics variable, a ∈ {NFT, HT, NP}. Models M_a yield three prediction probabilities P_a which are then grouped in new feature vectors v_P; (2nd Stage) Feature vector set {v_P} serves as input to a Logistic Regression classifier C_LR that outputs the final classification probabilities {P_f} denoting whether each typing session belongs to a PD patient or a healthy control. Finally, the mean of prediction probabilities P_f is used to categorise the subject as PD patient or healthy control.