System, computer-readable storage medium and method of deep learning of texture in short time series

A non-transitory computer-readable storage medium storing computer program instructions, which when executed by a computer perform a method of classification of short time series in order to detect a neurodegenerative disorder, comprising:

• receiving a plurality of sensor data collected from subjects with and without the neurodegenerative disorder over a period of a few seconds as the short time series;
• generating phase-space vectors from the plurality of sensor data in which each vector is a state of a dynamical system in space and time;
• transforming the phase-space vectors into a grayscale image representing recurrences of a state-space vector in the same area of the phase space, wherein the intensity distribution of the grayscale image takes values in the range of 0 to 1;
• for every pixel of the grayscale image, selecting a local window, where the respective pixel is at a center of the local window;
• extracting temporal texture features of each window to obtain a multi-dimensional time series of D texture dimensions;
• inputting the multidimensional time series, without the grayscale image, to a Long Short Term Memory (LSTM) network;
• training the LSTM network by:
• in an LSTM layer, inputting to a first LSTM block a first time step of the time series of D texture dimensions and determining a first output and an updated cell state,
• inputting to a next LSTM block a next time step of the time series of D texture dimensions, a current output and a current cell state and determining a next output and a next updated cell state, and
• continuing the inputting and the determining to LSTM blocks for each time step of the time series of D texture dimensions to output a final cell state;
• and
• classifying, by the LSTM network, the plurality of the sensor data as the neurodegenerative disorder or not.

A method of classification of short time series in order to detect a neurodegenerative disorder, comprising:

• receiving, by circuitry, a plurality of sensor data collected from subjects with and without the neurodegenerative disorder over a period of a few seconds as the short time series;
• generating, by the circuitry, phase-space vectors from the plurality of sensor data in which each vector is a state of a dynamical system in space and time;
• transforming, by the circuitry, the phase-space vectors into a grayscale image representing recurrences of a state-space vector in the same area of the phase space, wherein the intensity distribution of the grayscale image takes values in the range of 0 to 1;
• inputting, by the circuitry, the grayscale image representation to a Long Short Term Memory (LSTM) network;
• and
• classifying, by the LSTM network, the plurality of the sensor data as the neurodegenerative disorder or not;
• wherein the neurodegenerative disorder is Parkinson's disease (PD), and
• wherein the sensor data is of keystrokes collected from subjects with and without PD and is a time series in a range of 30 to 50 time steps of the keystroke hold duration.