Link articolo originale

Archivio di tutti i clip:
(Notebook di Evernote).

Cortical control of a tablet computer by people with paralysis

Materials and methods

Permission for these studies was granted by the US Food and Drug Administration (Investigational Device Exemption) and the Institutional Review Boards of Stanford University, Providence Veterans Affairs Medical Center, Brown University, and Massachusetts General Hospital. The participants in this study were enrolled in a pilot clinical trial of the BrainGate2 Neural Interface System ( Identifier: NCT00912041).

Participants were enrolled according to the inclusion and exclusion criteria of the clinical trial, and informed consent was obtained for all study-related protocols and procedures. Separate consent to publish photos and video was also obtained.
Participant T6 is a right-handed woman, 51 years old at time of study enrollment, diagnosed with ALS and with resultant motor impairment. In December 2012, a 96-channel intracortical microelectrode array (1.0-mm electrode length, 4 × 4 mm, Blackrock Microsystems, Salt Lake City, UT) was placed in the hand area of dominant motor cortex as previously described [26, 34]. At the time of this study, T6 retained speech and dexterous movements of her wrists and some fingers (ALSFRS(R) = 14). Data reported in this study are from T6’s post-implant trial days 1013, 1018, and 1034.
Participant T9 was a right-handed man, 51 years old at time of study enrollment, also diagnosed with ALS. In February 2015, he had two microelectrode arrays (1.5-mm electrode length, same manufacturer) placed in the hand area of dominant motor cortex. At the time of this study, T9 retained speech and had minimal and nonfunctional movement of the fingers (ALSFRS(R) = 6). Data reported in this study are from T9’s post-implant trial days 218, 222, and 225.
Participant T5 is a right-handed man, 63 years old at the time of study enrollment, with tetraplegia due to a C4 ASIA C cervical spinal cord injury. In August 2016, he had two microelectrode arrays (1.5-mm electrode length, same manufacturer) placed in the hand and arm area of dominant motor cortex. At the time of this study, T5 retained speech and had minimal and nonfunctional movement of the fingers. Data reported in this study are from T5’s post-implant trial days 121, 124, and 140. A fourth session (post-implant trial day 126) was also attempted, but was unsuccessful because of a cable malfunction (which was subsequently remedied).

Research setup
The research setup was similar to prior reports [26, 31–33, 35] for the purposes of data recording, processing, and analysis. A NeuroPort recording system (Blackrock Microsystems, Salt Lake City, UT) recorded neural signals from the participant’s motor cortex. These signals were routed into a custom real-time computer running the xPC/Simulink Real-Time operating system (Mathworks, Natick, MA) for processing and decoding. The output of the decoding algorithm was passed to a Bluetooth interface configured to work as a conventional wireless computer mouse using the Bluetooth Human Interface Device (HID) Profile. This virtual Bluetooth mouse was paired with a commercial Android tablet device (Google Nexus 9, Android OS 5.1) with no modifications to the operating system. Each participant viewed the device at their preferred comfortable distance, typically 40-60 cm from the eyes. No accessibility software was installed on the tablet, and no built-in accessibility features were enabled. Participants performed real-time “point-and-click” control over a cursor that appeared on the tablet computer once paired through the Bluetooth interface. Fig 1a details the flow of information from the participant to the tablet device. Advanced cursor features such as click-and-hold, multitouch, and gestures were not implemented in this study.

Fig 1. Research setup.

a Schematic of research setup with T6. We recorded from 96-channel electrode arrays implanted in motor cortex. The neural signals extracted from the arrays were passed into a decoding algorithm which output a two dimensional cursor velocity and a click signal. The output of the decoder was presented as a wireless Bluetooth mouse interface and paired with a computer tablet. The participants used this interface to control the tablet and perform common tasks like email and web browsing. b Example task timeline with T5 from trial day 124. Shortest vertical black lines represent general user interface clicks, shorter gray lines represent single character text entry, and taller gray lines represent autocompletion of text.

Videos of the study were captured in two ways. An external DSLR camera was positioned to record the participant as they controlled the tablet. Simultaneously, a screen capture program (AZ Screen Recorder, Hecorat) running on the tablet recorded all activity on the tablet as a video.

Neural decoders
In this study, intended cursor movements and clicks were decoded from neural activity using Kalman filters for cursor movement and state classifiers for click detection. 2D cursor velocities were estimated using a Recalibrated Feedback Intention Trained Kalman Filter (ReFIT-KF) for T6 and T5 [21, 31, 33] and a cumulative closed-loop decoder for participant T9 [35]. Briefly, the ReFIT-KF is a decoder built in a two-step fashion which attempts to correct the kinematics of first-pass iBCI control by assuming intention to move directly to the target, leading to improved performance. The cumulative closed-loop decoder is typically initialized using neural data recorded during an open-loop task. Additional data, recorded during closed-loop neural control, are then used to update decoder parameters, with the aim of refining the tuning model [35]. In order to reduce calibration time, it is also possible to seed the decoder with parameters from the previous research session, as was the case on T9’s trial days 222 and 225. Different decoders were used in this study because we aimed to highlight iBCI reliability and robustness. Being relatively decoder agnostic demonstrates that the performance achieved here is not intricately linked to the specifics of a single decoder, but that multiple decoding approaches can successfully drive a common communication device. Click intentions were classified using a hidden Markov model for T6 and T5 [24, 33] and a linear discriminant analysis classifier for T9 [34]. Participants each had their own imagery to enact a click. T6 attempted squeezing her left hand T5 attempted flexing his left arm. T9 attempted squeezing his right hand. The duration of the calibration blocks (excluding voluntary participant pauses between blocks) used to initialze the decoder, in minutes, for each day and each participant were: T6 (10, 12, 8), T9 (25, 4, 12), and T5 (20, 12, 16). Methods for further reducing this initial calibration period have been implemented more recently [36].
To initialize and calibrate the decoders, participants engaged in a center-out-back task described previously [31, 33, 35]. These decoders were built in a stepwise fashion, with the first stage of filter calibration performed as the cursor moved automatically to the targets while the participants imagined or attempted moving their hand as though they were controlling the cursor. This allowed the initialization of a decoder that was then improved upon in subsequent calibration blocks. The Kalman filters were also running bias correction algorithms throughout the task [32]. For T6 and T5, once core data collection began (see below), there were no decoder modifications or interruptions aside from voluntary inter-task breaks. Decoder bias re-estimation blocks were permitted as needed during the free-time period that followed core data collection when T6 and T5 were using the tablet to explore their interests. For T9, no decoder modifications or interruptions aside from voluntary inter-task breaks were performed once he started using the tablet.
Recorded signal quality can affect decoding performance, however this relationship was not specifically evaluated in this study. To better understand the signal quality of each participant’s neural data, plots of thresholded spiking activity for each participant were taken from the start of a research day. These appear in Fig 2. Participant T5 had the largest single units across his arrays while Participant T6’s array had the least number of distinguishable single units. Further detail on the relationship between signal quality and decoding performance can be found in prior reports [29, 32, 33, 35].

Fig 2. Thresholded spiking actvity of participants’ arrays.

Each panel, corresponding to a specified 96-channel array, shows the threshold crossing waveforms recorded over 60 seconds on the specified trial day. a is T6’s array. b and c are T5’s lateral and medial arrays, respectively. d and e are T9’s lateral and medial arrays, respectively. Scale bars represent 150 uV (vertical) and 500 us (horizontal). Data are from the following trial days: 1013 (T6), 124 (T5), and 218 (T9). Plot construction identical to that of Fig 5 of [33].

Task design
Once the decoder was calibrated, the tablet was paired with the BCI system. The technician ensured that the tablet displayed the home screen at the start of each session. Aside from ensuring that the cursor was active and under iBCI control by the participant, the technician did not otherwise intervene during tablet use. Participants used seven common applications on the tablet: an email client, a chat program, a web browser, a weather program, a news aggregator, a video sharing program, and a streaming music program. The applications used by the participants were either preinstalled with the tablet or downloaded by one of the research members from the Play Store (Google, Mountain View, CA) prior to the first day of the study. Participants were asked to launch each target application from the home screen, use as requested, and exit the program by returning to the home screen. Details of the specific tasks and programs appear in Table 1. Each participant completed the entire task design on each of three days. Tasks included periods of participant-determined actions (e.g., personal choice of typing topics) such that the number of clicks required for task completion varied across participants. For typing performance (assessed on email and chat tasks), duration was counted from the time the keyboard was activated by the participant to the time the last character or word was entered. Selections include all printed and non-printed characters (e.g., shift and delete keys). Effective characters are all printed characters that appeared as transmitted text. In addition to these structured tasks, each participant was asked in advance which additional consumer applications they would enjoy using. These applications were then downloaded from the Play Store. After completing the standard tasks, each participant proceeded to use their selected applications. On a separate day (implant day 1211 for T6 and implant day 416 for T9), T6 and T9 used the chat program to send messages to each other in real time. The research session ended at the participants’ discretion.