Fig 1.
(a) Schematic showing a subject lifting a weight from a table by following a posture sequence shown on the screen. The table height was adjusted such that trunk flexion was near 30° for each subject. (b) The sequence of posture images shown to the subject on the screen accompanied by a timed sound cue. At the start of each lift, the screen displayed a prompt informing the subject which weight to lift—no-weight, 10-lbs, or 24-lbs, then it displayed the posture sequence with a 1 second delay between each of the numbered (1)-(8) posture images. (c) Low-back muscle activity was measured from four surface EMG bipolar electrodes placed at L4/L5 vertebrae. (d) A subject performing a 24-lbs lift with a close up of the weights on the table. The no-weight case consisted of two sticks wrapped in foil and positioned to close a circuit between two charged foil railings. The force plate under the table can also be seen, flush with the floor, in the larger image.
Fig 2.
Flowcharts showing the steps involved in (a) data processing, segmentation and feature extraction, and (b) dimensionality reduction, feature selection, cross-validation and, finally, testing.
Fig 3.
Average muscle activity for different loading conditions.
The average muscle activity of the two left EMG channels for all nine subjects normalized to each subject’s maximum voluntary contraction (MVC) baseline. Time zero is the time of load-onset. The shaded regions show the region of ±1 standard deviation around the average. There is a clear spike in average activity around the load-onset time point that is more prominent with increasing lifted load values. The average of the right EMG channels showed similar activity patterns.
Fig 4.
Classification accuracy increases at time windows near load-onset.
The classification accuracy from the testing and the cross-validation steps for each 100 ms time window segment. The plot shows average accuracy from nine subjects. The shaded regions illustrate ± 1 standard deviation at each time point. The testing curve matches the validation curve, indicating that overfitting was avoided. Time zero indicates the time of load-onset.
Fig 5.
Distribution of classification results.
Confusion matrices showing the average classification results (with standard deviations in parenthesis) for the time windows with the highest accuracy during the regions (a) before and (b) after load-onset or full load support.
Fig 6.
Average classification recall for each weight class.
The average testing recall for each weight class at each time window for all nine test subjects. Time zero indicates the time of load-onset.
Fig 7.
Selected features and their selection frequency.
The percentage of times (out of 711) that each feature from each channel was selected for the optimal feature set as input by the greedy feedforward algorithm during cross-validation of the MLR classifier. The mean feature was selected much more frequently than the others.
Table 1.
A summary of relevant response times of muscles, classifiers, controllers and actuators commonly used in assistive device applications.
Fig 8.
Classification recall was affected by subject and weight class.
Classification recall for each subject during the optimal time windows before (Pre) and after (Post) load-onset. The interaction between subject and weight class had statistically significant affects on recall percentage.