Experimental study
Hand kinematics data from twenty-seven healthy subjects while performing the 20 tasks of SHFT was used [9] (Figure 1). Each subject performed these 20 ADL under controlled laboratory conditions, using real objects and following the original test instructions [5]. All the participants provided their informed consent to participate in the experiment (approved by the University Ethics Committees).
Sixteen joint angles were recorded (100 Hz) with an instrumented glove (Cyberglove Systems LLC; San Jose, CA (USA)) using a validated calibration protocol [10]: flexion of metacarpophalangeal joints (MCP1 to MCP5, 1 to 5 meaning thumb to little digits), flexion of interphalangeal thumb joint (IP1), flexion of proximal interphalangeal joints of the fingers (PIP2 to PIP5), flexion and abduction of the carpometacarpal thumb joint (CMC1), relative abduction between finger MCPs (index-middle, middle-ring, and ring-little), and palmar arching. Positive values of angles correspond to flexion and abduction. The recordings were filtered with a 2nd-order 2-way low-pass Butterworth filter with a cut-off frequency of 5 Hz.
We looked for the minimum number of ADLs that provides kinematic information similar to that considering all SHFT ADLs in terms of the most relevant kinematic parameters, by means of a synergy-based methodology, presented afterwards.
Analysis outline
Step 1. Extraction of subject-specific kinematic synergies underlying all the 20 ADLs from SHFT through computation of PCs. The resulting set of synergies (Ref PCs) are considered as reference for the next steps.
Step 2. Ordering of ADL from least to greatest effect on Ref PCs if removed. In each step, the ADL subtracted is the one having least effect on the resulting synergies when removed in comparison to Ref PCs. This process is repeated until only one ADL remains, and the result is a vector of ADLs sorted following the order of subtraction.
Step 3. For each subject, identification of the minimum number of ADLs, from the sorted ADLs of the previous step, that provide a range of movement of at least 85% that of the original data as well as keeping the joint coordination in terms of synergies (maximum angle between k-PCs and Ref PCs equal or lower than 30 degrees).
Step 4. Selection of the reduced set of ADLs for hand function assessment as those that appears more frequently in all subjects.
Step 5. Verification that the reduced set of ADLs is equivalent to the original set of ADLs in terms of motion strategies, range of motion and velocities of hand joints on two different populations; healthy subjects and patients with Hand Osteoarthritis (HOA).
Detailed analysis
Step 1 - Kinematic PC extraction
The Data Analysis was fully performed with a custom-developed software in Matlab. First, each SHFT task was resampled to 1000 frames so that all tasks weighed the same when looking for underlying synergies. Therefore, the data used throughout all the paper consist of 20 records of 1000 frames for each participant, resulting in a matrix with 16 columns (joint angles) and 20000 rows (20 ADL x 1000 frames). Four synergies were extracted for each subject (Ref PCs) using PCA following the methodology presented in previous studies, which considers normalized factors and varimax rotation [11–14]. The loadings of the Ref PCs (representing the kinematic synergies) and the corresponding scores (representing the dimensionality reduced kinematic data) are considered the descriptive data to be maintained in the ADL subtraction.
Step 2 - Ordering of ADL from least to greatest effect on Ref PCs if removed
An iterative subtraction method was implemented based on PCA. PCA is applied per subject to avoid unrealistic synergies that may appear when merging data from different subjects [15]. The scheme of the used method is shown in Fig 2. For each subject, the data of each ADL were tentatively removed one-by-one, and the resulting N datasets were used as input in N PCAs (one PCA, also with four PCs extracted, per each ADL removed; N=20-k, in the k-th step). Differently from the first kinematic PCA extraction, a non-standard scaling was applied [12], using the mean and SD of the original matrix, in order to allow the comparison between Ref PCs and PCs from k-th step (k-PCs). In each k-th step, the angles between the four k-PCs and Ref PCs vectors were computed to order the k-PCs in terms of similarity with Ref PCs as in previous works [15]: first, the k-PC having the smallest angle with Ref PC1 (α1), then the k-PC (from the three remaining ones) having the smallest angle with Ref PC2 (α2), and so on. The highest of the angles α1 to α4 was considered as an indicator of similarity (α_max) between the original set of Ref PCs and the new set of k-PCs. Then, after all the iterations (m=N+1), the ADL that provided the greatest similarity for the synergies, was removed, i.e., the one having the smallest α_max. This iteration was repeated until only one ADL remained (k=19). For each subject, the vector of ADLs contains the ADLs ordered from least to greatest effect on Ref PCs, if removed.
Step 3 - Selection of ADLs per subject
For each subject, the cut in the order removal was set at the maximum number of ADLs removed that still accomplish the following criteria: (1) the range (calculated as p95-p5) of the scores of each k-PC should be equal or higher than 85% of the range of the scores of the corresponding Ref PCs, and (2) the maximum angle between each four k-PCs and Ref PCs should be equal or lower than 30 degrees.
Step 4 - Selection of the reduced set of ADLs for hand function assessment
From the minimum set of ADLs of each subject, a selection of the reduced set of ADLs was performed. The ADL selected were the ones that were present in more than half of subjects, i.e. in more than 13 subjects and are the proposed ones for the BE-UJI set
Step 5 - Verification of main kinematic parameters on two different populations: healthy subjects and patients with HOA
Finally, motion strategies and main kinematic parameters such as range of motion and velocities of joints were compared between the 20 ADLs dataset and the BE-UJI set of ADLs, first, considering the same dataset used for the ADL reduction, and last, considering a dataset of 33 subjects with HOA. For both populations, healthy and HOA subjects, the same methodology was followed: first, for each subject, a PCA was performed by considering the BE-UJI set of ADLs (set PCs). Set PCs were reordered as in step 2 (grouping the set PCs more similar with ref PCs) and motion strategies of each subject were compared by means of the level of similarity (angle) between Ref PCs and set PCs. Secondly, range of motion of joints and velocities were obtained from all ADLs and from the BE-UJI set of ADLs, and were compared by means of box and whisker graphs.