Figures and data
A general outline of the proposed approach.
(1.1(a-b)) We propose a novel approach to mapping muscle couplings to the task space. Among current muscle synergy analysis approaches, muscle couplings are quantified in isolation of the task solely using dimensionality reduction. Using our approach, the functional characteristics of muscle interactions can be quantified in terms of the similarity of their encoded task information. We do so by determining the coupling between [mx, my] and a corresponding task parameter (τ) using mutual information (MI). From this perspective, task-redundant muscle couplings (pink shaded area in pink-orange intersection) represent muscles cooperating towards similar task goals, while task-synergistic muscle couplings (orange shaded area in pink-orange intersection) encapsulate the task information provided by a muscle pairing acting towards complementary task goals. Muscle couplings present across tasks (i.e., task-irrelevant) are quantified by conditioning the MI between [mx, my] pairs with respect to τ (yellow intersection). (1.2) A description of redundant and synergistic interactions. (a) Net redundant interactions are defined by a greater amount of information generated by the sum of individual observation of mx and my ([mx+ my]) than their simultaneous observation ([mx, my]). (b) In a net synergistic interaction, [mx, my] provides more information than [mx+ my]. (1.3(a-c)) An overview of the approach. Spatiotemporal muscle activation samples are extracted across trials from large-scale EMG datasets and concatenated into vectors, forming [mx, my] pairs.The derived muscle couplings are then run through the NIF pipeline [11], producing low-dimensional, multiplexed space-time muscle networks.
A summary of the NIF pipeline. (A) Large-scale datasets of EMG signals are captured while participants perform various motor tasks [26–28]. (B) The MI between all unique muscle-timepoint vector ([mx, my]) combinations with respect to a corresponding task parameter (τ) is determined [41], forming a network of functional connectivities. (C) These adjacency matrices are then analysed in terms of statistical significance and modular structure using percolation theory [11]. (D) The optimal spatial and temporal model-ranks are determined using generalised, consensus-based network community detection methods [32–34,36]. (E) The optimal model-ranks are used as input parameters for dimensionality reduction, where space-time muscle networks along with their underlying activation coefficients are concurrently extracted [26].
Graphical illustrations of each of the datasets analysed in the current study. (A) Dataset 1 consisted of participants executing table-top point-to-point reaching movements across four targets in forward (P1-P4) and backwards (P5-P8) directions at both fast and slow speeds [26]. The muscles recorded included the finger extensors (FE), brachioradialis (BR), biceps brachii (BI), medial-triceps (TM), lateral-triceps (TL), anterior deltoid (AD), posterior deltoid (PD), pectoralis major (PE), latissimus dorsi (LD) of the right, reaching arm. (B) For dataset 2, the activity of 30 muscles was recorded while participants performed whole-body point-to-point reaching movements across three different heights and bars and in various directions, accumulating to 72 unique reaching tasks [27]. (C) The circuit navigated by participants in dataset 3 as they executed various locomotion modes is illustrated, of which level-ground walking, stair- and ramp-ascent/descent were analysed in the current study [28]. Several sub-conditions were undertaken by participants for each locomotion mode including different walking-speeds, clockwise vs. counter-clockwise direction, different stair heights and ramp inclines etc. Participants executed these tasks while the EMG of 11 muscles on the right leg ((Gluteus medius (GlutM), right external oblique (Obl), semitendinosus (ST), gracilis (GR), biceps femoris (BF), rectus femoris (RF), vastus lateralis (VL), vastus medialis (VM), soleus (SO), tibialis anterior (TA), gastrocnemius medialis (GM)) along with kinematic, dynamic and IMU signals were captured.
A simplified example output from the proposed framework applied to a single trial of turning gait from Dataset 3. (A) Task-irrelevant, (B) Task-redundant and (C) Task-synergistic synchronous muscle couplings were quantified with respect to the heel kinematic marker (anterior-posterior direction). Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Three spatial (S1-S3) and two temporal task-irrelevant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 2 using the NIF pipeline (Panel A-B) [11,27]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Three spatial (S1-S3) and two temporal task-irrelevant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 3 using the NIF pipeline (Panel A-B) [11,28]. Activation coefficients are presented in supplementary materials (fig.4), indicating their task parameter-specific scaling averaged across participants in the dynamic, IMU and kinematic spaces. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Three spatial (S1-S3) and two temporal task-redundant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 2 using the NIF pipeline (Panel A-B) [11,27]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Three spatial (S1-S3) and two temporal task-redundant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 3 using the NIF pipeline (A-B) [11,28]. Activation coefficients are presented in supplementary materials document 2 (fig.2), indicating their task parameter-specific scaling averaged across participants in the dynamic, IMU and kinematic spaces. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Three spatial (S1-S3) and two temporal task-synergistic muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 2 using the NIF pipeline (Panel A-B) [11,27]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
A summary table illustrating the findings from an examination of the generalisability of the muscle networks extracted from each dataset. The spatial and temporal representations extracted from the full input data in each muscle-task information subspace were compared using Pearson’s correlation against functionally similar representations extracted from a subset of the input data.
Three spatial (S1-S3) and two temporal task-synergistic muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 3 using the NIF pipeline (Panel A-B) [11,28]. Activation coefficients are presented in supplementary materials (fig.6), indicating their task parameter-specific scaling averaged across participants in the dynamic, IMU and kinematic spaces. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [33], network centrality and connection strength respectively [39,40].
Co-I determines the difference between the sum total information shared with τ in mx and my when observed separately and the information shared with τ when they are observed together. The adjacency matrices show how this calculation is carried out for all unique [mx, my] combinations. Redundant and synergistic muscle couplings are then separated into two equivalently sized networks. The accompanying colour bars indicate the values present in the adjacency matrix.
Three spatial (S1-S3) and two temporal task-irrelevant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 1 using the NIF pipeline (Panel A-B) [10,23]. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [30], network centrality and connection strength respectively [35,43]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants.
Three spatial (S1-S3) and two temporal task-redundant muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 1 using the NIF pipeline (Panel A-B) [10,23]. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [30], network centrality and connection strength respectively [35,43]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants.
Three spatial (S1-S3) and two temporal task-synergistic muscle networks (T1-T2) were empirically identified and extracted across participants and task parameters from dataset 1 using the NIF pipeline (Panel A-B) [10,23]. Human body models accompanying each spatial network illustrate their respective submodular structure with node colour and size and edge width indicating community affiliation [30], network centrality and connection strength respectively [35,43]. (Panel C) Activation coefficients are presented to the right of the networks, indicating their task parameter-specific scaling averaged across participants.
Task-irrelevant activation coefficients (Dataset 3) [25]. Dynamic, inertial motion unit (IMU) and kinematic data were captured from the bilateral lower-limbs while 17 participants performed various locomotion modes (i.e. stair ascents/descents, ramp inclines/declines and level-ground walking). Activation coefficients are averaged across participants.
Task-redundant activation coefficients (Dataset 3) [25]. Dynamic, inertial motion unit (IMU) and kinematic data were captured from the bilateral lower-limbs while 17 participants performed various locomotion modes (i.e. stair ascents/descents, ramp inclines/declines and level-ground walking). Activation coefficients are averaged across participants.
Task-synergistic activation coefficients (Dataset 3) [25]. Dynamic, inertial motion unit (IMU) and kinematic data were captured from the bilateral lower-limbs while 17 participants performed various locomotion modes (i.e. stair ascents/descents, ramp inclines/declines and level-ground walking). Activation coefficients are averaged across participants.
