A DATA-DRIVEN OPTIMIZATION METHOD TO DETERMINE MUSCLE- TENDON PATHS OF THE INDEX FINGER 1 Jong Hwa Lee, 2 Deanna S. Asakawa, 1 Cecil A. Lozano, 3 Jack T. Dennerlein and 2 Devin L. Jindrich 1 Arizona State University, Tempe, AZ, USA 2 California State University, San Marcos, CA, USA 3 Northeastern University, Boston, MA, USA email: djindrich@csusm.edu web: www.limblab.org INTRODUCTION A musculoskeletal model of the human index finger can provide a better understanding of hand dynamics and control during multi-touch gestures used on hand-held touchscreen computing devices. As a first step toward a musculoskeletal model capable of describing the complex, multi-finger gestures used for touchscreen devices, we seek to incorporate intrinsic hand muscles into an existing musculoskeletal model of the human arm [5]. Although several studies have measured hand muscle moment arms in vivo and in situ [1-3], measurements of moment arms alone are insufficient for homeomorphic models like OpenSim [4]. Moment arms are indeterminate: many combinations of muscle- tendon paths could result in the same moment arm. We therefore developed techniques for data-driven optimization to determine attachments that reproduce measured moment arms for intrinsic and extrinsic muscles and tendons of the index finger. METHODS We added the following muscles to an upper- extremity model implemented in OpenSim [5]: terminal extensor (TE), extensor slip (ES), radial band (RB), ulnar band (UB), first dorsal interosseous or radial interosseous (RI), lumbricals (LU), and first palmar interosseous or ulnar interosseous (UI). Extensor indicis (EI) was omitted because the moment arms of EI and EDC (extensor digitorum communis) muscles were identical. Experimentally-measured values [1-3] were used as target moment arm relationships for all seven muscles at metacarpophalangeal (MCP), proximal interphalangeal (PIP), and distal interphalangeal (DIP) joints. We normalized moment arms by the length of middle phalanx for DIP and PIP joints, and by the MCP joint’s thickness and width for the MCP joint. We employed a simulated annealing [6] algorithm to optimize muscle attachment points. The objective function, f(x ) was defined as the root mean square (RMS) error between the experimentally-derived moment arms, r ! (q ! ) and the modeled-estimate moment arms, r ! (q ! , x ) as follows: Minimize   = ! ! ! ! !! ! (! ! ,! ) ! ! ! !!! Subject to  ! ≤  ! ≤  ! g ! x − ε ! ≤ 0 Where, x was 6×1 vector (described as x,y,z origin and x,y,z insertion points) to be optimized, q ! was the joint angle with respect to finger motions ( ), and  was each individual muscle. Boundary conditions constrained the path of muscle from violating a feasible region, expanded from bony segment (as a lower bound: lb ! ) to external hand dimensions (as an upper bound: ub ! ). ε ! was the maximum variations of experimental moment arms, and g ! x was RMS error of ab/adduction moment arms during flexion/extension movements. The inequality constraint imposed the optimal pathway on producing realistic ab/adduction moment arms during flexion/extension simulation. This inequality constraint g ! x , ε ! acted as a weight function; ε ! for the extrinsic tendons was a large number so the attachment points were less influenced by ab/adduction motions, while ε ! for the intrinsic muscles was a small number so these muscles were more influenced by ab/adduction movements. RESULTS AND DISCUSSION The data-driven optimization approach was able to determine the attachment points of extrinsic and intrinsic muscles, resulting in moment arms that