Information Science and Technology, The University of Tokyo
Yoshihiko Nakamura
(field1- Program3)

We have been developing an “Inverse system” [1] , which is a technology for estimating information in the human body from dynamics and feedback [2][3] based on an efficient computation of a rigid body system that was developed in the field of robotics. That is, we are trying to take a look at the information in the body by extraneous measurement and computation.

We conducted a project known as CREST, “Creating the Brain” (Research Supervisor: Prof. Shunichi Amari), from 1998 to 2003, and started research to generate and control a body movement pattern of humanoid robots by attracting symbolic nonlinear dynamics. Meanwhile, we developed a parallel computing method called Assembly and Disassembly Algorithm, in which O (log N) is usable for calculation, as a direct dynamics computing method [4], and we took a cue from that and started to compute human systemic skeletal musculature as a physics model. In the model, 997 wires are used to link a systemic skeletal system with 155 degrees of freedom [5]. It became one of the major projects through two Grants-in-Aid for Scientific Research (S): “Deployment of Dynamics-Based Information Processing Model of Intelligence” (2003 to 2007), and “Establishing Human-Machine Communication through Kinesiology and Linguistic Integration” (2008 to 2012).

Estimation starts with calculating the force applied to the skeleton with mass (the entire mass of the human body is allocated to the skeleton) from geometric information about motion, which is measured by motion capture, by a calculation scheme of inverse dynamics. The next computation is back calculation of systemic muscular tension expressed in the wire model without mass. There is an inequality constraint that the wire tension shall be positive or zero. The solution of a problem having high redundancy, 155 < 997, is explored as a secondary programming problem. In order to take the activity of antagonist muscles into account, wireless electromyography is attached to the body during motion capture, and muscle activities of up to 32 channels are measured. In the optimization problem, the secondary programming problem is solved using examples from the activity (Fig. 1).

Then, our interest was directed towards the deeper part of the body and construction of an inverse model for estimating nervous activity from muscle activity, and we are now studying the motor portion of the nervous system from the spinal cord toward the periphery. The spinal nerve consists of total 31 pairs of peripheral nerves emerging from the spinal cord (8 pairs of cervical nerves, 12 pairs of thoracic spinal nerves, 5 pairs of lumbar nerves, 5 pairs of sacral nerves and a pair of coccygeal nerves). As each nerve branches into an anterior branch and dorsal branch, the spinal nerve branches into 124 nerve fascicles. The function of skeletal muscle, to which each nerve fascicle projects, was clarified. The 124 nerve fascicles are supposed to represent the same number of signal pathways for estimating their activities. We placed a totally-coupled mathematical neural network in the spinal cord, and made it learn by activity of the nerve fascicles and voluntary signals (in which various hypotheses are used) descending from the brain. The spinal neural network obtained this way is regard as an optimized nerve system specialized in the anatomical body model and movement pattern typically found in humans, and represents the reflex arc by the spinal cord (Fig. 2).

Strategic Program, Program 3: In a hierarchical integrated simulation for predictive medicine (represented by Shu Takagi), a high-dimensional systemic neuromusculoskeletal system and its parallel computation algorithms are being developed by creating a model, in which the mass is dispersed among muscles and viscera, and computing the stress distribution of contraction and contact associated with elastic deformation. By integrating the multi-scale model from the muscle fibers of skeletal muscle studied by the Takagi Group and the basal ganglion model by the Doya Group, we are now aiming, for example, to duplicate a mechanism for altered dopamine metabolism in basal ganglions in Parkinson disease as altered motion in walking and reaching. In this way, we are going to establish basic computational technologies for predictive medicine helpful in drug discovery, diagnosis and treatment.

Fig. 1: Visualization of muscle activity	Fig 2 : Composition of spinal cord reflex model

[References]