Pattern recognition algorithms for upper-limb prosthetics control

In the last 70 years, the evolution of technological and surgical techniques in the field of the upper limb prosthetic have leaded even more advanced solutions to address the future research towards the development of prostheses that are functional and able to mimic the lost upper limb behavior, replicating the performance of the human arm. To this purpose, nowadays, the surface electromyography (sEMG) signals represent a promising approach for decoding the motor intention of amputees with different amputation level. Several approaches based on proportional amplitude methods or simple thresholds on sEMG signals have been proposed, in literature, to control a single degree of freedom (DoF) at time, without the possibility of increasing the number of controllable multiple DoFs in a natural manner. To address this relevant issue, Pattern Recognition (PR) strategies have been proposed to reach a more natural and intuitive control of myoelectric prostheses, compared to the conventional myoelectric control methods. In detail, the major potentiality of the PR methods has been to add multiple DoFs by keeping low the number of electrodes and allowing the discrimination of different muscular patterns for each class of motion. However, the use of PR algorithms to simultaneously decode both gestures and forces has never been studied deeply. Also the simultaneous control of a poliarticulated prostheses with several DoFs, related to the elbow, wrist, and hand joints, has to be yet investigated deeply to ensure greater dexterity than the conventional control strategies. This is considered a needed capability to restore upper limb functionality, especially for transhumeral and shoulder disarticulation amputees who have undergone Targeted Muscle Reinnervation (TMR) surgery. This surgical technique has been considered innovative and relevant for improving, together with PR strategies, prosthetic control by adding the number of controllable muscular sites. Indeed, the ultimate goal of the TMR is to obtain reinnervated areas that act as biological amplifiers of the motor control. Indeed, the ultimate goal of this surgical technique is to obtain reinnervated areas that act as biological amplifiers of the motor control. Also the simultaneous control of a poliarticulated prostheses with several DoFs, related to the elbow, wrist, and hand, has to be yet investigated deeply to ensure greater dexterity than the conventional control strategies. In this scenario, the potentiality of future clinical application of TMR and PR control strategies in the control of multifunctional prostheses was investigated to add value to the current knowledge in the field of upper-limb prosthetics. In detail, this thesis aims at providing promising PR-based strategies for (i) controlling simultaneously, with a hierarchical classification strategy, the hand/wrist gestures and exerted forces during grasping tasks; (ii) discriminating, with a parallel classification strategy, 27 motion classes related to the elbow, hand, and wrist joints. In detail, the parallel classification strategy uses three joint classifiers, one for each DoF. If only one of the three joints is involved in the desired movement, the parallel PR strategy will output a 1 DoF motion class; if instead, two or three joints are activated in a complex motion tasks, the parallel classification strategy will output a 2 or 3 DoFs motion class. To reach the first objective, a hierarchical classification strategy was developed and validated on 31 healthy subjects and 15 transradial amputees, with the aim to discriminate seven hand/wrist gestures, as well as the desired three force levels to exert during grasping tasks. In detail, the results from healthy showed an average F1Score about equals to 96% for the hand/wrist gestures and equals to 98% for the force classifiers, with both the Non Linear Logistic Regression (NLR) and Linear Discriminant Analysis (LDA) classifiers. To evaluate the robustness of the hierarchical PR system, both offline and in real-time, a prosthetic system composed of a hand (RoboLimb) and a wrist module (WristRotator) was employed by trans-radial amputees when they manage simultaneously the desired hand/wrist gestures and the three force levels. In detail, the results from transradial amputees reached an average F1Score values equals to 90% for the hand/wrist gestures and equals to 96% up to 98% for the force classifiers, when considering the Logistic Regression (LR), NLR and LDA classifiers. Also considering the real-time performance metrics, the Mann-Whitney test (U-test) with Bonferroni correction points out no statistically significant difference between the three algorithms. To the second purpose, a parallel classification strategy was developed and validated on 15 healthy subjects, to provide the simultaneous classification of 27 discrete and combined motion classes, by keeping the number of electrodes to a bare minimum and the classification error rates under 10%. In detail, the discrete 1 DoF motion classes involved only one joint, while the combined 2 or 3 DoFs movements, provided the simultaneous activation of two or all the three joints. In detail, the mean F1Score values were above 90% for all the joint classifiers, with both the LR and LDA algorithms. About the real-time results, the performance of the LR algorithm were statistically better than that obtained with the LDA, despite this last was considered the benchmark classifier for real-time employment. Then, also an analysis of the preliminary offline and real-time results, obtained from a TMR subject, was carried out. In this case, the classification performance of the TMR patient reached lower mean F1Score values than that on healthy subjects, and they were about equals to 86% for LR and LDA classifiers. From this preliminary results, there wasn't a statistical significance difference between the performance obtained with the LR and LDA algorithms. In conclusion, this thesis aims to provide useful insights into the choice of the suitable classification strategies to discriminate simultaneously hand/wrist gestures and grasping forces and to classify complex tasks involving multiple joints. The outcomes of these objectives reveal that the use of non linear classification algorithm, as NLR, is as much suitable as the benchmark LDA classifier for implementing a hierarchical sEMG-based PR system, able both to decode hand/wrist gestures and to associate different performed force levels to grasping actions. This result is also more appreciable if we consider that we have tested this PR strategy in clinical practice, by employing a robotic hand and wrist module. Regarding the second purpose, the obtained results strongly encourage further investigation of the parallel classification strategy's performance for others TMR patients. In this way, an even more level of robustness and reliability of the proposed PR system can be reached to control simultaneously and in a natural way different joints of a complex multi-DoFs prosthetic device.