Comparison of Machine Learning Techniques for Activities of Daily Living Classification with Electromyographic Data.

Advances in data science and wearable robotic devices present an opportunity to improve rehabilitation outcomes. Some of these devices incorporate electromyography (EMG) electrodes that sense physiological patient activity, making it possible to develop rehabilitation systems able to assess the patient's progress when performing activities of daily living (ADLs). However, additional research is needed to improve the ability to interpret EMG signals. To address this issue, an off-line classification approach for the 26 upper-limb ADLs included in the KIN-MUS UJI dataset is presented in this paper. The ADLs were performed by 22 subjects, while seven EMG signals were recorded from their forearms. From variable-length EMG time windows, 18 features were computed, and 13 features more were extracted from frequency domain windows. The classification performance of five different machine learning techniques, including Support Vector Machines (SVM), Convolutional Neural Networks (CNN), Gated Recurrent Unit (GRU) network, XGBoost, and Random Forests, were compared. CNN performed best amongst individual models, with an accuracy above 80%, compared to SVM with 77%, GRU with 73.9%, and the tree-based models below 64%. Ensemble learning with four CNN models achieved an even higher accuracy of 86%. These results suggest that the CNN ensemble model is capable of classifying EMG signals for most ADLs, which could be used in off-line quantitative assessment of robotic rehabilitation outcomes.

Machine learning and wearable sensors at preoperative assessments: Functional recovery prediction to set realistic expectations for knee replacements.

Unmet expectations contribute to a high patient dissatisfaction rate following total knee replacement but clinicians currently do not have the tools to confidently adjust expectations. In this study, supervised machine learning was applied to multi-variate wearable sensor data from preoperative timed-up-and-go tests. Participants (n=82) were instrumented three months after surgery and patients showing relevant improvement were designated as "responders" while the remainder were labelled "maintainers". Support vector machine, naïve Bayes, and random forest binary classifiers were developed to distinguish patients using sensor-derived features. Accuracy, sensitivity, specificity, and area under the receiver-operator curve (AUC) were compared between models using ten-fold out-of-sample testing. A high performance using only sensor-derived functional metrics was obtained with a random forest model (accuracy = 0.76 ± 0.11, sensitivity = 0.87 ± 0.08, specificity = 0.57 ± 0.26, AUC = 0.80 ± 0.14) but highly sensitive models were observed using naïve Bayes and SVM models after including patient age, sex, and BMI into the feature set (accuracy = 0.72, 0.73 ± 0.09, 0.12; sensitivity = 0.94, 0.95 ± 0.11, 0.11; specificity = 0.35, 0.37 ± 0.20, 0.18; AUC = 0.80, 0.74 ± 0.07, 0.11; respectfully). Including select patient-reported subjective measures increased the top random forest performance slightly (accuracy = 0.80 ± 0.10, sensitivity = 0.91 ± 0.14, specificity = 0.62 ± 0.23, AUC = 0.86 ± 0.09). The current work has demonstrated that prediction models developed from preoperative sensor-derived functional metrics can reliably predict expected functional recovery following surgery and this can be used by clinicians to help set realistic patient expectations.

Machine Learning Groups Patients by Early Functional Improvement Likelihood Based on Wearable Sensor Instrumented Preoperative Timed-Up-and-Go Tests.

BACKGROUND: Wearable sensors permit efficient data collection and unobtrusive systems can be used for instrumenting knee patients for objective assessment. Machine learning can be leveraged to parse the abundant information these systems provide and segment patients into relevant groups without specifying group membership criteria. The objective of this study is to examine functional parameters influencing favorable recovery outcomes by separating patients into functional groups and tracking them through clinical follow-ups. METHODS: Patients undergoing primary unilateral total knee arthroplasty (n = 68) completed instrumented timed-up-and-go tests preoperatively and at their 2-, 6-, and 12-week follow-up appointments. A custom wearable system extracted 55 metrics for analysis and a K-means algorithm separated patients into functionally distinguished groups based on the derived features. These groups were analyzed to determine which metrics differentiated most and how each cluster improved during early recovery. RESULTS: Patients separated into 2 clusters (n = 46 and n = 22) with significantly different test completion times (12.6 s vs 21.6 s, P < .001). Tracking the recovery of both groups to their 12-week follow-ups revealed 64% of one group improved their function while 63% of the other maintained preoperative function. The higher improvement group shortened their test times by 4.94 s, (P = .005) showing faster recovery while the other group did not improve above a minimally important clinical difference (0.87 s, P = .07). Features with the largest effect size between groups were distinguished as important functional parameters. CONCLUSION: This work supports using wearable sensors to instrument functional tests during clinical visits and using machine learning to parse complex patterns to reveal clinically relevant parameters.

Proposal and Validation of a Knee Measurement System for Patients With Osteoarthritis.

OBJECTIVE: Currently most measurements of knee joint function are obtained through observation and patient-reported outcomes. This paper proposes an implementation and validation of a knee monitor to measure quantitative joint data in multiple degrees of freedom. The proposed system is configurable with minimal patient interaction and no frame-alignment calibration procedure is required for measurement after visually placing/replacing sensors on patients. METHODS: A mobile software system was developed using a method of extracting clinical knee angles based on attitude estimations from independent wearable sensors. Validation was performed using a robot phantom and results were compared with a gold standard motion capture system. Two instrumentation placements (lateral and posterior) were examined. RESULTS: A posterior sensor placement was determined to provide the most repeatable results through multiple degrees of freedom and measurement accuracy approached a gold standard motion capture technology with low root-mean-square error (flexion: 3.34°, internal/external rotation: 2.18°, and varus/valgus: 1.44°). CONCLUSION: The proposed system is simple to use and convenient for use in ambulatory or unsupervised environments for joint measurement; however, it was shown that accuracy can be sensitive to sensor placement. SIGNIFICANCE: This system would be beneficial for obtaining quantitative patient data or tracking functional activity in variable environments, providing clinicians with indications of how patients' knees function during activity, potentially permitting more individualized care and recommendations.

Quantification of whole-body bradykinesia in Parkinson's disease participants using multiple inertial sensors.

Bradykinesia (slowness of movement) is a common motor symptom of Parkinson's disease (PD) that can severely affect quality of life for those living with the disease. Assessment and treatment of PD motor symptoms largely depends on clinical scales such as the Unified Parkinson's Disease Rating Scale (UPDRS). However, such clinical scales rely on the visual assessment by a human observer, naturally resulting in inter-rater variability. Although previous studies have developed objective means for measuring bradykinesia in PD patients, their evaluation was restricted by the type of movement and number of joints assessed. These studies failed to provide a more comprehensive, whole-body evaluation capable of measuring multiple joints simultaneously. This study utilizes wearable inertial measurement units (IMUs) to quantify whole-body movements, providing novel bradykinesia indices for walking (WBI) and standing up from a chair (sit-to-stand; SBI). The proposed bradykinesia indices include the joint angles at both upper and lower limbs and trunk motion to compute a complete, objective score for whole body bradykinesia. Thirty PD and 11 age-matched healthy control participants were recruited for the study. The participants performed two standard walking tasks that involved multiple body joints in the upper and lower limbs. The WBI and SBI successfully identified differences between control and PD participants. The indices also effectively identified differences within the PD population, distinguishing participants assessed with (ON) and without (OFF) levodopa; the gold-standard of treatment for PD. The goal of this study is to provide health professionals with an objective score for whole body bradykinesia by simultaneously measuring the upper and lower extremities along with truncal movement. This method demonstrates potential to be used in conjunction with current clinical standards for motor symptom assessment, and may also be promising for the remote assessment of PD patients and in cases where experienced clinicians may not be available. In conclusion, the intelligent use of this technology for the measurement of bradykinesia (among other symptoms) has vast implications for optimizing treatment in Parkinson's disease, ultimately leading to an improvement in quality of life.

Towards remote monitoring of Parkinson's disease tremor using wearable motion capture systems.

The management of movement disorders is shifting from a centralized-clinical assessment towards remote monitoring and individualized therapy. While a variety of treatment options are available, ranging from pharmaceutical drugs to invasive neuromodulation, the clinical effects are inconsistent and often poorly measured. For instance, the lack of remote monitoring has been a major limitation to optimize therapeutic interventions for patients with Parkinson's Disease (PD). In this work, we focus on the assessment of full-body tremor as the most recognized PD symptom. Forty PD and twenty two healthy participants were recruited. The main assessment tool was an inertial measurement unit (IMU)-based motion capture system to quantify full-body tremor and to separate tremor-dominant from non-tremor-dominant PD patients as well as from healthy controls. We developed a new measure and evaluated its clinical utility by correlating the results with the Unified Parkinson's Disease Rating Scale (UPDRS) scores as the gold standard. Significant correlation was observed between the UPDRS and the tremor severity scores for the selected tasks. The results suggest that it is feasible and clinically meaningful to utilize the suggested objective tremor score for the assessment of PD patients. Furthermore, this portable assessment tool could potentially be used in the home environment to monitor PD tremor and facilitate optimizing therapeutic interventions.

Using Wearable Technology to Generate Objective Parkinson's Disease Dyskinesia Severity Score: Possibilities for Home Monitoring.

A variety of clinical scales are available to assess dyskinesia severity in Parkinson's disease patients; however, such assessments are subjective, do not provide long term monitoring, and their use is subject to inter- and intra-rater variability. In this paper, an objective dyskinesia score was developed using an IMU -based motion capture system. Deep brain stimulation (DBS) surgery is currently the only acute intervention that results in the rapidly progressive reduction of dyskinesia's severity; hence, this form of therapy was selected as a model to validate the proposed method. Thirteen Parkinson's disease participants undergoing DBS surgery and 12 age-matched healthy control participants were assessed using the motion capture system. Concurrent Unified Dyskinesia Rating Scale (UDysRS) ratings were also performed. Parkinson's disease participants were assessed pre-operatively and for five visits post-operatively while seated at rest, during arms outstretched and while performing an action task. The kinematic data were used to develop an objective measure defined as the dyskinesia severity score. Generally, a strong correlation was observed between the UDysRS ratings and the full-body dyskinesia severity scores. The results suggest that it is feasible and clinically meaningful to utilize an objective full-body dyskinesia score for the assessment of dyskinesia. The portable motion capture system along with the developed software can be used remotely to monitor the full-body severity of dyskinesia, necessary for therapeutic optimization, especially in the patients home environment.

Characterization of multi-joint upper limb movements in a single task to assess bradykinesia.

Bradykinesia is a disabling symptom of Parkinson's disease (PD) which presents with slowness of movement. Visual assessment using clinical rating scales is currently the gold standard to assess bradykinesia. Such assessments require multiple separate movements, are subjective, and rely on the ability of the rater to determine frequency and amplitude features of excursion of multiple joints simultaneously. The current study introduces the use of wearable inertial measurement units (IMUs) to characterize full-arm repetitive movements and provide a new index score for bradykinesia severity (BKI) in the upper limbs. The BKI provides an approach to measuring bradykinesia reliably and objectively. Importantly, this index is needed to demonstrate separability between healthy individuals and PD participants, and also between bradykinetic and non-bradykinetic PD participants. Thirteen PD participants and ten age-matched healthy control participants were studied. Using a single upper limb task that activated multiple joints and recordings from angular displacements from all joints, features relevant to demonstrating bradykinesia were extracted and systematically combined to create the total BKI. A strong correlation coefficient was obtained comparing BKI to upper limb UPDRS bradykinesia scores (rs=-0.626, p=0.001). The BKI successfully identified differences between control and PD participants (p=0.018). The BKI was also sensitive enough to identify differences within the PD population, separating PD participants with and without bradykinesia (p<0.001). This study demonstrates the feasibility of using IMU-based motion capture systems and employing the new BKI for quantitative assessment of bradykinesia. This approach when generalized to lower extremity and truncal movements would be able to provide an objective and reproducible whole body bradykinesia index.

Dynamic 3-D virtual fixtures for minimally invasive beating heart procedures.

Two-dimensional or 3-D visual guidance is often used for minimally invasive cardiac surgery and diagnosis. This visual guidance suffers from several drawbacks such as limited field of view, loss of signal from time to time, and in some cases, difficulty of interpretation. These limitations become more evident in beating-heart procedures when the surgeon has to perform a surgical procedure in the presence of heart motion. In this paper, we propose dynamic 3-D virtual fixtures (DVFs) to augment the visual guidance system with haptic feedback, to provide the surgeon with more helpful guidance by constraining the surgeon's hand motions thereby protecting sensitive structures. DVFs can be generated from preoperative dynamic magnetic resonance (MR) or computed tomograph (CT) images and then mapped to the patient during surgery. We have validated the feasibility of the proposed method on several simulated surgical tasks using a volunteer's cardiac image dataset. Validation results show that the integration of visual and haptic guidance can permit a user to perform surgical tasks more easily and with reduced error rate. We believe this is the first work presented in the field of virtual fixtures that explicitly considers heart motion.

A potential field model using generalized sigmoid functions.

The lack of a potential field model capable of providing accurate representations of objects of arbitrary shapes is considered one major limitation in applying the artificial potential field method in many practical applications. In this correspondence, we propose a potential function based on generalized sigmoid functions. The generalized sigmoid model can be constructed from combinations of implicit primitives or from sampled surface data. The constructed potential field model can achieve an accurate analytic description of objects in two or three dimensions and requires very modest computation at run time. In this correspondence, applications of the generalized sigmoid model in path-planning tasks for mobile robots and in haptic feedback tasks are presented. The validation results in this correspondence show that the model can effectively allow the user or mobile robot to avoid penetrations of obstacles while successfully accomplishing the task.

Rendering of virtual fixtures for MIS using generalized sigmoid functions.

To avoid undesired collisions and improve the level of safety and precision, artificial potential field (APF) can be employed to generate virtual forces around protected tissue and to provide surgeons with real-time force refection through haptic feedback. In this paper, we propose a potential field-based force model using the generalized sigmoid function, and show that it can represent a large class of shapes. The proposed approach has several advantages such as computational efficiency, easily adjustable level of force reflection, and force continuity.