Explainable AI Elucidates Musculoskeletal Biomechanics: A Case Study Using Wrist Surgeries.

As datasets increase in size and complexity, biomechanists have turned to artificial intelligence (AI) to aid their analyses. This paper explores how explainable AI (XAI) can enhance the interpretability of biomechanics data derived from musculoskeletal simulations. We use machine learning to classify the simulated lateral pinch data as belonging to models with healthy or one of two types of surgically altered wrists. This simulation-based classification task is analogous to using biomechanical movement and force data to clinically diagnose a pathological state. The XAI describes which musculoskeletal features best explain the classifications and, in turn, the pathological states, at both the local (individual decision) level and global (entire algorithm) level. We demonstrate that these descriptions agree with assessments in the literature and additionally identify the blind spots that can be missed with traditional statistical techniques.

Sensitivity Analysis of Upper Limb Musculoskeletal Models During Isometric and Isokinetic Tasks.

Sensitivity coefficients are used to understand how errors in subject-specific musculoskeletal model parameters influence model predictions. Previous sensitivity studies in the lower limb calculated sensitivity using perturbations that do not fully represent the diversity of the population. Hence, the present study performs sensitivity analysis in the upper limb using a large synthetic dataset to capture greater physiological diversity. The large dataset (n = 401 synthetic subjects) was created by adjusting maximum isometric force, optimal fiber length, pennation angle, and bone mass to induce atrophy, hypertrophy, osteoporosis, and osteopetrosis in two upper limb musculoskeletal models. Simulations of three isometric and two isokinetic upper limb tasks were performed using each synthetic subject to predict muscle activations. Sensitivity coefficients were calculated using three different methods (two point, linear regression, and sensitivity functions) to understand how changes in Hill-type parameters influenced predicted muscle activations. The sensitivity coefficient methods were then compared by evaluating how well the coefficients accounted for measurement uncertainty. This was done by using the sensitivity coefficients to predict the range of muscle activations given known errors in measuring musculoskeletal parameters from medical imaging. Sensitivity functions were found to best account for measurement uncertainty. Simulated muscle activations were most sensitive to optimal fiber length and maximum isometric force during upper limb tasks. Importantly, the level of sensitivity was muscle and task dependent. These findings provide a foundation for how large synthetic datasets can be applied to capture physiologically diverse populations and understand how model parameters influence predictions.

The Center-Center Image Closely Approximates Other Methods for Syndesmosis Reduction Clamp Placement.

BACKGROUND: The optimal placement for a syndesmosis reduction clamp remains an open question. This study compared the center-center axis, which localizes clamp placement using only an internally rotated lateral ankle X-ray, with other common approaches, whose accuracy can only be confirmed using computed tomography (CT). METHODS: Bone models of anatomically aligned (n = 6) and malreduced (n = 48) limbs were generated from CT scans of cadaveric specimens. Four axes for guiding clamp placement (center-center, centroid, B2, and trans-syndesmotic) were then analyzed, using digitally reconstructed radiographs derived from the bone models. Each axis' location was defined using angle-height pairs that describe axis orientation along the full anatomical region where syndesmosis fixation occurs. RESULTS: In anatomically aligned limbs, the center-center axis was located on average (±95% CI [confidence interval]), 0.64° (±0.50°) internal rotation, 1.03° (±0.73°) internal rotation, and 2.09° (±7.29°) external rotation from the centroid, B2, and trans-syndesmotic axes, respectively. Fibular displacement altered the magnitude of limb rotation needed to identify the center-center axis. CONCLUSION: The center-center technique is a valid method that closely approximates previously described methods for syndesmosis clamp placement without using CT, and the magnitude of C-arm rotation needed to transition from a talar dome lateral to a center-center view may be a potential method for assessing syndesmosis reduction. LEVELS OF EVIDENCE: Level III: Retrospective comparative study.

Predictions of thumb, hand, and arm muscle parameters derived using force measurements of varying complexity and neural networks.

Subject-specific musculoskeletal models are a promising avenue for personalized healthcare. However, current methods for producing personalized models require dense, biomechanical datasets that include expensive and time-consuming physiological measurements. For personalized models to be clinically useful, we must be able to rapidly generate models from simple, easy to collect data. In this context, the objective of this paper is to evaluate if and how simple data, namely height/weight and pinch force data, can be used to achieve model personalization via machine learning. Using simulated lateral pinch force measurements from a synthetic population of 40,000 randomly generated subjects, we train neural networks to estimate four Hill-type muscle model parameters and bone density. We compare parameter estimates to the true parameters of 10,000 additional synthetic subjects. We also generate new personalized models using the parameter estimates and perform new lateral pinch simulations to compare predicted forces using these personalized models to those generated using a baseline model. We demonstrate that increasing force measurement complexity reduces the root-mean-square error in the majority of parameter estimates. Additionally, musculoskeletal models using neural network-based parameter estimates provide up to an 80% reduction in absolute error in simulated forces when compared to a generic model. Thus, easily obtained force measurements may be suitable for personalizing models of the thumb, although extending the method to more tasks and models involving other joints likely requires additional measurements.

Inverse distance weighting to rapidly generate large simulation datasets.

Obtaining large biomechanical datasets for machine learning is an ongoing challenge. Physics-based simulations offer one approach for generating large datasets, but many simulation methods, such as computed muscle control (CMC), are computationally costly. In contrast, interpolation methods, such as inverse distance weighting (IDW), are computationally fast. We examined whether IDW is a low-cost and accurate approach for interpolating muscle activations from CMC.IDW was evaluated using lateral pinch simulations in OpenSim. Simulated pinch data were organized into grids of varying sparsity (high, medium, and low density), where each grid point represented the muscle activations associated with a unique combination of mass and height of a young adult. For each grid, muscle activations were calculated via CMC and IDW for 108 random mass-height pairs that were not coincident with simulation grid vertices. We evaluated the interpolation errors from IDW for each grid, as well as the sensitivity of lateral pinch force to these errors. The root mean square error (RMSE) associated with interpolated muscle activations decreased with increasing grid density and never exceeded 4%. While CMC received a target thumb-tip force of 40 N, errors from the interpolated muscle activations never impacted the simulated force magnitude by more than 0.1 N. Furthermore, the computation time for CMC simulations averaged 4.22 core-minutes, while IDW averaged 0.95 core-seconds per mass-height pair.These results indicate IDW is a practical approach for rapidly estimating muscle activations from sparse CMC datasets. Future works could adapt our IDW approach to evaluate other tasks, biomechanical features, and/or populations.

Sim-to-real localization: Environment resilient deep ensemble learning for guided wave damage localization.

Guided ultrasonic wave localization systems use spatially distributed sensor arrays and wave propagation models to detect and locate damage across a structure. Environmental and operational conditions, such as temperature or stress variations, introduce uncertainty into guided wave data and reduce the effectiveness of these localization systems. These uncertainties cause the models used by each localization algorithm to fail to match with reality. This paper addresses this challenge with an ensemble deep neural network that is trained solely with simulated data. Relative to delay-and-sum and matched field processing strategies, this approach is demonstrated to be more robust to temperature variations in experimental data. As a result, this approach demonstrates superior accuracy with small numbers of sensors and greater resilience to spatially nonhomogeneous temperature variations over time.

Classifying muscle parameters with artificial neural networks and simulated lateral pinch data.

OBJECTIVE: Hill-type muscle models are widely employed in simulations of human movement. Yet, the parameters underlying these models are difficult or impossible to measure in vivo. Prior studies demonstrate that Hill-type muscle parameters are encoded within dynamometric data. But, a generalizable approach for estimating these parameters from dynamometric data has not been realized. We aimed to leverage musculoskeletal models and artificial neural networks to classify one Hill-type muscle parameter (maximum isometric force) from easily measurable dynamometric data (simulated lateral pinch force). We tested two neural networks (feedforward and long short-term memory) to identify if accounting for dynamic behavior improved accuracy. METHODS: We generated four datasets via forward dynamics, each with increasing complexity from adjustments to more muscles. Simulations were grouped and evaluated to show how varying the maximum isometric force of thumb muscles affects lateral pinch force. Both neural networks classified these groups from lateral pinch force alone. RESULTS: Both neural networks achieved accuracies above 80% for datasets which varied only the flexor pollicis longus and/or the abductor pollicis longus. The inclusion of muscles with redundant functions dropped model accuracies to below 30%. While both neural networks were consistently more accurate than random guess, the long short-term memory model was not consistently more accurate than the feedforward model. CONCLUSION: Our investigations demonstrate that artificial neural networks provide an inexpensive, data-driven approach for approximating Hill-type muscle-tendon parameters from easily measurable data. However, muscles of redundant function or of little impact to force production make parameter classification more challenging.

A SSTDR Methodology, Implementations, and Challenges.

Sequence time-domain reflectometry (STDR) and spread spectrum time-domain reflectometry (SSTDR) detect, locate, and diagnose faults in live (energized) electrical systems. In this paper, we survey the present SSTDR literature for discussions on theory, algorithms used in its analysis, and its more prominent implementations and applications. Our review includes both scientific litera-ture and selected patents. We also discuss future applications of SSTDR.

Physiology-Informed Real-Time Mean Arterial Blood Pressure Learning and Prediction for Septic Patients Receiving Norepinephrine.

OBJECTIVE: Septic shock is a life-threatening manifestation of infection with a mortality of 20-50% [1]. A catecholamine vasopressor, norepinephrine (NE), is widely used to treat septic shock primarily by increasing blood pressure. For this reason, future blood pressure knowledge is invaluable for properly controlling NE infusion rates in septic patients. However, recent machine learning and data-driven methods often treat the physiological effects of NE as a black box. In this paper, a real-time, physiology-informed human mean arterial blood pressure model for septic shock patients undergoing NE infusion is studied. METHODS: Our methods combine learning theory, adaptive filter theory, and physiology. We learn least mean square adaptive filters to predict three physiological parameters (heart rate, pulse pressure, and the product of total arterial compliance and arterial resistance) from previous data and previous NE infusion rate. These predictions are combined according to a physiology model to predict future mean arterial blood pressure. RESULTS: Our model successfully forecasts mean arterial blood pressure on 30 septic patients from two databases. Specifically, we predict mean arterial blood pressure 3.33 minutes to 20 minutes into the future with a root mean square error from 3.56 mmHg to 6.22 mmHg. Additionally, we compare the computational cost of different models and discover a correlation between learned NE response models and a patient's SOFA score. CONCLUSION: Our approach advances our capability to predict the effects of changing NE infusion rates in septic patients. SIGNIFICANCE: More accurately predicted MAP can lessen clinicians' workload and reduce error in NE titration.

Sim-to-Real: Employing ultrasonic guided wave digital surrogates and transfer learning for damage visualization.

Wavefield imaging is a powerful visualization tool in nondestructive evaluation for studying ultrasonic wave propagation and its interactions with damage. To isolate and study damage scattering, damage-free baseline data is often subtracted from a wavefield. This is often necessary because the damage wavefield can be orders of magnitude weaker than the incident waves. Yet, baselines are not always accessible. When the baselines are accessible, the experimental conditions for the baseline and test data must be extremely similar. Researchers have created several baseline-free approaches for isolating damage wavefields, but these often rely on specific experimental setups. In this paper, we discuss a flexible approach based on ultrasonic guided wave digital surrogates (i.e., numerical simulations of incident waves) and transfer learning. We demonstrate this approach with two setups. We first isolate reflections from a circular, 2 mm diameter half-thickness hole on a 10 × 10 cm steel plate. We then isolate 8 circular, half-thickness holes of various diameters from 1 mm to 40 mm on a 60 × 60 cm steel plate. The second plate has a non-square geometry and the data has multi-path reflections. With both data sets, we isolate damage reflections without explicit experimental baselines. We also briefly illustrate the comparison of our dictionary learning method with wavenumber filtering technique which is often used to enhance the defect wavefields.

Reduced Rank Least Squares for Real-Time Short Term Estimation of Mean Arterial Blood Pressure in Septic Patients Receiving Norepinephrine.

Norepinephrine (NE), an endogenous catecholamine, is a mainstay treatment for septic shock, which is a life-threatening manifestation of severe infection. NE counteracts the loss in blood pressure associated with septic shock. However, an NE infusion that is too low fails to counteract the blood pressure drop, and an NE infusion that is too high can cause a hypertensive crisis and heart attack. Ideally, the NE infusion rate should maintain a patient's mean arterial blood pressure (MAP) above 65 mmHg. There are a few data-driven, quantitative models to predict the MAP, and incorporate NE effects. This paper presents a model, driven by intensive care unit (ICU) measurable data and known NE inputs, to predict the future MAP of an ICU patient. We derive a least square estimation model for MAP based on available ICU data, including heart period, NE infusion rate, and respiration wave. We learn the parameters of our model from initial patient data and then use this information to predict future MAP data. We assess our model with data from 12 septic patients. Our model successfully predicts and tracks MAP when the NE infusion rate changes. Specifically, we predict MAP 3 to 20 min in the future with the mean error of less than 4 to 7 mmHg over 12 patients. Conclusion: this new approach creates the potential to advance methods for predicting NE infusion rate in septic patients. Significance: successfully predicted patients' MAP could reduce catastrophic human error and lessen clinicians' workload.

Sparse Wavenumber Recovery and Prediction of Anisotropic Guided Waves in Composites: A Comparative Study.

Guided wave methodologies are among the established approaches for structural health monitoring (SHM). For guided wave data, being able to accurately estimate wave properties in the absence of ample measurements can greatly facilitate the often time-consuming and potentially expensive data acquisition procedure. Nevertheless, inherent complexities of the guided waves, including their multimodal and frequency dispersive nature, hinder processing, analysis, and behavior prediction. The severity of these complexities is even higher in anisotropic media, such as composites. Several methods, including sparse wavenumber analysis (SWA), have been proposed in the literature to characterize guided wave propagation by extracting wave characteristics in a particular medium from the information contained in a few measurements, and subsequently using this information for full wavefield prediction. In this paper, we investigate the efficacy of guided wave reconstruction techniques, based on SWA, for predicting the behavior of guided waves in composite materials. We implement these techniques on several experimental and simulation data sets. We study their performance in estimating the frequency-dependent (dispersive) and anisotropic velocities of guided waves and in reconstructing full wavefields from limited available information.

Baseline-free guided wave damage detection with surrogate data and dictionary learning.

In guided wave structural health monitoring, damage detection is often accomplished by comparing measurements before damage (i.e., baseline data) and after damage (i.e., test data). Yet, in practical scenarios, baseline data is often unavailable. Data from surrogate structures (structures similar to the test structure) could replace baseline data, but due to small differences in material properties, such as thickness, temperature, and other effects, this data is often unreliable. In this paper, a dictionary learning framework overcomes this challenge and detects damage with surrogate information. The framework combines wave propagation characteristics of a test structure with geometric information from surrogate structures to create a synthetic damage-free baseline. The test data is compared with the synthetic baseline to detect damage. The framework is evaluated with four 108 mm ×108 mm plates: two 1.6-mm thick aluminum plates, one 1.6-mm thick steel plate, and one 6.25 mm thick aluminum plate. The framework is applied to each test structure after learning from plates with different material properties and thicknesses. With full wavefield data, this paper successfully isolates reflections from a mass without using explicit baseline data. Furthermore, with sparse guided wave data, this paper shows that a drop in a correlation coefficient can detect a mass without using explicit baseline data.

Dynamic Time Warping Temperature Compensation for Guided Wave Structural Health Monitoring.

Guided wave structural health monitoring is widely researched for remotely inspecting large structural areas. To detect, locate, and characterize damage, guided wave methods often compare data to a baseline signal. Yet, environmental variations create large differences between the baseline and the collected measurements. These variations hide damage signatures and cause false detection. Temperature compensation algorithms, such as baseline signal stretch and the scale transform have been used to optimally realign data to a baseline. While these methods are effective in some conditions, their performance deteriorates in the presence of large temperature variations, long propagation distances, and high frequencies. In this paper, we use dynamic time warping to better align guided wave data and to overcome these errors. When compared with stretch-based methods, we show that the dynamic time warping is more robust to the above-mentioned errors and more accurately detects damage with weak ultrasonic signatures.

Reconstruction of Lamb wave dispersion curves by sparse representation with continuity constraints.

Ultrasonic Lamb waves are a widely used research tool for nondestructive structural health monitoring. They travel long distances with little attenuation, enabling the interrogation of large areas. To analyze Lamb wave propagation data, it is often important to know precisely how they propagate. Yet, since wave propagation is influenced by many factors, including material properties, temperature, and other varying conditions, acquiring this knowledge is a significant challenge. In prior work, this information has been recovered by reconstructing Lamb wave dispersion curves with sparse wavenumber analysis. While effective, sparse wavenumber analysis requires a large number of sensors and is sensitive to noise in the data. In this paper, it extended and significantly improved by constraining the reconstructed dispersion curves to be continuous across frequencies. To enforce this constraint, it is included explicitly in a sparse optimization formulation, and by including in the reconstruction an edge detection step to remove outliers, and by using variational Bayesian Gaussian mixture models to predict missing values. The method is validated with simulation and experimental data. Significant improved performance is demonstrated over the original sparse wavenumber analysis approach in reconstructing the dispersion curves, in synthesizing noise-removed signals, in reducing the number of measurements, and in localizing damage.

Data-driven and calibration-free Lamb wave source localization with sparse sensor arrays.

Most Lamb wave localization techniques require that we know the wave's velocity characteristics; yet, in many practical scenarios, velocity estimates can be challenging to acquire, are unavailable, or are unreliable because of the complexity of Lamb waves. As a result, there is a significant need for new methods that can reduce a system's reliance on a priori velocity information. This paper addresses this challenge through two novel source localization methods designed for sparse sensor arrays in isotropic media. Both methods exploit the fundamental sparse structure of a Lamb wave's frequency-wavenumber representation. The first method uses sparse recovery techniques to extract velocities from calibration data. The second method uses kurtosis and the support earth mover's distance to measure the sparseness of a Lamb wave's approximate frequency-wavenumber representation. These measures are then used to locate acoustic sources with no prior calibration data. We experimentally study each method with a collection of acoustic emission data measured from a 1.22 m by 1.22 m isotropic aluminum plate. We show that both methods can achieve less than 1 cm localization error and have less systematic error than traditional time-of-arrival localization methods.

Robust ultrasonic damage detection under complex environmental conditions using singular value decomposition.

Guided wave ultrasonics is an attractive monitoring technique for damage diagnosis in large-scale plate and pipe structures. Damage can be detected by comparing incoming records with baseline records collected on intact structure. However, during long-term monitoring, environmental and operational conditions often vary significantly and produce large changes in the ultrasonic signals, thereby challenging the baseline comparison based damage detection. Researchers developed temperature compensation methods to eliminate the effects of temperature variation, but they have limitations in practical implementations. In this paper, we develop a robust damage detection method based on singular value decomposition (SVD). We show that the orthogonality of singular vectors ensures that the effect of damage and that of environmental and operational variations are separated into different singular vectors. We report on our field ultrasonic monitoring of a 273.05 mm outer diameter pipe segment, which belongs to a hot water piping system in continuous operation. We demonstrate the efficacy of our method on experimental pitch-catch records collected during seven months. We show that our method accurately detects the presence of a mass scatterer, and is robust to the environmental and operational variations exhibited in the practical system.

Dispersion curve recovery with orthogonal matching pursuit.

Dispersion curves characterize many propagation mediums. When known, many methods use these curves to analyze waves. Yet, in many scenarios, their exact values are unknown due to material and environmental uncertainty. This paper presents a fast implementation of sparse wavenumber analysis, a method for recovering dispersion curves from data. This approach, based on orthogonal matching pursuit, is compared with a prior implementation, based on basis pursuit denoising. In the results, orthogonal matching pursuit provides two to three orders of magnitude improvement in speed and a small average reduction in prediction capability. The analysis is demonstrated across multiple scenarios and parameters.

Data-driven matched field processing for Lamb wave structural health monitoring.

Matched field processing is a model-based framework for localizing targets in complex propagation environments. In underwater acoustics, it has been extensively studied for improving localization performance in multimodal and multipath media. For guided wave structural health monitoring problems, matched field processing has not been widely applied but is an attractive option for damage localization due to equally complex propagation environments. Although effective, matched field processing is often challenging to implement because it requires accurate models of the propagation environment, and the optimization methods used to generate these models are often unreliable and computationally expensive. To address these obstacles, this paper introduces data-driven matched field processing, a framework to build models of multimodal propagation environments directly from measured data, and then use these models for localization. This paper presents the data-driven framework, analyzes its behavior under unmodeled multipath interference, and demonstrates its localization performance by distinguishing two nearby scatterers from experimental measurements of an aluminum plate. Compared with delay-based models that are commonly used in structural health monitoring, the data-driven matched field processing framework is shown to successfully localize two nearby scatterers with significantly smaller localization errors and finer resolutions.

Sparse recovery of the multimodal and dispersive characteristics of Lamb waves.

Guided waves in plates, known as Lamb waves, are characterized by complex, multimodal, and frequency dispersive wave propagation, which distort signals and make their analysis difficult. Estimating these multimodal and dispersive characteristics from experimental data becomes a difficult, underdetermined inverse problem. To accurately and robustly recover these multimodal and dispersive properties, this paper presents a methodology referred to as sparse wavenumber analysis based on sparse recovery methods. By utilizing a general model for Lamb waves, waves propagating in a plate structure, and robust l1 optimization strategies, sparse wavenumber analysis accurately recovers the Lamb wave's frequency-wavenumber representation with a limited number of surface mounted transducers. This is demonstrated with both simulated and experimental data in the presence of multipath reflections. With accurate frequency-wavenumber representations, sparse wavenumber synthesis is then used to accurately remove multipath interference in each measurement and predict the responses between arbitrary points on a plate.