Explaining Machine Learning Models for Clinical Gait Analysis

Machine Learning (ML) is increasingly used to support decision-making in the healthcare sector. While ML approaches provide promising results with regard to their classification performance, most share a central limitation, their black-box character. This article investigates the usefulness of Explainable Artificial Intelligence (XAI) methods to increase transparency in automated clinical gait classification based on time series. For this purpose, predictions of state-of-the-art classification methods are explained with a XAI method called Layer-wise Relevance Propagation (LRP). Our main contribution is an approach that explains class-specific characteristics learned by ML models that are trained for gait classification. We investigate several gait classification tasks and employ different classification methods, i.e., Convolutional Neural Network, Support Vector Machine, and Multi-layer Perceptron. We propose to evaluate the obtained explanations with two complementary approaches: a statistical analysis of the underlying data using Statistical Parametric Mapping and a qualitative evaluation by two clinical experts. A gait dataset comprising ground reaction force measurements from 132 patients with different lower-body gait disorders and 62 healthy controls is utilized. Our experiments show that explanations obtained by LRP exhibit promising statistical properties concerning inter-class discriminativity and are also in line with clinically relevant biomechanical gait characteristics.

Short-term results of gait analysis with the Heidelberg foot measurement method and functional outcome after operative treatment of ankle fractures

Background Ankle fractures are common fractures in trauma surgery. Several studies have compared gait patterns between affected patients and control groups. However, no one used the Heidelberg Foot Measurement Method in combination with statistical parametric mapping of the entire gait cycle in this patient cohort. We sought to identify possible mobility deficits in the tibio-talar joint and medial arch in patients after ankle fractures as a sign of stiffness and pain that could result in a pathological gait pattern. We focused on the tibio-talar flexion as it is the main movement in the tibio-talar joint. Moreover, we examined the healing progress over time. Methods Fourteen patients with isolated ankle fractures were included prospectively. A gait analysis using the Heidelberg Foot Measurement Method was performed 9 and 26 weeks after surgery to analyse the tibio-talar dorsal flexion, the foot tibia dorsal flexion, the subtalar inversion and the medial arch as well as the cadence, the walking speed and the ground reaction force. The American Orthopedic Foot & Ankle Society ankle hindfoot score was used to obtain clinical data. Results were compared to those from 20 healthy participants. Furthermore, correlations between the American Orthopedic Foot & Ankle Society hindfoot score and the results of the gait analysis were evaluated. Results Statistical parametric mapping showed significant differences for the Foot Tibia Dorsal Flexion for patients after 9 weeks (53–75%: p = 0.001) and patients after 26 weeks (58–70%: p = 0.011) compared to healthy participants, respectively. Furthermore, significant differences regarding the tibio-talar dorsal flexion for patients 9 weeks after surgery (15–40%: p < 0.001; 56,5–70%: p = 0.007; 82–88%: p = 0.033; 97–98,5%: p = 0.048) as well as patients after 26 weeks (62,5–65%: p = 0.049) compared to healthy participants, respectively. There were no significant differences looking at the medial arch and the subtalar inversion. Moreover, significant differences regarding the ground reaction force were found for patients after 9 weeks (0–17%: p < 0.001; 21–37%: p < 0.001; 41–54%: p < 0.001; 60–64%: p = 0.013) as well as patients after 26 weeks (0–1,5%: p = 0.046; 5–15%: p < 0.001; 27–33%: p = 0.001; 45–49%: p = 0.005; 57–59%: p = 0.049) compared to healthy participants, respectively. In total, the range of motion in the tibio-talar joint and the medial arch was reduced in affected patients compared to healthy participants. Patients showed significant increase of the range of motion between 9 and 26 weeks. Conclusions This study shows, that patients affected by ankle fractures show limited mobility in the tibio-talar joint and the medial arch when compared to healthy participants. Even though the limitation of motion remains at least over a period of 26 weeks, a significant increase can be recognized over time. Furthermore, if we look at the absolute values, the patients’ values tend to get closer to those of the control group. Trial registration This study is registered at the German Clinical Trials Register (DRKS00023379).

Comparison of the kinematics, repeatability, and reproducibility of five different multi-segment foot models

Background Multi-segment foot models (MFMs) for assessing three-dimensional segmental foot motions are calculated via various analytical methods. Although validation studies have already been conducted, we cannot compare their results because the experimental environments in previous studies were different from each other. This study aims to compare the kinematics, repeatability, and reproducibility of five MFMs in the same experimental conditions. Methods Eleven healthy males with a mean age of 26.5 years participated in this study. We created a merged 29-marker set including five MFMs: Oxford (OFM), modified Rizzoli (mRFM), DuPont (DFM), Milwaukee (MiFM), and modified Shriners Hospital for Children Greenville (mSHCG). Two operators applied the merged model to participants twice, and then we analysed two relative angles of three segments: shank-hindfoot (HF) and hindfoot-forefoot (FF). Coefficients of multiple correlation (CMC) and mean standard errors were used to assess repeatability and reproducibility, and statistical parametric mapping (SPM) of the t-value was employed to compare kinematics. Results HF varus/valgus of the MiFM and mSHCG models, which rotated the segment according to radiographic or goniometric measurements during the reference frame construction, were significantly more repeatable and reproducible, compared to other models. They showed significantly more dorsiflexed HF and plantarflexed FF due to their static offset angles. DFM and mSHCG showed a greater range of motion (ROM), and some models had significantly different FF points of peak angle. Conclusions Under the same conditions, rotating the segment according to the appropriate offset angle obtained from radiographic or goniometric measurement increased reliability, but all MFMs had clinically acceptable reliability compared to previous studies. Moreover, in some models, especially HF varus/valgus, there were differences in ROM and points of peak angle even with no statistical difference in SPM curves. Therefore, based on the results of this study, clinicians and researchers involved in the evaluation of foot and ankle dysfunction need an understanding of the specific features of each MFM to make accurate decisions.

Comparison of Resting-State Brain Activation between Healthy Normal and Low Auditory-Verbal Working Memory Capacity Participants

Working memory (WM) capacity is the ability to maintain attention and store information briefly in the mind. However, each individual has a limited WM capacity that varies from one person to another. An individual can be categorized as having either normal or low WM capacity. This study aimed to evaluate and compare brain activations of healthy individuals with low and normal auditory-verbal WM capacity. A total of 39 healthy male young adults were recruited from local universities for this study. They were categorized into the normal and low auditory-verbal WM capacity group based on their score in the Malay Version of Auditory Verbal Learning Test (MVAVLT). All participants underwent resting-state functional magnetic resonance imaging (rs-fMRI) scans. The functional data were analyzed using Statistical Parametric Mapping (SPM) and Wake Forest University (WFU) Pickatlas softwares. Brain activations and resting-state amplitude fluctuation (rsAF) were contrasted between groups to determine whether there were any significant differences caused by the different auditory-verbal WM capacity. The findings indicated that the low auditory-verbal WM capacity group showed significantly higher cortical activations in the left lingual gyrus, bilateral middle temporal gyrus, left calcarine, left superior frontal gyrus, and left precuneus as compared to normal auditory-verbal WM capacity group. It is suggested that the higher activation of these brain areas in low verbal-auditory WM capacity participants was attributed to the lower neural adaptability of the brain at rest.

Assessment of Effective Network Connectivity among MEG None Contaminated Epileptic Transitory Events

Characterizing epileptogenic zones EZ (sources responsible of excessive discharges) would assist a neurologist during epilepsy diagnosis. Locating efficiently these abnormal sources among magnetoencephalography (MEG) biomarker is obtained by several inverse problem techniques. These techniques present different assumptions and particular epileptic network connectivity. Here, we proposed to evaluate performances of distributed inverse problem in defining EZ. First, we applied an advanced technique based on Singular Value Decomposition (SVD) to recover only pure transitory activities (interictal epileptiform discharges). We evaluated our technique’s robustness in separation between transitory and ripples versus frequency range, transitory shapes, and signal to noise ratio on simulated data (depicting both epileptic biomarkers and respecting time series and spectral properties of realistic data). We validated our technique on MEG signal using detector precision on 5 patients. Then, we applied four methods of inverse problem to define cortical areas and neural generators of excessive discharges. We computed network connectivity of each technique. Then, we confronted obtained noninvasive networks to intracerebral EEG transitory network connectivity using nodes in common, connection strength, distance metrics between concordant nodes of MEG and IEEG, and average propagation delay. Coherent Maximum Entropy on the Mean (cMEM) proved a high matching between MEG network connectivity and IEEG based on distance between active sources, followed by Exact low-resolution brain electromagnetic tomography (eLORETA), Dynamical Statistical Parametric Mapping (dSPM), and Minimum norm estimation (MNE). Clinical performance was interesting for entire methods providing in an average of 73.5% of active sources detected in depth and seen in MEG, and vice versa, about 77.15% of active sources were detected from MEG and seen in IEEG. Investigated problem techniques succeed at least in finding one part of seizure onset zone. dSPM and eLORETA depict the highest connection strength among all techniques. Propagation delay varies in this range [18, 25]ms, knowing that eLORETA ensures the lowest propagation delay (18 ms) and the closet one to IEEG propagation delay.

Three-Dimensional Upper Body Kinematics and Inter-articular Kinematic Sequence During a Canoe Polo Throw

Canoe polo is an increasingly popular discipline requiring both kayaking and ball-handling skills. While the kinematics of the upper body during throw has been investigated for several overhead sports, the canoe polo throw has still to be studied. Therefore, the aim of this study is to analyze the canoe polo throw kinematics in terms of angles and inter-articular sequencing to understand its specificity. A secondary aim was to investigate whether adding pelvis mobility has an impact. Nineteen male players of canoe polo were equipped with reflective body markers for the throw analysis. They performed 5 throws with the pelvis fixed and 5 throws with additional pelvic mobility in rotation around a vertical axis. Inverse kinematics was performed with OpenSim providing pelvis, trunk, and glenohumeral rotations. Angular velocities were calculated to build the inter-articular sequences relative to these throws. Statistical parametric mapping was used to assess the effect of pelvis mobility on the throwing kinematics. Similar kinematics patterns as in other overhead sports were observed, however, a different inter-articular sequence was found for the canoe polo throw with a maximal angular velocity occurring sooner for the thorax in axial rotation than for the pelvis in rotation. While the limitation of rotation of the pelvis around a vertical axis has an influence on the pelvis and trunk kinematics, it did not modify the kinematic sequence.

LeGUI: A Fast and Accurate Graphical User Interface for Automated Detection and Anatomical Localization of Intracranial Electrodes

Accurate anatomical localization of intracranial electrodes is important for identifying the seizure foci in patients with epilepsy and for interpreting effects from cognitive studies employing intracranial electroencephalography. Localization is typically performed by coregistering postimplant computed tomography (CT) with preoperative magnetic resonance imaging (MRI). Electrodes are then detected in the CT, and the corresponding brain region is identified using the MRI. Many existing software packages for electrode localization chain together separate preexisting programs or rely on command line instructions to perform the various localization steps, making them difficult to install and operate for a typical user. Further, many packages provide solutions for some, but not all, of the steps needed for confident localization. We have developed software, Locate electrodes Graphical User Interface (LeGUI), that consists of a single interface to perform all steps needed to localize both surface and depth/penetrating intracranial electrodes, including coregistration of the CT to MRI, normalization of the MRI to the Montreal Neurological Institute template, automated electrode detection for multiple types of electrodes, electrode spacing correction and projection to the brain surface, electrode labeling, and anatomical targeting. The software is written in MATLAB, core image processing is performed using the Statistical Parametric Mapping toolbox, and standalone executable binaries are available for Windows, Mac, and Linux platforms. LeGUI was tested and validated on 51 datasets from two universities. The total user and computational time required to process a single dataset was approximately 1 h. Automatic electrode detection correctly identified 4362 of 4695 surface and depth electrodes with only 71 false positives. Anatomical targeting was verified by comparing electrode locations from LeGUI to locations that were assigned by an experienced neuroanatomist. LeGUI showed a 94% match with the 482 neuroanatomist-assigned locations. LeGUI combines all the features needed for fast and accurate anatomical localization of intracranial electrodes into a single interface, making it a valuable tool for intracranial electrophysiology research.

Providing Evidence for the Null Hypothesis in Functional Magnetic Resonance Imaging Using Group-Level Bayesian Inference

Classical null hypothesis significance testing is limited to the rejection of the point-null hypothesis; it does not allow the interpretation of non-significant results. This leads to a bias against the null hypothesis. Herein, we discuss statistical approaches to ‘null effect’ assessment focusing on the Bayesian parameter inference (BPI). Although Bayesian methods have been theoretically elaborated and implemented in common neuroimaging software packages, they are not widely used for ‘null effect’ assessment. BPI considers the posterior probability of finding the effect within or outside the region of practical equivalence to the null value. It can be used to find both ‘activated/deactivated’ and ‘not activated’ voxels or to indicate that the obtained data are not sufficient using a single decision rule. It also allows to evaluate the data as the sample size increases and decide to stop the experiment if the obtained data are sufficient to make a confident inference. To demonstrate the advantages of using BPI for fMRI data group analysis, we compare it with classical null hypothesis significance testing on empirical data. We also use simulated data to show how BPI performs under different effect sizes, noise levels, noise distributions and sample sizes. Finally, we consider the problem of defining the region of practical equivalence for BPI and discuss possible applications of BPI in fMRI studies. To facilitate ‘null effect’ assessment for fMRI practitioners, we provide Statistical Parametric Mapping 12 based toolbox for Bayesian inference.

Population-specific brain [18F]-FDG PET templates of Chinese subjects for statistical parametric mapping

Statistical Parametric Mapping (SPM) is a computational approach for analysing functional brain images like Positron Emission Tomography (PET). When performing SPM analysis for different patient populations, brain PET template images representing population-specific brain morphometry and metabolism features are helpful. However, most currently available brain PET templates were constructed using the Caucasian data. To enrich the family of publicly available brain PET templates, we created Chinese-specific template images based on 116 [18F]-fluorodeoxyglucose ([18F]-FDG) PET images of normal participants. These images were warped into a common averaged space, in which the mean and standard deviation templates were both computed. We also developed the SPM analysis programmes to facilitate easy use of the templates. Our templates were validated through the SPM analysis of Alzheimer’s and Parkinson’s patient images. The resultant SPM t-maps accurately depicted the disease-related brain regions with abnormal [18F]-FDG uptake, proving the templates’ effectiveness in brain function impairment analysis.