Transparent Machine Learning Models to Diagnose Suspicious Thoracic Lesions Leveraging CT Guided Biopsy Data.

RATIONALE AND OBJECTIVES: To train and validate machine learning models capable of classifying suspicious thoracic lesions as benign or malignant and to further classify malignant lesions by pathologic subtype while quantifying feature importance for each classification. MATERIALS AND METHODS: 796 patients who had undergone CT guided thoracic biopsy for a concerning thoracic lesion (79.3% lung, 11.4% mediastinum, 6.5% pleura, 2.7% chest wall) were retrospectively enrolled. Lesions were classified as malignant or benign based on ground-truth pathology result, and malignant lesions were classified as primary or secondary cancer. Clinical variables were extracted from EMR and radiology reports. Supervised binary and multiclass classification models were trained to classify lesions based on the input features and evaluated on a held-out test set. Model specific feature analyses were performed to identify variables most predictive of each class, as well as to assess the independent importance of clinical, and imaging features. RESULTS: Binary classification models achieved a top accuracy of 80.6%, with predictive features included smoking history, age, lesion size, and lesion location. Multiclass classification models achieved a top weighted average f1-score of 0.73. Features predictive of primary cancer included smoking history, race, and age, while features predictive of secondary cancer included lesion location, and a history of cancer. CONCLUSION: Machine learning models enable classification of suspicious thoracic lesions based on clinical and imaging variables, achieving clinically useful performance while identifying importance of individual input features on a pathology-proven dataset. We believe models such as these are more likely to be trusted and adopted by clinicians.

CT Fluoroscopy Guided Thoracic Biopsies (CTTB) Are Highly Accurate and Safe: Outcomes and Predictive Modeling of Complications Utilizing Machine Learning.

PURPOSE: CT guided transthoracic biopsy (CTTB) is an established, minimally invasive method for diagnostic evaluation of a variety of thoracic diseases. We assessed a large CTTB cohort diagnostic accuracy, complication rates, and developed machine learning models to predict complications. MATERIALS AND METHODS: We retrospectively identified 796 CTTB patients in a tertiary hospital (5-year interval). We gathered and coded patient demographics, characteristics of each lesion biopsied, type of biopsy, diagnostic yield, type of diagnosis, and complication rates. Statistical analyses included summary statistics, multivariate logistic regression and machine learning (neural network) methods. RESULTS: Seven hundred ninety-six CTTBs were performed (43% fine needle aspirations, 5% core biopsies, 52% both). Diagnostic yield was 97.0% (73.9% malignant, 23.1% benign). Complications occurred in 14.7% (12.7% minor, 2.0% major). The most common complication was pneumothorax (13.1%), mostly minor. Multivariate logistic regression models could predict severity of complications with accuracies ranging from 65.5% to 83.5%, with smaller lesion dimension the strongest predictor. Type of biopsy was not a statistically significant predictor. A neural network model improved accuracy to 77.0%-94.2%. CONCLUSION: CTTB performed by thoracic radiologists in a tertiary hospital demonstrate excellent diagnostic yield (97.0%) with a low clinically important complication rate (2.0%). Machine learning methods including neural networks can accurately predict the likelihood of complications, offering pathways to potentially improve patient selection and procedural technique, in order to further optimize the risk-benefit ratio of CTTB.

Brain-state classification and a dual-state decoder dramatically improve the control of cursor movement through a brain-machine interface.

OBJECTIVE: It is quite remarkable that brain machine interfaces (BMIs) can be used to control complex movements with fewer than 100 neurons. Success may be due in part to the limited range of dynamical conditions under which most BMIs are tested. Achieving high-quality control that spans these conditions with a single linear mapping will be more challenging. Even for simple reaching movements, existing BMIs must reduce the stochastic noise of neurons by averaging the control signals over time, instead of over the many neurons that normally control movement. This forces a compromise between a decoder with dynamics allowing rapid movement and one that allows postures to be maintained with little jitter. Our current work presents a method for addressing this compromise, which may also generalize to more highly varied dynamical situations, including movements with more greatly varying speed. APPROACH: We have developed a system that uses two independent Wiener filters as individual components in a single decoder, one optimized for movement, and the other for postural control. We computed an LDA classifier using the same neural inputs. The decoder combined the outputs of the two filters in proportion to the likelihood assigned by the classifier to each state. MAIN RESULTS: We have performed online experiments with two monkeys using this neural-classifier, dual-state decoder, comparing it to a standard, single-state decoder as well as to a dual-state decoder that switched states automatically based on the cursor's proximity to a target. The performance of both monkeys using the classifier decoder was markedly better than that of the single-state decoder and comparable to the proximity decoder. SIGNIFICANCE: We have demonstrated a novel strategy for dealing with the need to make rapid movements while also maintaining precise cursor control when approaching and stabilizing within targets. Further gains can undoubtedly be realized by optimizing the performance of the individual movement and posture decoders.

Multimodal decoding and congruent sensory information enhance reaching performance in subjects with cervical spinal cord injury.

Cervical spinal cord injury (SCI) paralyzes muscles of the hand and arm, making it difficult to perform activities of daily living. Restoring the ability to reach can dramatically improve quality of life for people with cervical SCI. Any reaching system requires a user interface to decode parameters of an intended reach, such as trajectory and target. A challenge in developing such decoders is that often few physiological signals related to the intended reach remain under voluntary control, especially in patients with high cervical injuries. Furthermore, the decoding problem changes when the user is controlling the motion of their limb, as opposed to an external device. The purpose of this study was to investigate the benefits of combining disparate signal sources to control reach in people with a range of impairments, and to consider the effect of two feedback approaches. Subjects with cervical SCI performed robot-assisted reaching, controlling trajectories with either shoulder electromyograms (EMGs) or EMGs combined with gaze. We then evaluated how reaching performance was influenced by task-related sensory feedback, testing the EMG-only decoder in two conditions. The first involved moving the arm with the robot, providing congruent sensory feedback through their remaining sense of proprioception. In the second, the subjects moved the robot without the arm attached, as in applications that control external devices. We found that the multimodal-decoding algorithm worked well for all subjects, enabling them to perform straight, accurate reaches. The inclusion of gaze information, used to estimate target location, was especially important for the most impaired subjects. In the absence of gaze information, congruent sensory feedback improved performance. These results highlight the importance of proprioceptive feedback, and suggest that multi-modal decoders are likely to be most beneficial for highly impaired subjects and in tasks where such feedback is unavailable.

Benzalkonium chloride-induced denervation of orbicularis oculi muscle in rabbits.

PURPOSE: To examine the potential for benzalkonium chloride (BAK) to cause denervation of the orbicularis oculi muscle (OOM) in a rabbit model. METHODS: Pigmented rabbits were separated into five groups consisting of five rabbits each. Group 1 was injected with 1 mL of BAK 0.25% in the OOM of the upper eyelid. Group 2 was injected with 1 mL of BAK 0.5%. Group 3 included untreated controls. Groups 4 and 5 underwent surgical severing of the facial nerve (to cause complete paralysis of the OOM). Strength-duration curves for electrical stimulation of muscle twitches were measured for each group and chronaxie values were calculated to determine innervation status. Groups 1 and 4 were stimulated at 1 week postintervention while groups 2 and 5 were stimulated at 4 weeks postintervention. The rabbits were then sacrificed and the eyelids sent for histological analysis. RESULTS: In group 1, all five rabbits demonstrated denervation of the OOM in the injected area. In group 2, one rabbit developed an abscess at the injection site and was sacrificed at 1 week. Of the remaining four rabbits, two showed complete denervation and two showed denervation with evidence of partial reinnervation. The histology demonstrated marked atrophy of the OOM for BAK-treated rabbits when compared with controls. In group 3, all five rabbits showed normal OOM function. In groups 4 and 5, all rabbits showed denervation of the OOM and histological evidence of muscle atrophy similar to groups 1 and 2. CONCLUSIONS: BAK causes denervation when injected into the OOM in rabbits. The clinical relevance of this finding may be the onset of lagophthalmos and eyelid retraction in patients with chronic BAK exposure.

Functional connectivity and tuning curves in populations of simultaneously recorded neurons.

How interactions between neurons relate to tuned neural responses is a longstanding question in systems neuroscience. Here we use statistical modeling and simultaneous multi-electrode recordings to explore the relationship between these interactions and tuning curves in six different brain areas. We find that, in most cases, functional interactions between neurons provide an explanation of spiking that complements and, in some cases, surpasses the influence of canonical tuning curves. Modeling functional interactions improves both encoding and decoding accuracy by accounting for noise correlations and features of the external world that tuning curves fail to capture. In cortex, modeling coupling alone allows spikes to be predicted more accurately than tuning curve models based on external variables. These results suggest that statistical models of functional interactions between even relatively small numbers of neurons may provide a useful framework for examining neural coding.

Statistical assessment of the stability of neural movement representations.

In systems neuroscience, neural activity that represents movements or sensory stimuli is often characterized by spatial tuning curves that may change in response to training, attention, altered mechanics, or the passage of time. A vital step in determining whether tuning curves change is accounting for estimation uncertainty due to measurement noise. In this study, we address the issue of tuning curve stability using methods that take uncertainty directly into account. We analyze data recorded from neurons in primary motor cortex using chronically implanted, multielectrode arrays in four monkeys performing center-out reaching. With the use of simulations, we demonstrate that under typical experimental conditions, the effect of neuronal noise on estimated preferred direction can be quite large and is affected by both the amount of data and the modulation depth of the neurons. In experimental data, we find that after taking uncertainty into account using bootstrapping techniques, the majority of neurons appears to be very stable on a timescale of minutes to hours. Lastly, we introduce adaptive filtering methods to explicitly model dynamic tuning curves. In contrast to several previous findings suggesting that tuning curves may be in constant flux, we conclude that the neural representation of limb movement is, on average, quite stable and that impressions to the contrary may be largely the result of measurement noise.

Dealing with noisy gaze information for a target-dependent neural decoder.

We tend to look at targets prior to moving our hand towards them. This means that our eye movements contain information about the movements we are planning to make. This information has been shown to be useful in the context of decoding of movement intent from neural signals. However, this is complicated by the fact that occasionally, subjects may want to move towards targets that have not been foveated, or may be distracted and temporarily look away from the intended target. We have previously accounted for this uncertainty using a probabilistic mixture over targets, where the gaze information is used to identify target candidates. Here we evaluate how the accuracy of prior target information influences decoding accuracy. We also consider a mixture model where we assume that the target may be foveated or, alternatively, that the target may not be foveated. We found that errors due to inaccurate target information were reduced by including a generic model representing movements to all targets into the mixture.

Continuous movement decoding using a target-dependent model with EMG inputs.

Trajectory-based models that incorporate target position information have been shown to accurately decode reaching movements from bio-control signals, such as muscle (EMG) and cortical activity (neural spikes). One major hurdle in implementing such models for neuroprosthetic control is that they are inherently designed to decode single reaches from a position of origin to a specific target. Gaze direction can be used to identify appropriate targets, however information regarding movement intent is needed to determine when a reach is meant to begin and when it has been completed. We used linear discriminant analysis to classify limb states into movement classes based on recorded EMG from a sparse set of shoulder muscles. We then used the detected state transitions to update target information in a mixture of Kalman filters that incorporated target position explicitly in the state, and used EMG activity to decode arm movements. Updating the target position initiated movement along new trajectories, allowing a sequence of appropriately timed single reaches to be decoded in series and enabling highly accurate continuous control.

Development of a BIONic muscle spindle for prosthetic proprioception.

The replacement of proprioceptive function, whether for conscious sensation or feedback control, is likely to be an important aspect of neural prosthetic restoration of limb movements. Thus far, however, it has been hampered by the absence of unobtrusive sensors. We propose a method whereby fully implanted, telemetrically operated BIONs monitor muscle movement, and thereby detect changes in joint angle(s) and/or limb posture without requiring the use of secondary components attached to limb segments or external reference frames. The sensor system is designed to detect variations in the electrical coupling between devices implanted in neighboring muscles that result from changes in their relative position as the muscles contract and stretch with joint motion. The goal of this study was to develop and empirically validate mathematical models of the sensing scheme and to use computer simulations to provide an early proof of concept and inform design of the overall sensor system. Results from experiments using paired dipoles in a saline bath and finite element simulations have given insight into the current distribution and potential gradients exhibited within bounded anisotropic environments similar to a human limb segment and demonstrated an anticipated signal to noise ratio of at least 8:1 for submillimeter resolution of relative implant movement over a range of implant displacements up to 15 cm.

Electrical stimulation of the paralyzed orbicularis oculi in rabbit.

Dysfunction of the seventh cranial nerve often results in facial paralysis and loss of the ability to blink the eye, which can lead to corneal scarring, diminished vision, and potential loss of the eye. This study investigated the potential of electrical stimulation of the orbicularis oculi muscle as a means of restoring blink function. An animal model of orbicularis paralysis was created by sectioning the seventh cranial nerve in rabbit. Twenty paralyzed and five normal rabbits were acutely implanted with a subcutaneous stimulating electrode near the margin of the upper eyelid. Biphasic current controlled stimulation pulses were delivered between implanted contacts at the medial and lateral edges of the eyelid. Strength-duration curves for lid twitch threshold were generated, and quantitative measurements of lid closure were made for systematically varied parameters including pulse amplitude, pulse width, number of pulses delivered, and duration of paralysis prior to stimulation. Normal rabbits achieved a greater degree of lid closure due to electrical stimulation than rabbits that had been surgically paralyzed. Of rabbits that had been paralyzed, those demonstrating evidence of at least partial reinnervation achieved a greater degree of lid closure than those demonstrating persistent denervation. Trains of 10 ms biphasic pulses delivered at 50 Hz were found to be the most effective means of eliciting lid closure for the range of parameters tested.

Kinematics of electrically elicited eyelid movement.

Electrical stimulation has demonstrated potential for reanimating eye blink following facial paralysis caused by damage to the seventh cranial nerve. This study investigated the kinematics of lid movement caused by electrical stimulation of the orbicularis oculi muscle in both normal rabbit and rabbit with surgically induced seventh nerve lesion.

Virtual Visit: improving communication for those who need it most.

We are developing a videoconferencing system designed specifically to cater to the needs of the elderly and their distant friends and family. Focus group research has led to the incorporation of several key features that enhance the usability of the system. Audio and video transmission quality are optimized by separation, using the proven and familiar capability of analog phone for voice and the slightly less reliable, but faster data transmission of broadband Internet for video. A simple and intuitive user interface was designed based on the familiar steps of an actual household visit. The graphical interface is presented on the user's conventional television receiver and controlled by a simple seven-button remote control with integrated wireless microphone, which provides high quality audio pickup. The system can be remotely activated by a wearable wireless alarm button or by visitors with passkey privileges. A functional prototype has been developed and is currently undergoing field-testing. Preliminary response has been very encouraging. Future plans include extended focus group research, collaboration to integrate improved video transmission schemes, and in-home testing by seniors to gauge long-term user response.