Machine learning model comparison for freezing of gait prediction in advanced Parkinson's disease.

Introduction: Freezing of gait (FOG) is a paroxysmal motor phenomenon that increases in prevalence as Parkinson's disease (PD) progresses. It is associated with a reduced quality of life and an increased risk of falls in this population. Precision-based detection and classification of freezers are critical to developing tailored treatments rooted in kinematic assessments. Methods: This study analyzed instrumented stand-and-walk (SAW) trials from advanced PD patients with STN-DBS. Each patient performed two SAW trials in their OFF Medication-OFF DBS state. For each trial, gait summary statistics from wearable sensors were analyzed by machine learning classification algorithms. These algorithms include k-nearest neighbors, logistic regression, naïve Bayes, random forest, and support vector machines (SVM). Each of these models were selected for their high interpretability. Each algorithm was tasked with classifying patients whose SAW trials MDS-UPDRS FOG subscore was non-zero as assessed by a trained movement disorder specialist. These algorithms' performance was evaluated using stratified five-fold cross-validation. Results: A total of 21 PD subjects were evaluated (average age 64.24 years, 16 males, mean disease duration of 14 years). Fourteen subjects had freezing of gait in the OFF MED/OFF DBS. All machine learning models achieved statistically similar predictive performance (p < 0.05) with high accuracy. Analysis of random forests' feature estimation revealed the top-ten spatiotemporal predictive features utilized in the model: foot strike angle, coronal range of motion [trunk and lumbar], stride length, gait speed, lateral step variability, and toe-off angle. Conclusion: These results indicate that machine learning effectively classifies advanced PD patients as freezers or nonfreezers based on SAW trials in their non-medicated/non-stimulated condition. The machine learning models, specifically random forests, not only rely on but utilize salient spatial and temporal gait features for FOG classification.

Differential Responses to Low- and High-Frequency Subthalamic Nucleus Deep Brain Stimulation on Sensor-Measured Components of Bradykinesia in Parkinson's Disease.

INTRODUCTION: The current approach to assessing bradykinesia in Parkinson's Disease relies on the Unified Parkinson's Disease Rating Scale (UPDRS), which is a numeric scale. Inertial sensors offer the ability to probe subcomponents of bradykinesia: motor speed, amplitude, and rhythm. Thus, we sought to investigate the differential effects of high-frequency compared to low-frequency subthalamic nucleus (STN) deep brain stimulation (DBS) on these quantified facets of bradykinesia. METHODS: We recruited advanced Parkinson's Disease subjects with a chronic bilateral subthalamic nucleus (STN) DBS implantation to a single-blind stimulation trial where each combination of medication state (OFF/ON), electrode contacts, and stimulation frequency (60 Hz/180 Hz) was assessed. The Kinesia One sensor system was used to measure upper limb bradykinesia. For each stimulation trial, subjects performed extremity motor tasks. Sensor data were recorded continuously. We identified STN DBS parameters that were associated with improved upper extremity bradykinesia symptoms using a mixed linear regression model. RESULTS: We recruited 22 subjects (6 females) for this study. The 180 Hz STN DBS (compared to the 60 Hz STN DBS) and dopaminergic medications improved all subcomponents of upper extremity bradykinesia (motor speed, amplitude, and rhythm). For the motor rhythm subcomponent of bradykinesia, ventral contacts yielded improved symptom improvement compared to dorsal contacts. CONCLUSION: The differential impact of high- and low-frequency STN DBS on the symptoms of bradykinesia may advise programming for these patients but warrants further investigation. Wearable sensors represent a valuable addition to the armamentarium that furthers our ability to conduct objective, quantitative clinical assessments.

Towards Faster Gene Expression Prediction via Dimensionality Reduction and Feature Selection.

The majority of genes have a genetic component to their expression. Elastic nets have been shown effective at predicting tissue-specific, individual-level gene expression from genotype data. We apply principal component analysis (PCA), linkage disequilibrium pruning, or the combination of the two to reduce, or generate, a lower-dimensional representation of the genetic variants used as inputs to the elastic net models for the prediction of gene expression. Our results show that, in general, elastic nets attain their best performance when all genetic variants are included as inputs; however, a relatively low number of principal components can effectively summarize the majority of genetic variation while reducing the overall computation time. Specifically, 100 principal components reduce the computational time of the models by over 80% with only an 8% loss in R2. Finally, linkage disequilibrium pruning does not effectively reduce the genetic variants for predicting gene expression. As predictive models are commonly made for over 27,000 genes for more than 50 tissues, PCA may provide an effective method for reducing the computational burden of gene expression analysis.

Differential spatiotemporal gait effects with frequency and dopaminergic modulation in STN-DBS.

Objective: The spatiotemporal gait changes in advanced Parkinson's disease (PD) remain a treatment challenge and have variable responses to L-dopa and subthalamic deep brain stimulation (STN-DBS). The purpose of this study was to determine whether low-frequency STN-DBS (LFS; 60 Hz) elicits a differential response to high-frequency STN-DBS (HFS; 180 Hz) in spatiotemporal gait kinematics. Methods: Advanced PD subjects with chronic STN-DBS were evaluated in both the OFF and ON medication states with LFS and HFS stimulation. Randomization of electrode contact pairs and frequency conditions was conducted. Instrumented Stand and Walk assessments were carried out for every stimulation/medication condition. LM-ANOVA was employed for analysis. Results: Twenty-two PD subjects participated in the study, with a mean age (SD) of 63.9 years. Significant interactions between frequency (both LFS and HFS) and electrode contact pairs (particularly ventrally located contacts) were observed for both spatial (foot elevation, toe-off angle, stride length) and temporal (foot speed, stance, single limb support (SLS) and foot swing) gait parameters. A synergistic effect was also demonstrated with L-dopa and both HFS and LFS for right SLS, left stance, left foot swing, right toe-off angle, and left arm range of motion. HFS produced significant improvement in trunk and lumbar range of motion compared to LFS. Conclusion: The study provides evidence of synergism of L-dopa and STN-DBS on lower limb spatial and temporal measures in advanced PD. HFS and LFS STN-DBS produced equivalent effects among all other tested lower limb gait features. HFS produced significant trunk and lumbar kinematic improvements.

Evaluating the Impact of Health Care Data Completeness for Deep Generative Models.

BACKGROUND: Deep generative models (DGMs) present a promising avenue for generating realistic, synthetic data to augment existing health care datasets. However, exactly how the completeness of the original dataset affects the quality of the generated synthetic data is unclear. OBJECTIVES: In this paper, we investigate the effect of data completeness on samples generated by the most common DGM paradigms. METHODS: We create both cross-sectional and panel datasets with varying missingness and subset rates and train generative adversarial networks, variational autoencoders, and autoregressive models (Transformers) on these datasets. We then compare the distributions of generated data with original training data to measure similarity. RESULTS: We find that increased incompleteness is directly correlated with increased dissimilarity between original and generated samples produced through DGMs. CONCLUSIONS: Care must be taken when using DGMs to generate synthetic data as data completeness issues can affect the quality of generated data in both panel and cross-sectional datasets.

Adapting Random Forests to Predict Obesity-Associated Gene Expression.

Random forests (RFs) are effective at predicting gene expression from genotype data. However, a comparison of RF regressors and classifiers, including feature selection and encoding, has been under-explored in the context of gene expression prediction. Specifically, we examine the role of ordinal or one-hot encoding and of data balancing via oversam-pling in the prediction of obesity-associated gene expression. Our work shows that RFs compete with PrediXcan in the prediction of obesity-associated gene expression in subcutaneous adipose tissue, a highly relevant tissue to obesity. Additionally, RFs generate predictions for obesity-associated genes where PrediXcan fails to do so.

Using Verb Fluency, Natural Language Processing, and Machine Learning to Detect Alzheimer's Disease.

Alzheimer's disease (AD) causes significant impairments in memory and other cognitive domains. As there is no cure to the disease yet, early detection and delay of disease progression are critical for management of AD. Verbal fluency is one of the most common and sensitive neuropsychological methods used for detection and evaluation of the cognitive declines in AD, in which a subject is required to name as many items as possible in 30 or 60 seconds that belong to a certain category. In this study, we develop an approach to detect AD using a verb fluency (VF) task, a specific subset of verbal fluency analyzing the subjects' listing of verbs in a given time period. We use machine learning techniques including random forest (RF), neural network (NN), recurrent NN (RNN), and natural language processing (NLP) to detect the risk of AD. The results show that the developed models can stratify subjects into the corresponding AD and control groups with up to 76% accuracy using RF, but at a cost of having to preprocess the data. This accuracy is slightly lower, but not significantly, at 67% using RNN and NLP, which involves almost no manual preprocessing of the data. This study opens up a powerful approach of using simple VF tasks for early detection of AD.

Predicting Severity in People with Aphasia: A Natural Language Processing and Machine Learning Approach.

Speech language pathologists need an accurate assessment of the severity of people with aphasia (PWA) to design and provide the best course of therapy. Currently, severity is evaluated manually by an increasingly scarce pool of experienced and well-trained clinicians, taking considerable time resources. By analyzing the transcripts from three discourse elicitation methods, this study combines natural language processing (NLP) and machine learning (ML) to predict the severity of PWA, both by score and severity level. By engineering language features from PWA tasks, an unstructured k-means clustering presents distinct aphasia types, showing validity of the selected features. We develop regression models to predict severity scores along with a classification of severity by level (Mild, Moderate, Severe, and Very Severe) to assist clinicians to easily plan and monitor the course of treatment. Our best ML regression model uses a deep neural network and results in a mean absolute error (MAE) of 0.0671 and root mean squared error (RMSE) of 0.0922. Our best classification model uses a random forest and result in an overall accuracy of 73%, with the highest accuracy of 87.5% for mild severity. Our results suggest that using NLP and ML provides an accurate and cost-effective approach to evaluate the severity levels in PWA to consequently help clinicians determine rehabilitation procedures.

Evaluating the Fitness-to-Drive Using Evoked Visual Responses in Alzheimer's Disease.

Alzheimer's Disease (AD) is the sixth leading cause of death in the US. AD causes significant disability due to the devastating impact on the patients' day-to-day living activities and their loss of independence. One such day-to-day activity is driving, a complex task that requires attention, concentration, the ability to follow particular steps, react to stimuli promptly, and the ability to perceive and interpret visual-spatial information, all of which can be impaired in AD. Therefore, to ensure the safety of AD patients and other drivers, it is important to develop accurate and low-cost diagnostic tools to assess patients' fitness-to-drive. In this study, we develop machine learning (ML) models to predict fitness-to-drive using the electroencephalogram (EEG) technique of event-related potential (ERP). Specifically, we develop random forest (RF) models using EEG signals in early-stage AD patients and age-matched controls and conduct numerical experiments to predict fitness-to-drive and other driving performance metrics, collected from driving simulator data. Our results show that RF models predict patients' fitness-to-drive with AUC=0.83 and provide accurate measures of other driving performance metrics. Therefore, ML and ERP offer a valuable approach to assess driving safety for patients with early AD symptoms in the laboratory setting.

Improving Medication Regimen Recommendation for Parkinson's Disease Using Sensor Technology.

Parkinson's disease medication treatment planning is generally based on subjective data obtained through clinical, physician-patient interactions. The Personal KinetiGraph™ (PKG) and similar wearable sensors have shown promise in enabling objective, continuous remote health monitoring for Parkinson's patients. In this proof-of-concept study, we propose to use objective sensor data from the PKG and apply machine learning to cluster patients based on levodopa regimens and response. The resulting clusters are then used to enhance treatment planning by providing improved initial treatment estimates to supplement a physician's initial assessment. We apply k-means clustering to a dataset of within-subject Parkinson's medication changes-clinically assessed by the MDS-Unified Parkinson's Disease Rating Scale-III (MDS-UPDRS-III) and the PKG sensor for movement staging. A random forest classification model was then used to predict patients' cluster allocation based on their respective demographic information, MDS-UPDRS-III scores, and PKG time-series data. Clinically relevant clusters were partitioned by levodopa dose, medication administration frequency, and total levodopa equivalent daily dose-with the PKG providing similar symptomatic assessments to physician MDS-UPDRS-III scores. A random forest classifier trained on demographic information, MDS-UPDRS-III scores, and PKG time-series data was able to accurately classify subjects of the two most demographically similar clusters with an accuracy of 86.9%, an F1 score of 90.7%, and an AUC of 0.871. A model that relied solely on demographic information and PKG time-series data provided the next best performance with an accuracy of 83.8%, an F1 score of 88.5%, and an AUC of 0.831, hence further enabling fully remote assessments. These computational methods demonstrate the feasibility of using sensor-based data to cluster patients based on their medication responses with further potential to assist with medication recommendations.

Building resiliency to climate change uncertainty through bioretention design modifications.

Climate stationarity is a traditional assumption in the design of the urban drainage network, including green infrastructure practices such as bioretention cells. Predicted deviations from historic climate trends associated with global climate change introduce uncertainty in the ability of these systems to maintain service levels in the future. Climate change projections are made using output from coarse-scale general circulation models (GCMs), which can then be downscaled using regional climate models (RCMs) to provide predictions at a finer spatial resolution. However, all models contain sources of error and uncertainty, and predicted changes in future climate can be contradictory between models, requiring an approach that considers multiple projections. The performance of bioretention cells were modeled using USEPA's Storm Water Management Model (SWMM) to determine how design modifications could add resilience to these systems under future climate conditions projected for Knoxville, Tennessee, USA. Ten downscaled climate projections were acquired from the North American Coordinated Regional Downscaling Experiment program, and model bias was corrected using Kernel Density Distribution Mapping (KDDM). Bias-corrected climate projections were used to assess bioretention hydrologic function in future climate conditions. Several scenarios were evaluated using a probabilistic approach to determine the confidence with which design modifications could be implemented to maintain historic performance for both new and existing (retrofitted) bioretention cells. The largest deviations from current design (i.e., concurrently increasing ponding depths, thickness of media layer, media conductivity rates, and bioretention surface areas by 307%, 200%, 200%, and 300%, respectively, beyond current standards) resulted in the greatest improvements on historic performance with respect to annual volumes of infiltration and surface overflow, with all ten future climate scenarios across various soil types yielding increased infiltration and decreased surface overflow compared to historic conditions. However, lower performance was observed for more conservative design modifications; on average, between 13-82% and 77-100% of models fell below historic annual volumes of infiltration and surface overflow, respectively, when ponding zone depth, media layer thickness, and media conductivity were increased alone. Findings demonstrate that increasing bioretention surface area relative to the contributing catchment provides the greatest overall return on historic performance under future climate conditions and should be prioritized in locations with low in situ soil drainage rates. This study highlights the importance of considering local site conditions and management objectives when incorporating resiliency to climate change uncertainty into bioretention designs.

HeMA: A hierarchically enriched machine learning approach for managing false alarms in real time: A sepsis prediction case study.

Early detection of sepsis can be life-saving. Machine learning models have shown great promise in early sepsis prediction when applied to patient physiological data in real-time. However, these existing models often under-perform in terms of positive predictive value, an important metric in clinical settings. This is especially the case when the models are applied to data with less than 50% sepsis prevalence, reflective of the incidence rate of sepsis on the floor or in the ICU. In this study, we develop HeMA, a hierarchically enriched machine learning approach for managing false alarms in real time, and conduct a case study for early sepsis prediction. Specifically, we develop a two-stage framework, where a first stage machine learning model is paired with statistical tests, particularly Kolmogorov-Smirnov tests, in the second stage, to predict whether a patient would develop sepsis. Compared with machine learning models alone, the framework results in an increase in specificity and positive predictive value, without compromising F1 score. In particular, the framework shows improved performance when applied to data with 50% and 25% sepsis prevalence, collected from a large hospital system in the US, resulting in up to 18% and 7% increase in specificity and positive predictive value, respectively. Despite the significant improvements observed, and although F1 score is not negatively affected, because of the up to 6% decrease in sensitivity, further improvements and pilot studies may be necessary before deploying the framework in a clinical setting. Finally, external validation conducted using a publicly available dataset produces similar results, validating that the proposed framework is generalizable.

Improving Deep Reinforcement Learning With Transitional Variational Autoencoders: A Healthcare Application.

Reinforcement learning is a powerful tool for developing personalized treatment regimens from healthcare data. Yet training reinforcement learning agents through direct interactions with patients is often impractical for ethical reasons. One solution is to train reinforcement learning agents using an 'environment model,' which is learned from retrospective patient data, and can simulate realistic patient trajectories. In this study, we propose transitional variational autoencoders (tVAE), a generative neural network architecture that learns a direct mapping between distributions over clinical measurements at adjacent time points. Unlike other models, the tVAE requires few distributional assumptions, and benefits from identical training, and testing architectures. This model produces more realistic patient trajectories than state-of-the-art sequential decision-making models, and generative neural networks, and can be used to learn effective treatment policies.

Temporal Differential Expression of Physiomarkers Predicts Sepsis in Critically Ill Adults.

BACKGROUND: Sepsis is a life-threatening condition with high mortality rates. Early detection and treatment are critical to improving outcomes. Our primary objective was to develop artificial intelligence capable of predicting sepsis earlier using a minimal set of streaming physiological data in real time. METHODS AND FINDINGS: A total of 29,552 adult patients were admitted to the intensive care unit across five regional hospitals in Memphis, Tenn, over 18 months from January 2017 to July 2018. From these, 5,958 patients were selected after filtering for continuous (minute-by-minute) physiological data availability. A total of 617 (10.4%) patients were identified as sepsis cases, using the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3) criteria. Physiomarkers, a set of signal processing features, were derived from five physiological data streams including heart rate, respiratory rate, and blood pressure (systolic, diastolic, and mean), captured every minute from the bedside monitors. A support vector machine classifier was used for classification. The model accurately predicted sepsis up to a mean and 95% confidence interval of 17.4 ±â€Š0.22 h before sepsis onset, with an average test accuracy of 83.0% (average sensitivity, specificity, and area under the receiver operating characteristics curve of 0.757, 0.902, and 0.781, respectively). CONCLUSIONS: This study demonstrates that salient physiomarkers derived from continuous bedside monitoring are temporally and differentially expressed in septic patients. Using this information, minimalistic artificial intelligence models can be developed to predict sepsis earlier in critically ill patients.

Machine Learning's Application in Deep Brain Stimulation for Parkinson's Disease: A Review.

Deep brain stimulation (DBS) is a surgical treatment for advanced Parkinson's disease (PD) that has undergone technological evolution that parallels an expansion in clinical phenotyping, neurophysiology, and neuroimaging of the disease state. Machine learning (ML) has been successfully used in a wide range of healthcare problems, including DBS. As computational power increases and more data become available, the application of ML in DBS is expected to grow. We review the literature of ML in DBS and discuss future opportunities for such applications. Specifically, we perform a comprehensive review of the literature from PubMed, the Institute for Scientific Information's Web of Science, Cochrane Database of Systematic Reviews, and Institute of Electrical and Electronics Engineers' (IEEE) Xplore Digital Library for ML applications in DBS. These studies are broadly placed in the following categories: (1) DBS candidate selection; (2) programming optimization; (3) surgical targeting; and (4) insights into DBS mechanisms. For each category, we provide and contextualize the current body of research and discuss potential future directions for the application of ML in DBS.

Optimizing Individualized Treatment Planning for Parkinson's Disease Using Deep Reinforcement Learning.

More than one million people currently live with Parkinson's Disease (PD) in the U.S. alone. Medications, such as levodopa, can help manage PD symptoms. However, medication treatment planning is generally based on patient history and limited interaction between physicians and patients during office visits. This limits the extent of benefit that may be derived from the treatment as disease/patient characteristics are generally non-stationary. Wearable sensors that provide continuous monitoring of various symptoms, such as bradykinesia and dyskinesia, can enhance symptom management. However, using such data to overhaul the current static medication treatment planning approach and prescribe personalized medication timing and dosage that accounts for patient/care-giver/physician feedback/preferences remains an open question. We develop a model to prescribe timing and dosage of medications, given the motor fluctuation data collected using wearable sensors in real-time. We solve the resulting model using deep reinforcement learning (DRL). The prescribed policy determines the optimal treatment plan that minimizes patient's symptoms. Our results show that the model-prescribed policy outperforms the static a priori treatment plan in improving patients' symptoms, providing a proof-of-concept that DRL can augment medical decision making for treatment planning of chronic disease patients.

Predicting Alzheimer's Disease Using Driving Simulator Data.

Early detection of Alzheimer's Disease (AD) is critical in creating better outcomes for patients. Performance in complex tasks such as vehicular driving may be a sensitive tool for early detection of AD and serve as a good indicator of functional status. In this study, we investigate the classification of AD patients and controls using driving simulator data. Our results show that machine learning algorithms, especially random forest classifier, can accurately discriminate AD patients and controls (AUC = 0.96, Sensitivity = 87%, and Specificity = 93%). The model-identified most important features include Pothole Avoidance, Road Signs Recalled, Inattention Measurements, Reaction Time, and Detection Times, among others, all of which closely align with previous studies about cognitive functions that are affected by AD.

Improving Shared Decision-making and Treatment Planning Through Predictive Modeling: Clinical Insights on Ventral Hernia Repair.

Abdominal wall hernia repair, including ventral hernia repair, is one of the most common general surgical procedures. Nationally, at least 350 000 ventral hernia repairs are performed annually, and of those, 150 000 cases were identified as incisional hernias. Outcomes are reported to be poor, resulting in additional surgical repair rates of 12.3% at 5 years and as high as 23% at 10 years. Healthcare costs associated with ventral hernia repair are estimated to exceed $3 billion each year. Additionally, ventral hernia repair is often complex and unpredictable when there is a current infection or a history of infection and significant comorbidities. Accordingly, a predictive model was developed using a retrospectively collected dataset to associate the pre- and intra-operative characteristics of patients to their outcomes, with the primary goal of identifying patients at risk of developing complications a priori in the future. The benefits and implications of such a predictive model, however, extend beyond this primary goal. This predictive model can serve as an important tool for clinicians who may use it to support their clinical intuition and clarify patient need for lifestyle modification prior to abdominal wall reconstruction. This predictive model can also support shared decision-making so that a personalized plan of care may be developed. The outcomes associated with use of the predictive model may include surgical repair but may suggest lifestyle modification coupled with less invasive interventions.

Improving Prediction Performance Using Hierarchical Analysis of Real-Time Data: A Sepsis Case Study.

This paper presents a novel method for hierarchical analysis of machine learning algorithms to improve predictions of at risk patients, thus further enabling prompt therapy. Specifically, we develop a multi-layer machine learning approach to analyze continuous, high-frequency data. We illustrate the capabilities of this approach for early identification of patients at risk of sepsis, a potentially life-threatening complication of an infection, using high-frequency (minute-by-minute) physiological data collected from bedside monitors. In our analysis of a cohort of 586 patients, the model obtained from analyzing the output of a previously developed sepsis prediction model resulted in improved outcomes. Specifically, the original model failed to predict 11.76 ± 4.26% of sepsis patients earlier than Systemic Inflammatory Response Syndrome (SIRS) criteria, commonly used to identify patients at risk for rapid physiological deterioration resulting from sepsis. In contrast, the multi-layer model only failed to predict 3.21 ± 3.11% of sepsis patients earlier than SIRS. In addition, sepsis patients were predicted on average 204.87 ± 7.90 minutes earlier than SIRS criteria using the multi-layer model, which can potentially help reduce mortality and morbidity if implemented in the ICU.