A Review of Feature Selection Methods for Machine Learning-Based Disease Risk Prediction.

Machine learning has shown utility in detecting patterns within large, unstructured, and complex datasets. One of the promising applications of machine learning is in precision medicine, where disease risk is predicted using patient genetic data. However, creating an accurate prediction model based on genotype data remains challenging due to the so-called "curse of dimensionality" (i.e., extensively larger number of features compared to the number of samples). Therefore, the generalizability of machine learning models benefits from feature selection, which aims to extract only the most "informative" features and remove noisy "non-informative," irrelevant and redundant features. In this article, we provide a general overview of the different feature selection methods, their advantages, disadvantages, and use cases, focusing on the detection of relevant features (i.e., SNPs) for disease risk prediction.

Detection of sleep apnea from single-channel electroencephalogram (EEG) using an explainable convolutional neural network (CNN).

Sleep apnea (SA) is a common disorder involving the cessation of breathing during sleep. It can cause daytime hypersomnia, accidents, and, if allowed to progress, serious, chronic conditions. Continuous positive airway pressure is an effective SA treatment. However, long waitlists impede timely diagnosis; overnight sleep studies involve trained technicians scoring a polysomnograph, which comprises multiple physiological signals including multi-channel electroencephalography (EEG). Therefore, it is important to develop simplified and automated approaches to detect SA. In the present study, we have developed an explainable convolutional neural network (CNN) to detect SA events from single-channel EEG recordings which generalizes across subjects. The network architecture consisted of three convolutional layers. We tuned hyperparameters using the Hyperband algorithm, optimized parameters using Adam, and quantified network performance with subjectwise 10-fold cross-validation. Our CNN performed with an accuracy of 69.9%, and a Matthews correlation coefficient (MCC) of 0.38. To explain the mechanisms of our trained network, we used critical-band masking (CBM): after training, we added bandlimited noise to test recordings; we parametrically varied the noise band center frequency and noise intensity, quantifying the deleterious effect on performance. We reconciled the effects of CBM with lesioning, wherein we zeroed the trained network's 1st-layer filter kernels in turn, quantifying the deleterious effect on performance. These analyses indicated that the network learned frequency-band information consistent with known SA biomarkers, specifically, delta and beta band activity. Our results indicate single-channel EEG may have clinical potential for SA diagnosis.

Identifying the lungs as a susceptible site for allele-specific regulatory changes associated with type 1 diabetes risk.

Type 1 diabetes (T1D) etiology is complex. We developed a machine learning approach that ranked the tissue-specific transcription regulatory effects for T1D SNPs and estimated their relative contributions to conversion to T1D by integrating case and control genotypes (Wellcome Trust Case Control Consortium and UK Biobank) with tissue-specific expression quantitative trait loci (eQTL) data. Here we show an eQTL (rs6679677) associated with changes to AP4B1-AS1 transcript levels in lung tissue makes the largest gene regulatory contribution to the risk of T1D development. Luciferase reporter assays confirmed allele-specific enhancer activity for the rs6679677 tagged locus in lung epithelial cells (i.e. A549 cells; C > A reduces expression, p = 0.005). Our results identify tissue-specific eQTLs for SNPs associated with T1D. The strongest tissue-specific eQTL effects were in the lung and may help explain associations between respiratory infections and risk of islet autoantibody seroconversion in young children.

Machine Learning Identifies Six Genetic Variants and Alterations in the Heart Atrial Appendage as Key Contributors to PD Risk Predictivity.

Parkinson's disease (PD) is a complex neurodegenerative disease with a range of causes and clinical presentations. Over 76 genetic loci (comprising 90 SNPs) have been associated with PD by the most recent GWAS meta-analysis. Most of these PD-associated variants are located in non-coding regions of the genome and it is difficult to understand what they are doing and how they contribute to the aetiology of PD. We hypothesised that PD-associated genetic variants modulate disease risk through tissue-specific expression quantitative trait loci (eQTL) effects. We developed and validated a machine learning approach that integrated tissue-specific eQTL data on known PD-associated genetic variants with PD case and control genotypes from the Wellcome Trust Case Control Consortium. In so doing, our analysis ranked the tissue-specific transcription effects for PD-associated genetic variants and estimated their relative contributions to PD risk. We identified roles for SNPs that are connected with INPP5P, CNTN1, GBA and SNCA in PD. Ranking the variants and tissue-specific eQTL effects contributing most to the machine learning model suggested a key role in the risk of developing PD for two variants (rs7617877 and rs6808178) and eQTL associated transcriptional changes of EAF1-AS1 within the heart atrial appendage. Similarly, effects associated with eQTLs located within the Brain Cerebellum were also recognized to confer major PD risk. These findings were replicated in two additional, independent cohorts (the UK Biobank, and NeuroX) and thus warrant further mechanistic investigations to determine if these transcriptional changes could act as early contributors to PD risk and disease development.

Ankle torque forecasting using time-delayed neural networks.

A method for ankle torque prediction ahead of the current time is proposed in this paper. The mean average value of EMG signals from four muscles, alongside the joint angle and angular velocity of the right ankle, were used as input parameters to train a time-delayed artificial neural network. Data collected from five healthy subjects were used to generate the dataset to train and test the model. The model predicted ankle torque for five different future times from zero to 2 seconds. Model predictions were compared to torque calculated from inverse dynamics for each subject. The model predicted ankle torque up to 1 second ahead of time with normalized root mean squared error of less than 15 percent while the coefficient of determination was over 0.85.Clinical Relevance- the potential of the model for predicting joint torque ahead of time is helpful to establish an intuitive interaction between human and assistive robots. This model has application to assist patients with neurological disorders.

Healthcare pathway discovery and probabilistic machine learning.

BACKGROUND AND PURPOSE: Healthcare pathways define the execution sequence of clinical activities as patients move through a treatment process, and they are critical for maintaining quality of care. The aim of this study is to combine healthcare pathway discovery with predictive models of individualized recovery times. The pathway discovery has a particular emphasis on producing pathway models that are easy to interpret for clinicians without a sufficient background in process mining. The predictive model takes the stochastic volatility of pathway performance indicators into account. METHOD: This study utilizes the business process-mining software ProM to design a process mining pipeline for healthcare pathway discovery and enrichment using hospital records. The efficacy of combining learned healthcare pathways with probabilistic machine learning models is demonstrated via a case study that applies the proposed process mining pipeline to discover appendicitis pathways from hospital records. Machine learning methodologies based on probabilistic programming are utilized to explore pathway features that influence patient recovery time. RESULTS: The produced appendicitis pathway models are easy for clinical interpretation and provide an unbiased overview of patient movements through the treatment process. Analysis of the discovered pathway model enables reasons for longer than usual treatment times to be explored and deviations from standard treatment pathways to be identified. A probabilistic regression model that estimates patient recovery time based on the information extracted by the process mining pipeline is developed and has the potential to be very useful for hospital scheduling purposes. CONCLUSION: This study establishes the application of the business process modelling tool ProM for the improvement of healthcare pathway mining methods. The proposed pipeline for healthcare pathway discovery has the potential to support the development of probabilistic machine learning models to further relate healthcare pathways to performance indicators such as patient recovery time.