Browsing by Author "Forkert, Nils Daniel"

Now showing 1 - 16 of 16
  • Thumbnail Image
    ItemOpen Access
    Accelerated Quantitative Magnetization Transfer (qMT) Imaging
    (2018-10-24) Mclean, Melany Ann; Pike, G. Bruce; Forkert, Nils Daniel; Lebel, Robert Marc
    Quantitative magnetization transfer (qMT) is an advanced magnetic resonance imaging (MRI) technique with enhanced specificity to myelin. The acquisition of many images with unique magnetization transfer (MT) saturation results in a signal response curve known as the z-spectrum. The two-pool tissue model, which describes properties of nuclei with free and restricted motion, can be fit to the z-spectrum to provide details of macromolecular tissue content (including myelin) beyond what can be seen from conventional single saturation approaches (e.g. MT ratio). Widespread use of qMT has been hindered by long acquisition times inherent to z-spectrum-based imaging techniques including qMT and chemical exchange saturation transfer (CEST). This thesis uses sparseSENSE, a combined parallel imaging and compressed sensing technique, to accelerate MT-weighted images. In this thesis, sparsifying reconstruction algorithms are shown to enable high-quality image reconstruction from 4D qMT datasets, retrospectively under-sampled by factors of up to 32. MT-weighted images demonstrate exceptional image quality at high acceleration factors, which is shown to translate well to accelerated z-spectra. However, qMT parametric maps produced from accelerated z-spectra are shown to be sensitive to acceleration artifacts and can only be accelerated by a factor of 4 with minimal loss of image quality. Nonetheless, this acceleration can yield a significant acquisition time savings when applied to prospectively under-sampled data. In addition, time savings created by acceleration can be used to increase spatial resolution or collect more MT-weighted images, enabling even higher acceleration factors. Long acquisition times have often been cited as a limitation of the qMT method. This work has addressed that limitation, making qMT protocols more feasible for in vivo research studies, particularly in youth and patient populations.
  • Thumbnail Image
    ItemOpen Access
    Acute Ischemic Stroke Analysis Using Deep Learning-based Image-to-image Translation
    (2023-08) Gutierrez Munoz, Jose Alejandro; Forkert, Nils Daniel; Pike, G. Bruce; Lee, Joon; LeVan, Pierre; Almekhlafi, Mohammed
    Acute ischemic stroke occurs due to the sudden occlusion of a cerebral artery, leading to a disruption in metabolic homeostasis and cell damage. Accurate diagnosis and informed treatment decision-making rely on clinical assessments accompanied by medical imaging. Deep learning methods offer the potential to enhance this decision-making by enabling complex pattern recognition. However, they often rely on large amounts of data, which poses a challenge in stroke centers due to the diverse range of imaging modalities employed. Moreover, specialized processing is often required for the meaningful interpretation of valuable imaging methods like perfusion imaging. Recent advancements in deep learning have made it easier to process and analyze perfusion data by predicting the follow-up tissue outcome. However, these models rely on manual binary lesion segmentations as prediction targets, which may hinder interpretability and limit the amount of available data. To address these limitations, the work described in this thesis utilizes a set of deep learning techniques known as image-to-image translation networks in two distinct ways. First, a method was developed to simulate computed tomography datasets based on magnetic resonance imaging scans and vice versa. The results showed that the proposed approach produces realistic outputs, effectively changing the modality while preserving stroke lesions and brain morphology in follow-up scans. This increases the availability of single-modality data and provides an alternative imaging option for follow-up stroke evaluation. Second, a method was developed to predict stroke tissue outcomes from perfusion scans, without relying on manual lesion segmentations and predicting the follow-up image instead. The results show that the proposed method is able to capture the effects of different treatments, highlighting its potential as a tool for treatment guidance or efficacy evaluation. In conclusion, the application of image-to-image generative modelling proves to be valuable for enhancing acute ischemic stroke analysis and care.
  • Thumbnail Image
    ItemOpen Access
    The assessment of fragility fracture risk using HR-pQCT as a novel tool for diagnosis of osteoporosis
    (2021-08) Whittier, Danielle Elizabeth Wein; Boyd, Steven Kyle; Schneider, Prism Steorra; Manske, Sarah Lynn; Edwards, William Brent; Forkert, Nils Daniel; Hallgrimsson, Benedikt; Jepsen, Karl
    Osteoporosis is a systemic skeletal disease, characterized by reduced bone density and deterioration of bone microarchitecture, leading to increased fracture risk. However, current diagnosis using dual-energy X-ray absorptiometry (DXA) only accounts for density and consequently fails to capture most individuals who fracture. High-resolution peripheral quantitative computed tomography (HR-pQCT) is a medical imaging modality capable of characterizing three-dimensional bone microarchitecture at peripheral skeletal sites, and has demonstrated that bone microarchitecture can improve prediction of fracture risk. However, to date the improvement is modest, as interpretation of the interaction between fracture and the numerous parameters provided by HR-pQCT is complex. The objective of this dissertation was to elucidate the key microarchitectural characteristics that underpin bone fragility, and use these insights to improve assessment of fracture risk with HR-pQCT. First, reference data in the form of centile curves was established for HR-pQCT parameters using a population-based cohort (n=1,236, age 18–90 years), and a new intuitive parameter called void space was developed to capture localized regions of bone loss in HR-pQCT images. In a separate prospective multi-center cohort (n=5,873, age 40–90 years), unsupervised machine learning was implemented to identify common groupings (i.e., phenotypes) of bone microarchitecture in older adults. Three phenotypes were identified and characterized as low density, structurally impaired, and healthy bone, where the low density phenotype had the strongest association with incident osteoporotic fractures (hazard ratio = 3.28). Using the same cohort, a fracture risk assessment tool, called µFRAC, was developed using supervised machine learning methods to provide a 5-year risk of major osteoporotic fracture based on HR-pQCT parameters, and was demonstrated to significantly outperform DXA in predicting fracture risk. Finally, a new retrospective cohort of patients with fragility fractures at the hip (n=108, age 56–96 years) was used to characterize bone fragility. Hip fracture patients were significantly associated with the low density phenotype and had bone void spaces that were 2–3 times larger than controls. Together, these findings provide insight into the characteristics of bone that lead to osteoporotic fractures and introduces tools that enable insightful interpretation of HR-pQCT data for clinical use.
  • Thumbnail Image
    ItemOpen Access
    Automated Performance Assessment of Virtual Temporal Bone Dissection
    (2020-07-21) Sachan, Surbhi; Chan, Sonny; Alim, Usman R.; Boyd, Jeffrey Edwin; Forkert, Nils Daniel
    Mastoidectomy is a surgical procedure in which a portion of the temporal bone is removed by using fine microsurgical skills. Development of virtual reality simulators with high-fidelity visual, auditory, and force feedback has allowed trainees to learn this skill in a safe environment without the limitations associated with the traditional way of learning, i.e., cadaveric specimens. However, without an automatic feedback mechanism, an expert's presence is required to assess the performance, placing a heavy burden on their time. This investigation focuses on automating the performance evaluation obviating the need for an expert's time. This is accomplished by automating the criteria based on a well-established and validated assessment instrument known as the Welling Scale, to score the mastoidectomy performed on a virtual surgery simulator. Image processing algorithms are devised and run on the output of the virtual surgery to automatically score these criteria. The criteria are described in terms of four functional categories: Identification, Skeletonization, Intactness and No cells. Algorithms are devised for each of these categories. This work further validates the accuracy of these algorithms by doing a study where these criteria are evaluated by two experts, as well as the work done in this thesis. The results of the study show that automatic performance assessment of virtual mastoidectomy surgery is feasible.
  • Thumbnail Image
    ItemOpen Access
    Automated Video-Based Rodent Behavior Analysis
    (2024-01-30) Le, Van Anh; Murari, Kartikeya; Forkert, Nils Daniel; Yanushkevich, Svetlana; Bento, Mariana Pinheiro; Ravichandran, Avinash
    Rodents represent more than 95% of the laboratory animals used in preclinical and neuroscience research. Mouse behavior analysis is an important step to evaluate disease states and normal brain processes. This thesis focuses on developing automatic video-based mouse behavior analysis tools, which allow high throughput assessments and alleviate the limitations of manual analysis. Particularly, we investigated multiple machine-learning based approaches to fill the gaps of existing studies regarding rodent behavior measurements and create reliable computer-assisted frameworks. Firstly, we introduced MaSoMoTr which is a markerless mice tracking tool for social experiments. The tracking workflow incorporated deep-learning-based techniques with conventional handcrafted tracking methods to simultaneously track two mice of the same appearance in controlled settings. The proposed method achieved significant improvement compared to the state-of-the-art pose-estimation-based tracking frameworks. Following that, we developed a social behavior recognition system integrating our tracking tool to identify a set of mouse behaviors in continuous videos recording two interacting mice. Datasets collected and annotated during these two studies have been made publicly available for further research and development. Finally, two approaches were proposed for automatically recognizing single mouse behaviors in two different settings. We investigated the possibility of extracting spatio-temporal features from single mouse recordings using a deep learning structure which combined a 3D convolutional network and a recurrent neural network with Long Short-Term Memory cells. These extracted features were tested to recognize 8 single mouse behaviors in videos belonging to the largest public single mouse dataset and attained promising performance. Next, we proposed a noninvasive video-based method for mouse sleep assessment. The results obtained were highly correlated with commonly used invasive methods
  • Thumbnail Image
    ItemOpen Access
    Automatic Classification of Idiopathic Parkinsonian Disease and Progressive Supranuclear Palsy using Multi-Spectral MRI Datasets: A Machine Learning Approach
    (2018-09-19) Talai, Aron Sahand; Forkert, Nils Daniel; Monchi, Oury; Chan, Sonny
    Parkinson's disease, which is characterized by a range of motor and non-motor symptoms is categorized into classical Parkinsonian disease (PD) and atypical Parkinsonian syndromes (APS), such as progressive supranuclear palsy Richardson’s syndrome (PSP-RS). The differential diagnosis between PD and PSP-RS is often challenged by similarity of early symptoms, effectively resulting in considerable misclassification rates. The aim of this thesis is to assess the benefits of using biomarkers from multi-modal MRI datasets in the accurate classification of PD vs. PSP-RS. Multi-spectral information form T1-, T2-, and diffusion-weighted (DWI) MRI from 38 healthy controls (HC), 45 PD, and 20 PSP-RS subjects were available for this study. In detail, morphological (category 1), brain iron marker (category 2), and diffusion features (category 3) were employed. In the last category, all feature types were combined (combinational) for the development of a machine learning model. Nested leave-one-out-cross validation was used to evaluate the classification performance in each category followed by a 1000 permutation test to assess classification significance. The results suggest that, the DWI based classifier tied with the combinational approach in terms of overall accuracy. However, in the former, the specificity was lower by 10%. In detail, 4 PSP-RS and 1 PD subjects are incorrectly classified as PD and PSP-RS in the combinational approach resulting in a sensitivity and specificity of 91.67% and 94.12%, respectively. The obtained results indicate that features extracted from T1- and T2-weighted MRI perform worst based on overall accuracy. All classification categories were statistically significant (p<0.001). In conclusion, combination of features from different MRI modalities such as T1-, T2-, and diffusion-weighted datasets improves the multi-level classification performance of HC vs. PD vs.PSP-RS compared to single modality features, particularly in terms of PD vs. other differentiation. The results and concepts discussed in this research thesis have wide ranging implication for future developments of computer-aided diagnosis of PD sub-syndromes.
  • Thumbnail Image
    ItemOpen Access
    Bone as an Orientable, Smooth Surface: Distance Transforms, Morphometry, and Adaptation
    (2021-08) Besler, Bryce Albert Alphonsus; Boyd, Steven Kyle; Fear, Elise Carolyn; Forkert, Nils Daniel; Manske, Sarah Lynn; Cooper, David Michael Lane; Nielsen, Jorgen S
    Age-related changes in bone fundamentally occur at the surface. Understanding and modeling these changes is the primary means of understanding and preventing age-related fractures. However, this is a challenging task, as the bone microarchitecture changes topology during adaption when rods disconnect and plates form holes. The primary objective is to handle topological changes mathematically and develop computational methods for the simulation of bone adaptation. This thesis develops a model of age-related bone loss based on the axioms that the bone surface is orientable and smooth. First, a novel artifact is discovered and described for the distance transform of sampled signals that limits their applicability in simulation and morphometry. Second, a new transform is defined termed the ``high-order signed distance transform'' that is better than the so-called exact signed distance transform in the sense that it has an order of accuracy greater than one. However, this transform does not permit a unique solution on sampled binary images, and another method is needed. Third, an algorithm is presented for computing the unique, high-order signed distance transform of biphasic materials from computed tomography data. Fourth, a method of performing morphometry on closed surfaces is described that relates existing global bone morphometric techniques to local curvature values. This method works on binary images without the need for signed distance transforms when small changes in the bone volume are permitted. Finally, the morphometry and high-order signed distance transform are integrated into a model of age-related bone loss. Principally, this work establishes bone adaptation as a geometric flow, simulated using level set methods that are efficient and naturally handle topological changes. The contribution of this thesis is the establishment of a strong mathematical foundation for modeling bone adaptation. High accuracy computational methods are defined to integrate the theory into practice. The theory and methods form a rigorous basis for biological theories of bone adaptation and provide techniques for measuring and falsifying theories.
  • Thumbnail Image
    ItemOpen Access
    Combined Segmentation and Hemodynamic Analysis of Cerebrovascular Structures using Arterial Spin Labeling
    (2020-01-17) Phellan Aro, Renzo; Forkert, Nils Daniel; Frayne, Richard; Walker, Richard E. A.; Lebel, Robert Marc; Far, Behrouz H.; Duong, Luc
    Spatiotemporal arterial spin labeling magnetic resonance angiography (4D ASL MRA) is a non-invasive imaging modality used to acquire dynamic images of cerebrovascular structures. It can achieve high spatial and temporal resolution, while capturing morphological and blood flow data. Recent scientific studies have used 4D ASL MRA to analyze the cerebrovascular system for characterization, diagnosis, and post-treatment assessment of different cerebrovascular diseases, such as aneurysms, arteriovenous malformations, and moyamoya disease. However, this image sequence generates a considerable amount of data, which can be tedious to analyze by simple visual inspection, a problem also present with other 4D imaging methods. In this case, medical image processing methods can be used to extract the morphological and blood flow data contained in 4D ASL MRA datasets and present it in a more useful format to clinicians and researchers. The aim of this work was to develop and evaluate novel image processing methods for advanced analysis of 4D MRA datasets. The overreaching idea for the development of the corresponding methods is to use blood flow information for improving the vessel segmentation while the vessel segmentation is used to improve the results of the hemodynamic analysis. It was hypothesized that this combined analysis improves the vessel segmentation and results the hemodynamic analysis at the same time. The methods were developed and evaluated using 15 datasets of healthy volunteers, flow phantom measurements, and two datasets of patients with a stenosis. The findings of this work indicate that the proposed combined segmentation and hemodynamic analysis can improve the overall accuracy of the segmentation and blood flow parameter estimation while first experiments also show that the proposed methods can be applied to patients with a cerebrovascular disease. The methods developed in this work could help translating 4D ASL MRA datasets into clinical practice and support clinical research of various cerebrovascular diseases using 4D ASL MRA while the developed methods have also the potential to be useful for other 4D imaging sequences.
  • Thumbnail Image
    ItemOpen Access
    Deep learning methods for classifying disease subtypes in multiple sclerosis based on clinical imaging and non-imaging data
    (2024-09-23) Soleymani, Mahshid; Zhang, Yunyan; Bento, Mariana Pinheiro; Forkert, Nils Daniel; MacDonald, Matthew Ethan
    Multiple sclerosis (MS) is a common inflammatory demyelinating and neurodegenerating disease of the central nervous system impacting over 2.8 million people worldwide. Most people start MS with a relapsing-remitting form (RRMS), yet no two persons have the same disease course. Many of them will develop a secondary-progressive course (SPMS) despite treatment, causing dramatic health and socioeconomic consequences. Early accurate measurement of disease activity will permit early effective treatment for improved prognosis. But there is no established method to classify these two subtypes beforehand clinically. By leveraging the power of deep learning such as convolutional neural networks (CNNs), this project aims to optimize personalized disease characterization using standard clinical data especially brain magnetic resonance imaging (MRI). Specifically, based on 140 clinical participants with RRMS or SPMS, the research targets phenotype prediction through a series of development and validation processes. These included data optimization, model development and testing based on both 2D- and 3D- CNN models, and model interpretation using a recognized method called gradient-class activation mapping (Grad- CAM). Results showed that axial images normalized with a Z-score like approach were most feasible. Both the 2D and 3D models achieved >80% accuracy in predicting RRMS and SPMS, where combining both MRI and clinical variables appeared to perform better than either data type alone. The Grad-CAM analysis helped discern critical brain areas related to each MS subtype. These findings underscore the potential of deep learning based completely on clinical care data to detect disease activity, marking early diagnosis and personalized treatment possible.
  • Thumbnail Image
    ItemEmbargo
    Distributed Deep Learning Methods for Medical Imaging Analysis
    (2024-10-29) Souza De Andrade, Raissa Cristina; Forkert, Nils Daniel; Wilms, Matthias; Pike, G. Bruce; Barker, Ken
    Recent advancements in deep learning have equipped healthcare professionals with valuable tools to support clinical decision-making and reduce workloads. However, many medical centers lack sufficient datasets to train deep learning models, especially for rare diseases or centers in remote or underserved areas. Although collecting and curating datasets from multiple centers into a centralized repository is commonly employed to solve this problem, this approach is often infeasible due to its costs and privacy regulations that prohibit data sharing. Consequently, many centers and populations will not benefit from cutting-edge artificial intelligence. The distributed deep learning framework proposed in this work addresses these challenges by training accurate models while patient data remains securely stored within its center. Thus, privacy concerns are addressed while collaborative multi-center training is facilitated. A key innovation of this work is the development and evaluation of the travelling model, a method well-suited for scenarios where individual centers have very limited data availability. The travelling model is evaluated across various scenarios, including extreme cases where centers contribute only a single medical image, and is applied to critical medical imaging tasks such as brain age prediction, disease classification, and tumour segmentation. In general, the travelling model effectively increases the overall dataset quantity and diversity without compromising patient data privacy. However, solutions for the inherent acquisition shift biases caused by variations in equipment and protocols across centers and decentralized data quality control are needed to leverage its full potential. Therefore, this work also developed and integrated two novel methods into the travelling model approach, a data harmonization for reducing acquisition shift biases and automated decentralized data quality control. The results of this work demonstrate that the travelling model framework achieved performances comparable to models trained on a centralized repository across all evaluated tasks. Moreover, it performed better than the commonly used federated learning in cases where centers contributed fewer than five datasets. Additionally, the proposed data harmonization method reduced scanner variability by 23%, improving disease classification accuracy by 4%. Finally, the automated decentralized quality control method effectively identified incorrect and low-quality datasets, enabling more robust and reliable model performance.
  • Thumbnail Image
    ItemOpen Access
    Explainable prediction of Parkinson’s disease in a large multimodal database
    (2023-09-08) Camacho Camacho, Milton Ivan; Forkert, Nils Daniel; Monchi, Oury; Ismail, Zahinoor; LeVan, Pierre
    Parkinson’s disease (PD) is the second most common neurodegenerative disease affecting millions of people all over the world. Accurate diagnosis is important to enable prompt interventions that can improve disease prognosis. However, the heterogeneity of PD renders an accurate diagnosis challenging, especially early in the disease phase when symptoms are known to be subtle. Thus, it is imperative to obtain reliable, non-invasive, in-vivo biomarkers for PD diagnosis. Within this context, the overarching aim of this work is to develop accurate explainable deep learning models trained from a large multimodal magnetic resonance imaging (MRI) database to classify PD and healthy controls and capable of identifying structural changes associated with PD. Therefore, the objectives of this thesis are: (1) to investigate the use of T1-weighted brain MRI as a biomarker of macro-structural changes associated with PD; (2) to investigate a combination of T1-weighted and diffusion tensor MRI as a fusion of micro- and macro-structural brain morphology and its relationship with PD. To achieve these objectives, one of the largest multi-center imaging databases of over 2,000 PD and control subjects was pooled and preprocessed as the first step. Second, an explainable deep learning model was developed to accurately classify PD patients while revealing important brain regions. Third, a multimodal explainable deep learning model was trained enabling a more in-depth understanding of the interplay between the micro- and macro-structural properties of specific brain regions and the disease. The results of this work offer an important insight into structural brain changes as non-invasive, in-vivo biomarkers of PD through an in-depth analysis of associated brain regions using large multicenter data. Deep learning models are proposed to provide generalizable and robust PD classification resulting in a 0.87 and 0.89 area under the receiver operating characteristic curve for the best unimodal and multimodal approaches, respectively. Lastly, explainability methods identified brain regions in line with current knowledge of the disease providing further evidence for the clinical utility of the developed methods. This work presents relevant findings and novel methodologies that could aid improving PD diagnosis and the acceptance of computer-aided diagnosis systems targeting PD.
  • Thumbnail Image
    ItemOpen Access
    Multimodal Imaging of Cortical Networks Controlling Lower Limb Locomotion: Towards the Development of Brain-Computer Interfaces
    (2018-07-11) Kline, Adrienne; Ronsky, Janet L.; Goodyear, Bradley G.; Forkert, Nils Daniel; Syed, Naweed L.
    In 2015 the National Spinal Cord Injury Association of Canada reported that 30,000 Canadians suffer from paralysis in two or more limbs. In many cases this takes away the fundamental ability to walk. Walking, an intricate sensorimotor task, involves the interactions of both dynamic and balancing neurological processes. Brain computer interfaces (BCIs) are attempting to bridge the gap that will allow persons with compromised mobility to interact with the world via control of prosthetic devices that can ‘act’ by using solely neural input (i.e. thoughts). The goal of this thesis was to aid in the development of a BCI for lower limb locomotion by identifying similarities and differences between cortical activity associated with executed and imagined left and right lower limb movements using electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI). Data from 16 participants showed that it was possible to differentiate between right versus left executed and imagined thought processes for lower limb locomotion using solely information from an EEG, and that these patterns of brain activity were generalizable across time points and trials. It was also found, through the use of fMRI, that areas of brain activation in executed and imagined conditions were similar for some areas but showed unique activation areas as well. A novel paradigm to co-register EEG and fMRI data was developed that can easily be utilized in other contexts. Finally, using EEG and fMRI data allowed for an efficient model to use in a machine learning paradigm that successfully predicted left versus right lower limb movement. This research adds to the existing body of knowledge in understanding psychomotor brain activity associated with thought coordination processes involved in the task of walking in normal persons represented by algorithmic patterns.
  • Thumbnail Image
    ItemOpen Access
    Optimal Control of Nonlinear Networks Dynamics with Applications to Brain Stimulation in Alzheimer's Disease
    (2017-12-20) Sánchez Rodríguez, Lázaro Miguel; Sotero-Diaz, Roberto Carlos; Monchi, Oury; Forkert, Nils Daniel; Kiss, Zelma; Vasudevan, Kris
    Brain stimulation can modulate the activity of neural circuits impaired by Alzheimer’s disease (AD), having promising clinical benefit. However, all individuals with the same condition currently receive identical brain stimulation, with limited theoretical basis for this generic approach. In this study, we introduce a control theory framework for obtaining exogenous signals that revert pathological electroencephalographic activity in AD at a minimal energetic cost, while reflecting patients’ biological variability. By considering nonlinearities in our model, we identified regions for which control inputs fail to correct abnormal activity. We also found that limbic system and basal ganglia structures constitute the top target locations for stimulation in AD. Patients with highly integrated anatomical networks are the most suitable candidates for the propagation of stimuli and consequent success on the control task. Other diseases associated with alterations in brain dynamics and the self-control mechanisms of the brain can be addressed through our framework.
  • Thumbnail Image
    ItemOpen Access
    Predicting the Side Effects of Antiseizure Medications Using Machine Learning Models
    (2024-01-02) Lin, Chantelle Qing Yang; Josephson, Colin Bruce; Sajobi, Tolulope; Klein, Karl Martin; Forkert, Nils Daniel
    With over 20 anti-seizure medications (ASMs), identifying the ideal drug is often imprecise and time-consuming. Developing predictive models to expedite optimal drug selection is challenging due to the minimal differences in efficacy among adult patients with epilepsy. However, side-effects vary considerably between medications, and are one of the main reasons for discontinuation of ASM treatment. The aim was to (1) assess the prognostic utility of high- dimensional data such as genetic features with clinical features to predict ASM discontinuation, and (2) determine the optimal regression/machine learning model for predicting ASM discontinuation. This retrospective cohort study included 4,853 exposures to any ASM, and 624 patients exposed to valproic acid (VPA) from the RAISE-GENIC study during the years 2006-2020. The predicted outcome was defined as ASM discontinuation due to any side-effect reported by the patient. Clinical features included age of onset, patient age, sex, comorbidities, seizure type, EEG variables, and imaging variables. Network analysis of mRNA expression data from VPA-exposed neurons derived from control induced pluripotent stem cells (iPSCs) was leveraged to extract exome sequencing and genome-wide single nucleotide polymorphism data. Features were selected for model inclusion based on relevance as determined by the ReliefF algorithm. Penalized logistic regression, support vector machine, random forest, and k-nearest neighbor models were trained on the normalized bootstrapped dataset and model quality was assessed using stratified 10-fold cross validation. Models with only clinical and combined clinical and genetic features were compared by quantitative as well as visual discrimination and calibration metrics. The results showed that the best performing model was the penalized logistic regression using the VPA dataset with genetic and clinical features. The accuracy was 0.75 [95% confidence interval 0.74-0.76], area under the receiver operating characteristic curve was 0.66 [0.66-0.67], Brier score was 0.20 [0.19-0.21], sensitivity was 0.42 [0.41-0.42], and specificity 0.82 [0.82-0.83]. Machine learning using clinical and genetic features can moderately predict treatment-ending side- effects to VPA with moderate performance, discrimination, and calibration. If these results can be validated and improved upon, decision tools can be incorporated into clinical routines, simplifying drug prescriptions, saving time, and improving patient quality of life.
  • Thumbnail Image
    ItemEmbargo
    Using Domain Adaptation and Inductive Transfer Learning to Improve Patient Outcome Prediction in the Intensive Care Unit
    (2023-12-07) Mutnuri, Maruthi Kumar; Lee, Joon; Stelfox, Thomas; Forkert, Nils Daniel; MacDonald, Matthew Ethan; Parhar, Ken Kuljit
    Predicting patient outcomes in the intensive care unit (ICU) can allow appropriate allocation of resources, minimize costs, and provide better patient care. Machine learning and deep learning models can predict patient outcomes with a high degree of accuracy, but training those models is both data- and resource-intensive. Deep learning models trained on small datasets tend to overfit and generalize poorly, and transfer learning (TL) helps in such situations by leveraging the knowledge learned from pre-trained models. Transfer learning is a machine learning technique where a model pre-trained on source task is adapted for a different but related target task. Here, source task is trained with a large dataset whereas a small dataset is sufficient for training target task. Notably, TL is widely used in medical image analysis and natural language processing, but it is uncommon in electronic health record (EHR) analysis. Within the TL literature, domain adaptation (DA) is most common, whereas inductive transfer learning (ITL) is rare. This study explores both DA and ITL using EHR data. To investigate the effectiveness of these TL models, we compared them with baseline models of logistic regression (LR), lasso regression, and fully connected neural networks (FCNN) in the prediction of 30-day mortality, acute kidney injury (AKI), hospital length of stay (H_LOS), and ICU length of stay (ICU_LOS). We used two cohorts: (1) eCritical, a multicenter ICU data linked with administrative data from ICUs in Alberta, Canada between March 2013 and December 2019, which has 55,689 unique admission records from 48,672 unique patients admitted to 15 medical-surgical ICUs, and (2) MIMIC-III, a single-center publicly available ICU dataset from Boston, USA between 2001 and 2012. The first admission of adult patient records with more than 24-hour ICU stays were included in this retrospective study. We included common features from both the cohorts. Random data subsets of training data, ranging from 1% to 75%, and the full dataset were used to compare the performances of DA and ITL with FCNN, LR, and lasso. Overall, ITL outperformed baseline FCNN, LR, and lasso models in 55 of the 56 comparisons (7 data subsets, 4 outcomes, and 2 baseline models) using BA and MSE metrics. However, DA models outperformed the baseline models 45 out of the 56 times. ITL performance was comparatively better than DA considering the number of times it outperformed baseline models and the margin with which it outperformed baseline models. Also, in 11 out of the 16 cases (8 of 8 for ITL and 3 of 8 for DA) TL models outperformed baseline models at 1% data subset. This is significant because TL models are useful in data-scarce scenarios. When using EHR data, the similarity of data distributions in source and target domains was crucial, as evident from ITL performing much better than DA, mostly because of the domain mismatch in the two cohorts concerning AKI, H_LOS, and ICU_LOS outcomes. As the pre-trained models will be made available publicly, further research can be conducted with additional outcomes and different cohorts to make these pre-trained models more robust using incremental or cumulative transfer learning. These pre-trained models can be used for predicting patient outcomes at ICU.
  • Thumbnail Image
    ItemOpen Access
    Using Machine Learning for Prognostication of Diagnosis and Identifying Neural Correlates of Impulse Dyscontrol in Preclinical and Prodromal Dementia
    (2019-08-22) Gill, Sascha Charlene; Ismail, Zahinoor; Smith, Eric Edward; Forkert, Nils Daniel; MacMaster, Frank P.
    Introduction: Mild Behavioural impairment (MBI) is a validated syndrome that describes neuropsychiatric symptoms (NPS) in preclinical and prodromal dementia. This thesis uses machine learning (ML) and traditional statistical models to: 1) Explore the utility of NPS for predicting diagnostic status 2) Identify the neural correlates of MBI impulse dyscontrol (ID) domain. Methods: Data from the Alzheimer’s Disease Neuroimaging (ADNI) database were extracted. All subjects enrolled in ADNI were between the age of 55-90 years, English or Spanish speakers, and accompanied by study partners who completed the NPI-Q 1) To address the first objective, the logistic model tree classifier combined with an information gain feature selection was trained to predict follow-up diagnosis (normal cognition [NC], MCI, or AD-dementia) using baseline neuroimaging, neuropsychiatric, and demographic data. 2) To address the second objective, ID was identified as behavioural symptoms of agitation/aggression, irritability, and aberrant motor behaviour. Linear mixed effect models were used to assess if ID was related to regional diffusion tensor imaging (DTI) and volumetric parameters. Additionally, ML modeling used a rule-based classification algorithm combined with an information gain feature selector to predict ID using neuroimaging variables. Results: 1) MBI total scores and volume of the left hippocampus were identified as the most important features to predict follow-up diagnostic status. 2) Cingulum, fornix, inferior/superior fronto-occipital fasciculus, superior cerebellar peduncle, and corpus callosum, were the white matter tracts associated with ID. Grey matter regions associated with ID included the parahippocampal gyrus supramarginal gyrus, superior frontal regions, and hippocampus. Conclusion: NPS are early indicators of neurodegenerative disease and can be used predict cognitive decline and dementia.