Large Clinical Datasets Research Articles

ClusterMLD: An Efficient Hierarchical Clustering Method for Multivariate Longitudinal Data

Longitudinal data clustering is challenging because the grouping has to account for the similarity of individual trajectories in the presence of sparse and irregular times of observation. This paper puts forward a hierarchical agglomerative clustering method based on a dissimilarity metric that quantifies the cost of merging two distinct groups of curves, which are depicted by B-splines for the repeatedly measured data. Extensive simulations show that the proposed method has superior performance in determining the number of clusters, classifying individuals into the correct clusters, and in computational efficiency. Importantly, the method is not only suitable for clustering multivariate longitudinal data with sparse and irregular measurements but also for intensely measured functional data. Towards this end, we provide an R package for the implementation of such analyses. To illustrate the use of the proposed clustering method, two large clinical data sets from real-world clinical studies are analyzed.

Using ICD-9 diagnostic codes for external validation of topic models derived from primary care electronic medical record clinical text data.

Background/Objectives: Unsupervised topic models are often used to facilitate improved understanding of large unstructured clinical text datasets. In this study we investigated how ICD-9 diagnostic codes, collected alongside clinical text data, could be used to establish concurrent-, convergent- and discriminant-validity of learned topic models. Design/Setting: Retrospective open cohort design. Data were collected from primary care clinics located in Toronto, Canada between 01/01/2017 through 12/31/2020. Methods: We fit a non-negative matrix factorization topic model, with K = 50 latent topics/themes, to our input document term matrix (DTM). We estimated the magnitude of association between each Boolean-valued ICD-9 diagnostic code and each continuous latent topical vector. We identified ICD-9 diagnostic codes most strongly associated with each latent topical vector; and qualitatively interpreted how these codes could be used for external validation of the learned topic model. Results: The DTM consisted of 382,666 documents and 2210 words/tokens. We correlated concurrently assigned ICD-9 diagnostic codes with learned topical vectors, and observed semantic agreement for a subset of latent constructs (e.g. conditions of the breast, disorders of the female genital tract, respiratory disease, viral infection, eye/ear/nose/throat conditions, conditions of the urinary system, and dermatological conditions, etc.). Conclusions: When fitting topic models to clinical text corpora, researchers can leverage contemporaneously collected electronic medical record data to investigate the external validity of fitted latent variable models.

The development of the pediatric stroke neuroimaging platform (PEDSNIP)

Childhood stroke occurs from birth to 18 years of age, ranks among the top ten childhood causes of death, and leaves lifelong neurological impairments. Arterial ischemic stroke in infancy and childhood occurs due to arterial occlusion in the brain, resulting in a focal lesion. Our understanding of mechanisms of injury and repair associated with focal injury in the developing brain remains rudimentary. Neuroimaging can reveal important insights into these mechanisms. In adult stroke population, multi-center neuroimaging studies are common and have accelerated the translation process leading to improvements in treatment and outcome. These studies are centered on the growing evidence that neuroimaging measures and other biomarkers (e.g., from blood and cerebrospinal fluid) can enhance our understanding of mechanisms of risk and injury and be used as complementary outcome markers. These factors have yet to be studied in pediatric stroke because most neuroimaging studies in this population have been conducted in single-centred, small cohorts. By pooling neuroimaging data across multiple sites, larger cohorts of patients can significantly boost study feasibility and power in elucidating mechanisms of brain injury, repair and outcomes. These aims are particularly relevant in pediatric stroke because of the decreased incidence rates and the lack of mechanism-targeted trials.Toward these aims, we developed the Pediatric Stroke Neuroimaging Platform (PEDSNIP) in 2015, funded by The Brain Canada Platform Support Grant, to focus on three identified neuroimaging priorities. These were: developing and harmonizing multisite clinical protocols, creating the infrastructure and methods to import, store and organize the large clinical neuroimaging dataset from multiple sites through the International Pediatric Stroke Study (IPSS), and enabling central searchability. To do this, developed a two-pronged approach that included building 1) A Clinical-MRI Data Repository (standard of care imaging) linked to clinical data and longitudinal outcomes and 2) A Research-MRI neuroimaging data set acquired through our extensive collaborative, multi-center, multidisciplinary network. This dataset was collected prospectively in eight North American centers to test the feasibility and implementation of harmonized advanced Research-MRI, with the addition of clinical information, genetic and proteomic studies, in a cohort of children presenting with acute ischemic stroke.Here we describe the process that enabled the development of PEDSNIP built to provide the infrastructure to support neuroimaging research priorities in pediatric stroke. Having built this Platform, we are now able to utilize the largest neuroimaging and clinical data pool on pediatric stroke data worldwide to conduct hypothesis-driven research. We are actively working on a bioinformatics approach to develop predictive models of risk, injury and repair and accelerate breakthrough discoveries leading to mechanism-targeted treatments that improve outcomes and minimize the burden following childhood stroke. This unique transformational resource for scientists and researchers has the potential to result in a paradigm shift in the management, outcomes and quality of life in children with stroke and their families, with far-reaching benefits for other brain conditions of people across the lifespan.

Molecular dynamics analysis of superoxide dismutase 1 mutations suggests decoupling between mechanisms underlying ALS onset and progression

Mutations in the superoxide dismutase 1 (SOD1) gene are the second most common known cause of ALS. SOD1 variants express high phenotypic variability and over 200 have been reported in people with ALS. It was previously proposed that variants can be broadly classified in two groups, ‘wild-type like’ (WTL) and ‘metal binding region’ (MBR) variants, based on their structural location and biophysical properties. MBR variants, but not WTL variants, were associated with a reduction of SOD1 enzymatic activity. In this study we used molecular dynamics and large clinical datasets to characterise the differences in the structural and dynamic behaviour of WTL and MBR variants with respect to the wild-type SOD1, and how such differences influence the ALS clinical phenotype. Our study identified marked structural differences, some of which are observed in both variant groups, while others are group specific. Moreover, collecting clinical data of approximately 500 SOD1 ALS patients carrying variants, we showed that the survival time of patients carrying an MBR variant is generally longer (∼6 years median difference, p < 0.001) with respect to patients with a WTL variant. In conclusion, our study highlighted key differences in the dynamic behaviour between WTL and MBR SOD1 variants, and between variants and wild-type SOD1 at an atomic and molecular level, that could be further investigated to explain the associated phenotypic variability. Our results support the hypothesis of a decoupling between mechanisms of onset and progression of SOD1 ALS, and an involvement of loss-of-function of SOD1 with the disease progression.

3D-QCNet - A pipeline for automated artifact detection in diffusion MRI images.

Artifacts are a common occurrence in Diffusion MRI (dMRI) scans. Identifying and removing them is essential to ensure the accuracy and viability of any post-processing carried out on these scans. This makes quality control (QC) a crucial first step prior to any analysis of dMRI data. Several QC methods for artifact detection exist, however they suffer from problems like requiring manual intervention and the inability to generalize across different artifacts and datasets. In this paper, we propose an automated deep learning (DL) pipeline that utilizes a 3D-Densenet architecture to train a model on diffusion volumes for automatic artifact detection. Our method is validated on 9000 volumes sourced from 7 large clinical datasets spanning different acquisition protocols (with different gradient directions, high and low b-values, single-shell and multi-shell acquisitions) from multiple scanners. Additionally, they represent diverse subject demographics including age, sex and the presence or absence of pathologies. Our QC method is found to accurately generalize across this heterogenous data by correctly detecting 92% artifacts on average across our test set. This consistent performance over diverse datasets underlines the generalizability of our method, which currently is a significant barrier hindering the widespread adoption of automated QC techniques. Thus, 3D-QCNet can be integrated into diffusion pipelines to effectively automate the arduous and time-intensive process of artifact detection.

Theoretical Schemas to Guide BACPAC Chronic Low Back Pain Clinical Research.

Chronic low back pain (cLBP) is a complex with a heterogenous clinical presentation. A better understanding of the factors that contribute to cLBP is needed for accurate diagnosis, optimal treatment, and identification of mechanistic targets for new therapies. The Back Pain Consortium (BACPAC) Research Program provides a unique opportunity in this regard as it will generate large clinical datasets including a diverse set of harmonized measurements. The Theoretical Model Working Group (TMWG) was established to guide BACPAC research, and to organize new knowledge within a mechanistic framework. This article summarizes the initial works of the TMWG. It includes a three-stage integration of expert opinion and an umbrella literature review of factors that affect cLBP severity and chronicity. During Stage 1, experts from across BACPAC established a taxonomy for risk and prognostic factors (RPFs) and preliminary graphical depictions. During Stage 2, a separate team conducted a literature review according to PRISMA guidelines to establish working definitions, associated data elements (ADEs), and overall strength-of-evidence (SOE) for identified RPFs. These were subsequently integrated with expert opinion during Stage 3. The majority (∼80%) of RPFs had little SOE confidence, while 7 factors had substantial confidence for either a positive association (pain-related anxiety, serum c-reactive protein, diabetes, and anticipatory/compensatory postural adjustments) or no association (serum interleukin 1-beta/interleukin 6, transversus muscle morphology/activity, and quantitative sensory testing) with cLBP. This theoretical perspective will evolve over time as BACPAC investigators link empirical results to theory, challenge current ideas of the biopsychosocial model, and utilize a systems approach to develop tools and algorithms that disentangle the dynamic interactions between cLBP factors.

Estimating the In Vivo Function of CYP2D6 Alleles through Population Pharmacokinetic Modeling of Brexpiprazole.

Accurate prediction of CYP2D6 phenotype from genotype information is important to support safe and efficacious pharmacotherapy with CYP2D6 substrates. To facilitate accurate CYP2D6 genotype-phenotype translation, there remains a need to investigate the enzyme activity associated with individual CYP2D6 alleles using large clinical data sets. This study aimed to quantify and compare the in vivo function of different CYP2D6 alleles through population pharmacokinetic (PopPK) modeling of brexpiprazole using data from 13 clinical studies. A PopPK model of brexpiprazole and its two metabolites, DM-3411 and DM-3412, was developed based on plasma concentration samples from 826 individuals. As the minor metabolite, DM-3412, is formed via CYP2D6, the metabolic ratio of DM-3412:brexpiprazole calculated from the PopPK parameter estimates was used as a surrogate measure of CYP2D6 activity. A CYP2D6 genotype-phenotype analysis based on 496 subjects showed that the CYP2D6*2 allele (n=183) was associated with only 10% enzyme activity relative to the wild-type allele (CYP2D6*1) and a low enzyme activity was consistently observed across genotypes containing CYP2D6*2. Among the decreased function alleles, the following enzyme activities relative to CYP2D6*1 were estimated: 23% for CYP2D6*9 (n=20), 32% for CYP2D6*10 (n=62), 64% for CYP2D6*14 (n=1), 4% for CYP2D6*17 (n=37), 4% for CYP2D6*29 (n=13), and 9% for CYP2D6*41 (n=64). These findings imply that a lower functional value would more accurately reflect the in vivo function of many reduced function CYP2D6 alleles in the metabolism of brexpiprazole. The low enzyme activity observed for CYP2D6*2, which has also been reported by others, suggests that the allele exhibits substrate-specific enzyme activity.

DeepHeartCT: A fully automatic artificial intelligence hybrid framework based on convolutional neural network and multi-atlas segmentation for multi-structure cardiac computed tomography angiography image segmentation.

Cardiac computed tomography angiography (CTA) is an emerging imaging modality for assessing coronary artery as well as various cardiovascular structures. Recently, deep learning (DL) methods have been successfully applied to many applications of medical image analysis including cardiac CTA structure segmentation. However, DL requires a large amounts of data and high-quality labels for training which can be burdensome to obtain due to its labor-intensive nature. In this study, we aim to develop a fully automatic artificial intelligence (AI) system, named DeepHeartCT, for accurate and rapid cardiac CTA segmentation based on DL. The proposed system was trained using a large clinical dataset with computer-generated labels to segment various cardiovascular structures including left and right ventricles (LV, RV), left and right atria (LA, RA), and LV myocardium (LVM). This new system was trained directly using high-quality computer labels generated from our previously developed multi-atlas based AI system. In addition, a reverse ranking strategy was proposed to assess the segmentation quality in the absence of manual reference labels. This strategy allowed the new framework to assemble optimal computer-generated labels from a large dataset for effective training of a deep convolutional neural network (CNN). A large clinical cardiac CTA studies (n = 1,064) were used to train and validate our framework. The trained model was then tested on another independent dataset with manual labels (n = 60). The Dice score, Hausdorff distance and mean surface distance were used to quantify the segmentation accuracy. The proposed DeepHeartCT framework yields a high median Dice score of 0.90 [interquartile range (IQR), 0.90-0.91], a low median Hausdorff distance of 7 mm (IQR, 4-15 mm) and a low mean surface distance of 0.80 mm (IQR, 0.57-1.29 mm) across all segmented structures. An additional experiment was conducted to evaluate the proposed DL-based AI framework trained with a small vs. large dataset. The results show our framework also performed well when trained on a small optimal training dataset (n = 110) with a significantly reduced training time. These results demonstrated that the proposed DeepHeartCT framework provides accurate and rapid cardiac CTA segmentation that can be readily generalized for handling large-scale medical imaging applications.

HMOX1 STR polymorphism and malaria: an analysis of a large clinical dataset

BackgroundInducible expression of heme oxygenase-1 (encoded by the gene HMOX1) may determine protection from heme released during malaria infections. A variable length, short tandem GT(n) repeat (STR) in HMOX1 that may influence gene expression has been associated with outcomes of human malaria in some studies. In this study, an analysis of the association between variation at the STR in HMOX1 on severe malaria and severe malaria subtypes is presented in a large, prospectively collected dataset (MalariaGEN).MethodsThe HMOX1 STR was imputed using a recently developed reference haplotype panel designed for STRs. The STR was classified by total length and split into three alleles based on an observed trimodal distribution of repeat lengths. Logistic regression was used to assess the association between this repeat on cases of severe malaria and severe malaria subtypes (cerebral malaria and severe malarial anaemia). Individual analyses were performed for each MalariaGEN collection site and combined for meta-analysis. One site (Kenya), had detailed clinical metadata, allowing the assessment of the effect of the STR on clinical variables (e.g. parasite count, platelet count) and regression analyses were performed to investigate whether the STR interacted with any clinical variables.ResultsData from 17,960 participants across 11 collection sites were analysed. In logistic regression, there was no strong evidence of association between STR length and severe malaria (Odds Ratio, OR: 0.96, 95% confidence intervals 0.91–1.02 per ten GT(n) repeats), although there did appear to be an association at some sites (e.g., Kenya, OR 0.90, 95% CI 0.82–0.99). There was no evidence of an interaction with any clinical variables.ConclusionsMeta-analysis suggested that increasing HMOX1 STR length is unlikely to be reliably associated with severe malaria. It cannot be ruled out that repeat length may alter risk in specific populations, although whether this is due to chance variation, or true variation due to underlying biology (e.g., gene vs environment interaction) remains unanswered.

SurvivalCNN: A deep learning-based method for gastric cancer survival prediction using radiological imaging data and clinicopathological variables

Radiological images have shown promising effects in patient prognostication. Deep learning provides a powerful approach for in-depth analysis of imaging data and integration of multi-modal data for modeling. In this work, we propose SurvivalCNN, a deep learning structure for cancer patient survival prediction using CT imaging data and non-imaging clinical data. In SurvivalCNN, a supervised convolutional neural network is designed to extract volumetric image features, and radiomics features are also integrated to provide potentially different imaging information. Within SurvivalCNN, a novel multi-thread multi-layer perceptron module, namely, SurvivalMLP, is proposed to perform survival prediction from censored survival data. We evaluate the proposed SurvivalCNN framework on a large clinical dataset of 1061 gastric cancer patients for both overall survival (OS) and progression-free survival (PFS) prediction. We compare SurvivalCNN to three different modeling methods and examine the effects of various sets of data/features when used individually or in combination. With five-fold cross validation, our experimental results show that SurvivalCNN achieves averaged concordance index 0.849 and 0.783 for predicting OS and PFS, respectively, outperforming the compared state-of-the-art methods and the clinical model. After future validation, the proposed SurvivalCNN model may serve as a clinical tool to improve gastric cancer patient survival estimation and prognosis analysis.

Application of a convolutional neural network to the quality control of MRI defacing

Large-scale neuroimaging datasets present unique challenges for automated processing pipelines. Motivated by a large clinical trials dataset with over 235,000 MRI scans, we consider the challenge of defacing — anonymisation to remove identifying facial features. The defacing process must undergo quality control (QC) checks to ensure that the facial features have been removed and that the brain tissue is left intact. Visual QC checks are time-consuming and can cause delays in preparing data. We have developed a convolutional neural network (CNN) that can assist with the QC of the application of MRI defacing; our CNN is able to distinguish between scans that are correctly defaced and can classify defacing failures into three sub-types to facilitate parameter tuning during remedial re-defacing. Since integrating the CNN into our anonymisation pipeline, over 75,000 scans have been processed. Strict thresholds have been applied so that ambiguous classifications are referred for visual QC checks, however all scans still undergo an efficient verification check before being marked as passed. After applying the thresholds, our network is 92% accurate and can classify nearly half of the scans without the need for protracted manual checks. Our model can generalise across MRI modalities and has comparable performance when tested on an independent dataset. Even with the introduction of the verification checks, incorporation of the CNN has reduced the time spent undertaking QC checks by 42% during initial defacing, and by 35% overall. With the help of the CNN, we have been able to successfully deface 96% of the scans in the project whilst maintaining high QC standards. In a similarly sized new project, we would expect the model to reduce the time spent on manual QC checks by 125 h. Our approach is applicable to other projects with the potential to greatly improve the efficiency of imaging anonymisation pipelines.

Derivation and validation of a machine learning-based risk prediction model for in-hospital mortality in patients with acute heart failure

Abstract Background Although risk stratification is important in patients with acute heart failure (AHF) to predict patient prognosis, pre-existing risk models have not often been used due to its complexity. Recently, machine learning methods have been presented as an alternative approach to analyzing the predictive probability of large clinical datasets. Purpose The aim of this study is to develop a user-friendly risk score developed by one of machine learning methods and compare the performance of the new risk score to the existing conventional risk models. Methods A machine-learning-based risk model was developed using least absolute shrinkage and selection operator (LASSO) regression by identifying predictors of in-hospital mortality in the derivation cohort (REALITY-AHF) and externally validating and comparing its performance with two pre-existing risk models: the Get With The Guidelines risk score incorporating brain natriuretic peptide and hypochloremia (GWTG-BNP-Cl-RS) and the acute decompensated heart failure national registry (ADHERE) risk model. Results In-hospital deaths in the derivation and validation (NARA-HF) cohorts were 76 (5.1%) and 61 (4.9%), respectively. The risk score comprised four variables (systolic blood pressure, blood urea nitrogen, serum chloride, and C-reactive protein) and was developed according to the results of the LASSO regression weighting the coefficient for selected variables using a logistic regression model (4V-RS). Even though 4V-RS comprised fewer variables, In the validation cohort, it showed a higher area under the receiver operating characteristic curve (AUC) than the ADHERE risk model (AUC, 0.783 vs. 0.740; P=0.059) and a significant improvement in net reclassification (0.359; 95% CI, 0.10–0.67; p=0.006). 4V-RS performed similarly to GWTG-BNP-Cl-RS in terms of discrimination (AUC, 0.783 vs. 0.759; p=0.426) and net reclassification (0.176; 95% CI, −0.08–0.43; p=0.178). Conclusions The 4V-RS model comprising only four readily available data points at the time of admission performed similarly to the more complex pre-existing risk model in patients with AHF. Funding Acknowledgement Type of funding sources: Foundation. Main funding source(s): Cardiovascular Research Fund

A tree-structure-guided graph convolutional network with contrastive learning for the assessment of parkinsonian hand movements.

Bradykinesia is one of the core motor symptoms of Parkinson's disease (PD). Neurologists typically perform face-to-face bradykinesia assessment in PD patients according to the Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS). As this human-expert assessment lacks objectivity and consistency, an automated and objective assessment scheme for bradykinesia is critically needed. In this paper, we propose a tree-structure-guided graph convolutional network with contrastive learning scheme to solve the challenge of difficulty in fine-grained feature extraction and insufficient model stability, finally achieving the video-based automated assessment of Parkinsonian hand movements, which represent a vital MDS-UPDRS component for examining upper-limb bradykinesia. Specifically, a tri-directional skeleton tree scheme is proposed to achieve effective fine-grained modeling of spatial hand dependencies. In this scheme, hand skeletons are extracted from videos, and then the spatial structures of these skeletons are constructed through depth-first tree traversal. Afterwards, a tree max-pooling module is employed to establish remote exchange between outer and inner nodes, hierarchically gather the most salient motion features, and hence achieve fine-grained mining. Finally, a group-sparsity-induced momentum contrast is also developed to learn similar motion patterns under different interference through contrastive learning. This can promote stable learning of discriminative spatial-temporal features with invariant motion semantics. Comprehensive experiments on a large clinical video dataset reveal that our method achieves competitive results, and outperforms other sensor-based and RGB-depth methods. The proposed method leads to accurate assessment of PD bradykinesia through videos collected by low-cost consumer cameras of limited capabilities. Hence, our work provides a convenient tool for PD telemedicine applications with modest hardware requirements.

STAR sleep recording export software for automatic export and anonymization of sleep studies

Sleep research often relies on large retrospective clinical datasets. However, as the data is usually stored in proprietary formats specific for each sleep software, the raw data cannot be easily accessed and analyzed with external tools. While the raw data can usually be exported to more common data formats, this is often a cumbersome and labor-intensive task as it is not required for clinical purposes. Additionally, the recordings often include sensitive patient information which must be removed before the data can be shared or analyzed externally. This anonymization can be difficult to perform manually without the correct tools or knowledge of the file types and how they internally store the data. The STAR sleep recording export software provides a simple tool that can be used to perform the sleep study exports automatically. This allows the user to easily export a batch of sleep studies with minimal effort. In addition, the software can also be used to automatically anonymize the exported sleep recordings allowing researchers to save time and personnel resources as these do not need to be allocated for exporting and anonymizing sleep studies. The software supports Noxturnal, RemLogic, Profusion PSG and Sleepware G3 and it is free and openly available for anyone to download and use.

Functional association of NR4A3 downregulation with impaired differentiation in myeloid leukemogenesis

The coincident downregulation of NR4A1 and NR4A3 has been implicated in myeloid leukemogenesis, but it remains unknown how these two genes function in myeloid cells and how their combined downregulation promotes myeloid leukemogenesis. Since NR4A1 abrogation is thought to confer a survival and proliferation advantage to myeloid cells, we hypothesized that downregulation of NR4A3 may have a complementary effect on myeloid cell differentiation. First, we tested the association between differentiation status of leukemic cells and NR4A3 expression using two large clinical datasets from patients with different acute myeloid leukemia (AML) subtypes. The analysis revealed a close association between differentiation status and different subtypes of AML Then, we probed the effects of differentiation-inducing treatments on NR4A3 expression and NR4A3 knockdown on cell differentiation using two myeloid leukemia cell lines. Differentiation-inducing treatments caused upregulation of NR4A3, while NR4A3 knockdown prevented differentiation in both cell lines. The cell culture findings were validated using samples from chronic myeloid leukemia (CML) patients at chronic, accelerated and blastic phases, and in acute promyelocytic leukemia (APL) patients before and after all trans-retinoic acid (ATRA)-based differentiation therapy. Progressive NR4A3 downregulation was coincident with impairments in differentiation in patients during progression to blastic phase of CML, and NR4A3 expression was increased in APL patients treated with ATRA-based differentiating therapy. Together, our findings demonstrate a tight association between impaired differentiation status and NR4A3 downregulation in myeloid leukemias, providing a plausible mechanistic explanation of how myeloid leukemogenesis might occur upon concurrent downregulation of NR4A1 and NR4A3.

SIG-Former: monocular surgical instruction generation with transformers

Purpose:Automatic surgical instruction generation is a crucial part for intra-operative surgical assistance. However, understanding and translating surgical activities into human-like sentences are particularly challenging due to the complexity of surgical environment and the modal gap between images and natural languages. To this end, we introduce SIG-Former, a transformer-backboned generation network to predict surgical instructions from monocular RGB images.Methods:Taking a surgical image as input, we first extract its visual attentive feature map with a fine-tuned ResNet-101 model, followed by transformer attention blocks to correspondingly model its visual representation, text embedding and visual–textual relational feature. To tackle the loss-metric inconsistency between training and inference in sequence generation, we additionally apply a self-critical reinforcement learning approach to directly optimize the CIDEr score after regular training.Results:We validate our proposed method on DAISI dataset, which contains 290 clinical procedures from diverse medical subjects. Extensive experiments demonstrate that our method outperforms the baselines and achieves promising performance on both quantitative and qualitative evaluations.Conclusion:Our experiments demonstrate that SIG-Former is capable of mapping dependencies between visual feature and textual information. Besides, surgical instruction generation is still at its preliminary stage. Future works include collecting large clinical dataset, annotating more reference instructions and preparing pre-trained models on medical images.

In-Hospital Mortality of Sepsis Differs Depending on the Origin of Infection: An Investigation of Predisposing Factors.

Sepsis is a heterogeneous syndrome characterized by a variety of clinical features. Analysis of large clinical datasets may serve to define groups of sepsis with different risks of adverse outcomes. Clinical experience supports the concept that prognosis, treatment, severity, and time course of sepsis vary depending on the source of infection. We analyzed a large publicly available database to test this hypothesis. In addition, we developed prognostic models for the three main types of sepsis: pulmonary, urinary, and abdominal sepsis. We used logistic regression using routinely available clinical data for mortality prediction in each of these groups. The data was extracted from the eICU collaborative research database, a multi-center intensive care unit with over 200,000 admissions. Sepsis cohorts were defined using admission diagnosis codes. We used univariate and multivariate analyses to establish factors relevant for outcome prediction in all three cohorts of sepsis (pulmonary, urinary and abdominal). For logistic regression, input variables were automatically selected using a sequential forward search algorithm over 10 dataset instances. Receiver operator characteristics were generated for each model and compared with established prognostication tools (APACHE IV and SOFA). A total of 3,958 sepsis admissions were included in the analysis. Sepsis in-hospital mortality differed depending on the cause of infection: abdominal 18.93%, pulmonary 19.27%, and renal 12.81%. Higher average heart rate was associated with increased mortality risk. Increased average Mean Arterial Pressure (MAP) showed a reduced mortality risk across all sepsis groups. Results from the LR models found significant factors that were relevant for specific sepsis groups. Our models outperformed APACHE IV and SOFA scores with AUC between 0.63 and 0.74. Predictive power decreased over time, with the best results achieved for data extracted for the first 24 h of admission. Mortality varied significantly between the three sepsis groups. We also demonstrate that factors of importance show considerable heterogeneity depending on the source of infection. The factors influencing in-hospital mortality vary depending on the source of sepsis which may explain why most sepsis trials have failed to identify an effective treatment. The source of infection should be considered when considering mortality risk. Planning of sepsis treatment trials may benefit from risk stratification based on the source of infection.

Thyroid nodule segmentation and classification in ultrasound images through intra- and inter-task consistent learning.

Thyroid nodule segmentation and classification in ultrasound images are two essential but challenging tasks for computer-aided diagnosis of thyroid nodules. Since these two tasks are inherently related to each other and sharing some common features, solving them jointly with multi-task leaning is a promising direction. However, both previous studies and our experimental results confirm the problem of inconsistent predictions among these related tasks. In this paper, we summarize two types of task inconsistency according to the relationship among different tasks: intra-task inconsistency between homogeneous tasks (e.g., both tasks are pixel-wise segmentation tasks); and inter-task inconsistency between heterogeneous tasks (e.g., pixel-wise segmentation task and categorical classification task). To address the task inconsistency problems, we propose intra- and inter-task consistent learning on top of the designed multi-stage and multi-task learning network to enforce the network learn consistent predictions for all the tasks during network training. Our experimental results based on a large clinical thyroid ultrasound image dataset indicate that the proposed intra- and inter-task consistent learning can effectively eliminate both types of task inconsistency and thus improve the performance of all tasks for thyroid nodule segmentation and classification.

Semi-supervised approach to event time annotation using longitudinal electronic health records.

Large clinical datasets derived from insurance claims and electronic health record (EHR) systems are valuable sources for precision medicine research. These datasets can be used to develop models for personalized prediction of risk or treatment response. Efficiently deriving prediction models using real world data, however, faces practical and methodological challenges. Precise information on important clinical outcomes such as time to cancer progression are not readily available in these databases. The true clinical event times typically cannot be approximated well based on simple extracts of billing or procedure codes. Whereas, annotating event times manually is time and resource prohibitive. In this paper, we propose a two-step semi-supervised multi-modal automated time annotation (MATA) method leveraging multi-dimensional longitudinal EHR encounter records. In step I, we employ a functional principal component analysis approach to estimate the underlying intensity functions based on observed point processes from the unlabeled patients. In step II, we fit a penalized proportional odds model to the event time outcomes with features derived in step I in the labeled data where the non-parametric baseline function is approximated using B-splines. Under regularity conditions, the resulting estimator of the feature effect vector is shown as root-n consistent. We demonstrate the superiority of our approach relative to existing approaches through simulations and a real data example on annotating lung cancer recurrence in an EHR cohort of lung cancer patients from Veteran Health Administration.

Quantification of Digital Body Maps for Pain: Development and Application of an Algorithm for Generating Pain Frequency Maps.

BackgroundPain is an unpleasant sensation that signals potential or actual bodily injury. The locations of bodily pain can be communicated and recorded by freehand drawing on 2D or 3D (manikin) surface maps. Freehand pain drawings are often part of validated pain questionnaires (eg, the Brief Pain Inventory) and use 2D templates with undemarcated body outlines. The simultaneous analysis of drawings allows the generation of pain frequency maps that are clinically useful for identifying areas of common pain in a disease. The grid-based approach (dividing a template into cells) allows easy generation of pain frequency maps, but the grid’s granularity influences data capture accuracy and end-user usability. The grid-free templates circumvent the problem related to grid creation and selection and provide an unbiased basis for drawings that most resemble paper drawings. However, the precise capture of drawn areas poses considerable challenges in producing pain frequency maps. While web-based applications and mobile-based apps for freehand digital drawings are widely available, tools for generating pain frequency maps from grid-free drawings are lacking.ObjectiveWe sought to provide an algorithm that can process any number of freehand drawings on any grid-free 2D body template to generate a pain frequency map. We envisage the use of the algorithm in clinical or research settings to facilitate fine-grain comparisons of human pain anatomy between disease diagnosis or disorders or as an outcome metric to guide monitoring or discovery of treatments.MethodsWe designed a web-based tool to capture freehand pain drawings using a grid-free 2D body template. Each drawing consisted of overlapping rectangles (Scalable Vector Graphics <rect> elements) created by scribbling in the same area of the body template. An algorithm was developed and implemented in Python to compute the overlap of rectangles and generate a pain frequency map. The utility of the algorithm was demonstrated on drawings obtained from 2 clinical data sets, one of which was a clinical drug trial (ISRCTN68734605). We also used simulated data sets of overlapping rectangles to evaluate the performance of the algorithm.ResultsThe algorithm produced nonoverlapping rectangles representing unique locations on the body template. Each rectangle carries an overlap frequency that denotes the number of participants with pain at that location. When transformed into an HTML file, the output is feasibly rendered as a pain frequency map on web browsers. The layout (vertical-horizontal) of the output rectangles can be specified based on the dimensions of the body regions. The output can also be exported to a CSV file for further analysis.ConclusionsAlthough further validation in much larger clinical data sets is required, the algorithm in its current form allows for the generation of pain frequency maps from any number of freehand drawings on any 2D body template.